Concurrent Entanglement Routing for Quantum Networks:

Model and Designs

Shouqian Shi Chen Qian
[email protected] [email protected]
University of California, Santa Cruz University of California, Santa Cruz
Alice Bob Alice Bob
ABSTRACT
qubit
Quantum entanglement enables important computing applications
(a) Before teleportation (b) After teleportation
such as quantum key distribution. Based on quantum entanglement, entanglement
quantum networks are built to provide long-distance secret sharing Alice repeater Bob Alice repeater Bob
between two remote communication parties. Establishing a multi- data qubit
hop quantum entanglement exhibits a high failure rate, and existing (c) Before swapping (d) After swapping
quantum networks rely on trusted repeater nodes to transmit quan-
tum bits. However, when the scale of a quantum network increases, Figure 1: (a-b) Quantum teleportation to transmit a qubit (consum-
ing a local or distant entanglement). (c-d) Entanglement swapping
it requires end-to-end multi-hop quantum entanglements in order
to build a long-distance entanglement.
to deliver secret bits without letting the repeaters know the se-
cret bits. This work focuses on the entanglement routing problem, 1 INTRODUCTION
whose objective is to build long-distance entanglements via un- Secure information exchange via quantum networks has been pro-
trusted repeaters for concurrent source-destination pairs through posed, studied, and validated since 1980s [5, 17, 18, 34, 37, 47, 58] and
multiple hops. Different from existing work that analyzes the tradi- many experimental studies have demonstrated that long-distance
tional routing techniques on special network topologies, we present secrete sharing via quantum networks can become successful in
a comprehensive entanglement routing model that reflects the dif- reality, such as the DARPA quantum network [18], SECOQC Vi-
ferences between quantum networks and classical networks as well enna QKD network [37], the Tokyo QKD network [47], and the
as a new entanglement routing algorithm that utilizes the unique satellite quantum network in China [58]. A quantum network (also
properties of quantum networks. Evaluation results show that the called a quantum Internet) is an interconnection of quantum pro-
proposed algorithm Q-CAST increases the number of successful cessors and repeaters that can generate, exchange, and process
long-distance entanglements by a big margin compared to other quantum information [8, 10, 24, 56]. It transmits information in the
methods. The model and simulator developed by this work may form of quantum bits, called qubits, and stores qubits in quantum
encourage more network researchers to study the entanglement memories1 . Quantum networks are not meant to replace the classi-
routing problem. cal Internet communication. In fact, they supplement the classical
Internet and enable a number of important applications such as
CCS CONCEPTS quantum key distribution (QKD) [5, 17, 40], clock synchronization
[25], secure remote computation [7], and distributed consensus [15],
· Networks → Network protocol design; Routing protocols;
most of which cannot be easily achieved by the classical Internet.
· Hardware → Quantum communication and cryptography;
Most applications of quantum networks are developed based on
· Computer systems organization → Quantum computing.
two important features of quantum entanglement. 1) Quantum en-
tanglement is inherently private by the laws of quantum mechanics
KEYWORDS
such as the łno-cloning theoremž [36] and hence prevents a third
Quantum Internet; Quantum Networks; Entanglement Routing; party from eavesdropping the communication [17]. Quantum en-
Network Modeling tanglement enables a perfect solution to the fundamental problem
ACM Reference Format: of network security: key distribution (also known as key agree-
Shouqian Shi and Chen Qian. 2020. Concurrent Entanglement Routing for ment) [16]. Compared to public key cryptography [45], quantum
Quantum Networks: Model and Designs. In Annual conference of the ACM key distribution (QKD) has provable security in terms of informa-
Special Interest Group on Data Communication on the applications, technolo- tion theory and forward secrecy [56], instead of relying on the
gies, architectures, and protocols for computer communication (SIGCOMM computational complexity of certain functions such as factoriza-
’20), August 10ś14, 2020, Virtual Event, NY, USA. ACM, New York, NY, USA, tion. 2) Quantum entanglement provides strong correlation and
14 pages. https://doi.org/10.1145/3387514.3405853 instantaneous coordination of the communication parties. Hence,
quantum entanglement can achieve tasks that are difficult to co-
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed ordinate in classical networks, and a well-known one is quantum
for profit or commercial advantage and that copies bear this notice and the full citation teleportation, as shown in Fig. 1 (a) and (b). If a pair of entangled
on the first page. Copyrights for third-party components of this work must be honored. qubits are shared by Alice and Bob, then Alice can send one bit of
For all other uses, contact the owner/author(s).
SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA secret information to Bob with the help of quantum measurement
© 2020 Copyright held by the owner/author(s).
1 e.g.,
transmitting a pair of entangled photons and storing the entanglement state into
ACM ISBN 978-1-4503-7955-7/20/08.
https://doi . org/10 . 1145/3387514 . 3405853 a pair of nitrogen-vacancy centers in two remote diamonds [11, 13]
SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA Shouqian Shi and Chen Qian

and the classical Internet [44]. Hence, QKD can be achieved via link-layer models [13, 23, 35]. While the physical layer and link
quantum entanglement. layer studies of quantum networks require experimental valida-
We note that quantum networks will become practical in the tions on special and expensive hardware, the entanglement routing
near future, and they do not rely on the success of well-functioning algorithms can be comprehensively evaluated via simulations as
quantum computers. Both academia and industry have a time-to- long as the link-layer model reflects the practical physical facts,
time debate on when a practical quantum computer will be available similar to prior studies for intra-/inter-domain routing, wireless
with a sufficient amount of qubits to implement the proposed quan- multi-hop routing [19], data center routing [3, 50, 59], etc.
tum algorithms, such as Shor’s integer factorization [49]. It seems To our knowledge, this is the first work of a comprehensive
that well-functioning quantum computers might not become avail- protocol design specifically for entanglement routing in quantum
able in the near future. However, many applications of quantum networks, with new models, new metrics, and new algorithms,
networks can be implemented with one or two qubits. Considering working on arbitrary network topologies. We present a compre-
the QKD example, we are able to distribute a secret bit with only hensive entanglement routing model that reflects the difference
one entanglement pair. By repeating the 1-pair QKD process we between quantum networks and classical networks and propose
can generate secret keys with a sufficient length. new entanglement routing designs that utilize the unique prop-
To generate a quantum entanglement between two parties Alice erties of quantum networks. We propose a few routing metrics
and Bob, an entangled pair of photons are created, and each pho- that particularly fit quantum networks instead of using hop-count
ton is sent to a party through a channel, such as an optical fiber. and physical distance. The proposed algorithms include realistic
However, the optical fiber is inherently lossy and the success rate protocol-design considerations such as arbitrary network topolo-
p of establishing an entanglement pair decays exponentially with gies, multiple concurrent sources and destinations to compete for
the physical distance between the two parties [41, 42]. Hence, to resources, link state exchanges, and limited qubit capacity of each
increase the success rate of long-distance quantum entanglement, node, most of which have not been considered by prior studies.
a number of quantum repeaters need to be deployed between two Evaluation results show that the proposed algorithm Q-CAST in-
long-distance communication parties [42, 56]. Many existing quan- creases the number of successful long-distance entanglements by a
tum networks [18, 37, 47, 58] rely on łtrusted repeatersž to relay big margin compared to other known methods. More importantly,
entanglements. Each trusted repeater gets the actual data qubit tele- this study may encourage more network researchers to study the
ported from the sender and teleports the data qubit to the receiver, entanglement routing problem. We present and clarify the mod-
similar to the łstore-and-forwardž process in classical networks. A els and problems of entanglement routing, with the comparison
more attractive approach is to use quantum swapping [6, 33, 34]. of similar terms and concepts used in classical network research.
As shown in Fig. 1 (c) and (d), via entanglement swapping, a quan- A simulator with algorithm implementation, topology generation,
tum repeater that holds entanglements to both Alice and Bob can statistics, and network visualization functions is available on this
turn the two one-hop entanglements into one direct entanglement link [1].
between Alice and Bob. Multi-hop swapping is also possible with a The rest of this paper is organized as follows. Section 2 presents
path of repeaters holding entanglements with their predecessors the related work of quantum network routing and Section 3 in-
and successors. During quantum swapping, a repeater does not troduces the network model. We present the algorithm designs in
know the qubit information hence it does not have to be trusted. Section 4. The evaluation results are shown in Section 5. We discuss
This work focuses on a key problem called entanglement routing, some related issues in Section 6 and conclude this work in Section 7.
whose objective is to build long-distance entanglements through 2 RELATED WORK
multiple hops of quantum repeaters and entanglement swapping, Quantum information exchange has been proposed, studied, and
even if the repeaters may be untrusted or corrupted [23, 35]. En- validated for more than 20 years [5, 17, 18, 34, 37, 47, 58]. The
tanglement routing has not been thoroughly investigated but is concept of quantum networks is first introduced by the DARPA
necessary in future large-scale quantum networks: When a quan- quantum network project aiming to implement secure communi-
tum network scale increases, similar to the Internet, users do not cation in the early 2000s [18]. Recent implementations include the
always trust all forwarding devices between the source and des- SECOQC Vienna QKD network [37], the Tokyo QKD network [47],
tination or some trusted repeaters may be corrupted. In addition, and China’s satellite quantum network [58]. These experimental
a large number of trusted repeaters increase the attack space and works rely on trusted repeaters.
the vulnerability of the whole system. Entanglement routing finds In order to design future large quantum networks in which re-
an end-to-end path of concurrent quantum entanglement through peaters may not trust each other, one fundamental problem is to
a number of repeaters and performs quantum swapping without route quantum entanglements with high reliability in quantum
letting the repeaters know the data bits. This can be considered repeater networks [53]. Van Meter et al. studies applying Dijkstra
on the network layer of a quantum network [13]. Existing works algorithm to repeater network [30]. Pirandola et al. discuss the
that investigate the entanglement routing problem of quantum net- limits of repeaterless quantum communication [42] and propose
works are limited to analyzing the traditional routing algorithms multi-path routing in a diamond topology [39]. Schoute et al. [48]
(Dijkstra shortest paths, multipath routing, and greedy routing) on propose a framework to study quantum network routing. However,
special network topologies (ring, sphere, or grid), such as the very their discussion is only limited to ring or sphere topology. Das et al.
recent ones [23, 35]. [14] compare different special topologies for entanglement routing.
Similar to other network routing problems, entanglement routing Caleffi [9] studies the optimal routing problem in a chain of re-
is a distributed algorithm design problem to utilize the underlying peaters. Pant et al. [35] propose solutions for entanglement routing
Concurrent Entanglement Routing for Quantum Networks: Model and Designs SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA

in grid networks. [23] proposes virtual-path based greedy routing quantum processors by quantum channels to form a quantum net-
in ring and grid networks. Vardoyan et al. [54] study a quantum work and run the network applications to communicate with each
entanglement switch in a star topology. All these studies assume other. Different from classical end hosts, each quantum processor
specialized network topologies such as a grid or ring, which may is equipped with a certain number of memory qubits and necessary
be over-simplified. The topologies of practical quantum networks hardware to perform quantum entanglement and teleportation on
may be arbitrary graphs because 1) the end hosts in quantum net- the qubits. All quantum processors are connected via the classical
works must exist on specified locations according to application Internet and are able to freely exchange classical information.
requirements, instead of following certain topologies; 2) deploying 2. Quantum repeaters. As it is difficult to directly establish
unnecessary devices just to create a certain topology is a waste of an entanglement between two remote quantum processors, quan-
resource. tum repeaters are used as relays. Quantum repeaters support long-
The above studies are limited to analyzing the traditional routing distances entanglement sharing via quantum swapping. A quantum
algorithms on special network topologies. Compared to them, this repeater may also connect to other repeaters and quantum proces-
paper is the first work of a comprehensive protocol design specif- sors via the classical Internet to exchange the control messages.
ically for entanglement routing in quantum networks, with new Every quantum processor also includes the complete function of
models, new metrics, and new algorithms, working on arbitrary a repeater. Hereafter we call both quantum processors and repeaters
network topologies. It includes three unique improvements: 1) We as nodes.
present a practical network model that clearly specifies the net- 3. Quantum channels. A quantum channel connecting two
work information that is locally known to each node, includes more nodes supports the transmission of qubits. The physical material of
practical network topologies such as arbitrary network graphs, and quantum channels may be polarization-maintaining optical fibers.
present locally executed protocols on every single node. 2) This A quantum channel is inherently lossy: the success rate of each
work considers concurrent source-destination pairs that may cause attempt to create an entanglement of a quantum channel c is pc ,
contention on quantum links. Concurrent routing is one of the which decreases exponentially with the physical length of the chan-
most important design challenges of quantum networks because nel: pc = e −α L , where L is the physical length of the channel and
each quantum link can only be used for one source-destination pair, α is a constant depending on the physical media [35, 42, 52, 56].2
unlike packet switching. We believe our solution matches practical If an attempt is successful, the two quantum processors share an
quantum network applications. Prior methods are not specifically entanglement pair, and there is a quantum link on this channel.
designed for concurrent source-destination pairs and might become Network topology. Consider a network of quantum nodes de-
sub-optimal in practical situations. 3) We propose a few routing scribed by a multigraph G = ⟨V , E, C⟩. V is the set of n nodes. Each
metrics that particularly fit quantum networks instead of using node u is a quantum processor, equipped with a limited number
hop-count and physical distance. These metrics are important to Qu of qubits to build quantum links. All nodes are connected via
select good paths in quantum networks and can also be used for classical networks, i.e., the Internet, and every node has a certain
future studies. level of classical computing and storage capacity, such as a desktop
Recently, Dahlberg et al. [13] provide a reference model of the server. E is the set of edges in the graph. An edge existing between
quantum network stack, which contains the physical layer, link two nodes means that the two nodes share one or more quantum
layer, network layer, and transport layer. Based on that, they provide channels. C is the set of all quantum channels, each of which is
a reliable physical and link layer protocol for quantum networks identified by its two end nodes. The number of channels on an edge
on the NV hardware platform. The routing algorithms proposed is called the width W of the edge.
in our paper fit in the ‘network layer’ [13] to provide the concur- A node can assign/bind each of its quantum memory qubits
rent entanglement routing solutions, leveraging the services in the to a quantum channel [27, 28], such that no qubit is assigned to
quantum link layer. more than one channel, and no channel is assigned more than one
qubits at the same end of it. Channels that are assigned qubits
3 NETWORK AND SECURITY MODELS at its both ends are bound channels, other channels are unbound
The network model used in this study follows the facts from ex-
channels. There could be more than one bound channels between
isting physical experiments [6, 31, 33, 34] and the corresponding
two nodes. And two neighbor nodes may share multiple quantum
studies [14, 35, 48] to reflect a practical quantum network. Com-
links. To create a quantum entanglement, two neighbor nodes make
pared to prior models used in existing studies of quantum network
a number of quantum entanglement attempts at the same time on
performance [14, 23, 35, 48], this model includes many practical con-
the bound channels connecting them.
siderations, e.g. the dynamics of quantum links, definition and com-
parison of different routing metrics, concurrent source-destination 3.2 Communication and security model
pairs, limited qubit capacity of each node, clear differentiation of For each round of communication, the source and destination are
the network topology and link state information, and limited link two mutually trusted quantum processors, but they may not trust
state propagation in a time slot. other nodes. The source aims to deliver secret bits to the destina-
tion without letting other repeaters know, via a path of quantum
3.1 Network components
There are three main components in a quantum network [51, 56], 2 The success rate of a link is determined by the physical layer and link layer, taking
explained as follows. into account the channel transmissivity, fidelity of transmitted entanglements, number
of permitted entanglement trials in one phase, and the link layer algorithm [13, 35]. In
1. Quantum processors are similar to the end hosts in classi- the link layer, a channel is allowed multiple attempts to build a link, and the link is
cal networks, which are connected to a certain number of other established on the first successful attempt. The pc here is the overall success rate.
SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA Shouqian Shi and Chen Qian

P1 P1 A C D B A C D B
Receives SD pairs Receives SD pairs

P2: qubit assignment P2: qubit assignment E E

One time slot

Input: network topology and

(a) P1: All nodes are informed of (b) P2 (external phase): paths are
SD pairs
Output: selected paths, qubit Same as the left the S-D pair: A-B, and qubits are found (solid lines) from A to B, and
to channel assignment not bounded to any quantum each node assigns qubits to
channels (dashed lines) channels according to the paths
P3 P3
A C D B A C D B
Exchange link state Exchange link state

P4: place internal link P4: place internal link

Input: selected paths and
E E
local link state Same as the left
Output: internal links (c) P3: each channel may fail or (d) P4 (internal phase): each node
succeed to build an entanglement locally decides swapping (bold
Node 1 Node 2 (solid line). These link states are solid lines in C and D). Then, A and
shared within 𝒌 hops. B share a remote entanglement.
Figure 2: Phases in a time slot. Nodes 1 and 2 are two arbitrary
neighbor nodes and run the same algorithm. Figure 3: Phases in a time slot. Entanglement routing aims to build
end-to-end paths for S-D pairs (A-B in this example).
swapping (explained in Sec. 3.3). All nodes will follow the protocol
running on all nodes that produces consistency results. Each node
but may seek to get the secret information sent from the source to
then binds its qubits to channels and attempts to generate quantum
the destination, similar to the łhonest-but-curiousž model in classic
entanglements with neighbors on the bound channels [27, 27]. As
network security. Once an intermediate node measures the infor-
an example in Fig. 3(b), two paths (solid curves) are calculated to
mation to perform passive eavesdropping, such behavior will be
connect A and B. A path is identified by the sequence of the nodes
detected by the two endpoints according to the no-cloning theorem.
along the path v 0 , v 1 , · · · , vh and the path width W , meaning each
In addition, an external classical ‘network information server’
edge of the path has at least W parallel channels. The path ⟨(v 0 ,
may be trusted to maintain the following information and send
v 1 , · · · , vh ), W ⟩ is also called a (W , h)-path, or a W -path. C, D, and
delta updates to all nodes in the network when necessary: 1) the
E are nodes on the paths and work as repeaters. Since qubits are
network topology and 2) the current source-destination pairs (S-
limited resources, some channels are not assigned qubits and thus
D pairs) that need to establish long-distance entanglements. The
not used in this time slot. During P2, each channel can make a
network information server may work in an honest-but-curious
number nc of attempts [20], nc ≥ 1, until a link is built or timeout.
way and it will not know the communication content. Hence even
After P2, some quantum links may be created as shown in Fig. 3(b).
if a network information server may be comprised ś which can
We call the information of these links as link states. Compared to
be detected by classical auditing methods ś it will not hurt the
the same term in link-state routing of classical networks [32], the
confidentiality for previous, on-going, and future communications.
quantum link states are highly dynamic and nondeterministic.
3.3 Quantum swapping via a path In Phase Three (P3), each node knows its own link states via
Time slots. For multi-hop entanglement swapping, all nodes on classical communications with its neighbors [35] and shares its
the path need to establish and hold quantum entanglements with link states via the classical network, as shown in Fig. 3(c). Since
its predecessor and successor at the same time. Hence, some level entanglements will quickly decay, each node can only exchange the
of time synchronization among all nodes is necessary, which can be link states with a subset of other nodes. P3 only includes classical
achieved by existing current synchronization protocols via Internet information exchange.
connections. Time is loosely synchronized in time slots [35]. Each In Phase Four (P4), also called the internal phase [35], nodes per-
time slot is a device-technology-dependent constant and set to an form entanglement swapping to establish long-distance quantum
appropriate duration such that the established entanglements do entanglement using the successful quantum links. Each node locally
not discohere within one time slot [35]. The global network topol- determines the swapping of successful entanglements, which can
ogy G = ⟨V , E, C⟩, which is relatively stable, should be common be considered as placing an internal link between two qubits as
knowledge for all nodes before any time slot. shown in Fig. 3(d). Each swapping succeeds at a device-dependent
As shown in Fig. 2 and Fig. 3, each time slot includes four phases probability q. A and B can successfully share an entanglement qubit
as an extended model from [35]. In Phase One (P1), via the Internet, pair (an ebit) if there is an end-to-end path with both external and
all nodes receive the information of the current S-D pairs that need internal links as shown in Fig. 1(d).
to establish long-distance entanglements. As an example in Fig. 3(a), After P4, the secret bit can be teleported from the source to
each node has a number of qubits (red dots) and multiple quantum the destination. Eavesdropping attempts at any repeater will be
channels (dashed lines) connecting neighbors. Two neighbors may detected hence the confidentiality is preserved.
share multiple channels. Suppose ⟨A, B⟩ is the only S-D pair for this Local knowledge of link-state. P3 and P4 should be short such
time slot, and all nodes are informed of the S-D pair. that the successful entanglements built in P2 do not decay. Hence,
Phase Two (P2) is called the external phase [35]. In P2, paths are it is impractical for a node to know the global link states within
found for the S-D pairs, according to an identical routing algorithm such a short time as the classical network has latencies [35]. A
Concurrent Entanglement Routing for Quantum Networks: Model and Designs SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA

practical model is to allow each node to know the link states of its 3.5 Compared with classic network routing
k-hop neighbors, k ≥ 1 [23]. The swapping decisions in P4 thus We summarize the differences between quantum entanglement
include the k-hop link-state information as the input. It is obvious routing and classic network routing. We show that existing routing
that the routing path selection could be sub-optimal without global techniques are not sufficient to solve the entanglement routing
link-state knowledge. problem.
Exclusive qubit/channel reservation. In P2 of each time slot, Term clarification. Edges, channels, and links have different
to establish a single link on a channel, each end of the channel is definitions in this model, although they are used interchangeably
assigned a qubit. This qubit-channel assignment is exclusive: one in classic networks. Besides, the network topology and global link
qubit cannot be shared by other channels, and no more qubits can states may be considered as similar information in classic routing
be assigned to a channel. In P4, to generate distant entanglements such as OSPF [32]. However, in a quantum network, while the
from local ones, quantum swapping is performed on pairs of links. network topology (nodes and channels) is stable and known to all
This quantum swapping is also exclusive and a single link cannot be nodes, the link states (whether the entanglements succeeded) are
used for more than one swapping. Hence, the qubits and channels dynamic and only shared locally in P3 and P4 of each time slot.
are precious routing resources and should be carefully managed. Versus routing in wired packet-switching networks. Link-
Physical parameters. We show the physical parameters of typ- state and distance-vector are two main types of routing protocols for
ical quantum networks, which provide several insights into our packet-switching networks. Main differences: 1) Packet switching
model and design. 1) The short entanglement persistent time de- relies on either link-state broadcast or multi-round distance vector
termines the nodes should be synchronized to ensure all links are exchanges to compute the shortest paths. However, in a quantum
available simultaneously for selected paths. 2) The short entangle- network, link states are probabilistic and vary in different time
ment persistent time T sets the limit that t 2 + t 3 + t 4 < T in Fig. 3. slots. There is no time for global link-state broadcast or distance
3) The local link state cannot be propagated to the whole network. vector convergence, because entanglements on the links will quickly
4) The qubit capacity is bounded in a node so that the dynamic decay. 2) Quantum links are highly unreliable while wired links are
binding of qubits and channels are necessary in P2. A most recent relatively reliable. 3) A quantum link cannot be shared by multiple
quantum processor can have up to 8 qubits [13]. The typical time S-D pairs, which is allowed in classic packet switching. If a link
for an entanglement to discohere is 1.46s [13]. The entanglement is claimed by multiple S-D pairs, it can only satisfy one of them.
establishment time is ~165µs for concurrent trying [13]. The suc- Hence, the łshortest pathsž computed by classic routing will not
cess rate of a single entanglement try is dependent on the length of always be available. 4) Classic packets can be buffered on any node
the optical fiber and is typically ~0.01% [13]. Multiple concurrent for future transmission. In quantum networks, links on a path must
entanglement tries are possible within a time slot to have a reason- be successful in the same time slot.
able channel success rate in P2. A typical classical communication Versus routing in multi-hop wireless networks, such as mo-
finishes at ~1ms, in a dedicated optical fiber network. Balancing bile ad hoc networks [19] and wireless sensor networks [2]. Main
the time for P2 and P3 are necessary to have both high channel differences: 1) For an ad hoc wireless node, neither the network
success rate in P2 and a large enough local view of link states in P3. topology nor global link state is known. For a quantum node, al-
The entanglement readout time is typically < 3µs [13], negligible though link state is local information, the network topology is
in routing algorithm design. known in advance via the Internet. 2) An ad hoc wireless node
3.4 The entanglement routing problem can keep sending a packet until the transmission is successful or a
This work studies the entanglement routing problem: we are given preferred receiver moves close to it. Each quantum link can only
a quantum network with an arbitrary network graph G = ⟨V , E, C⟩ be used once and all links on an end-to-end path must be available
and a number of source-destination (S-D) pairs ⟨s 1, d 1 ⟩, ⟨s 2, d 2 ⟩, simultaneously. 3) Existing wireless ad hoc routing methods such
· · · , ⟨sm , dm ⟩, where si , di ∈ V . The number of memory qubits of as DSR [21], AODV [38], and geographic routing [22, 26, 43] are all
a node u ∈ V is Qu , and each edge e ∈ E consists of one or more packet-switching protocols and do not fit quantum networks. Also,
channels from C. For each bound channel c, a link is successfully they do not take the global network topology information.
built at a probability pc in P2. In P3, each node gets the link-state Versus circuit-switching, virtual circuit, and flow schedul-
information of its k-hop neighbors. Each node decides the swapping ing in SDN. Circuit switching, virtual circuit, and flow scheduling
of its internal qubits in P4 locally, and each swapping succeeds in in software defined networks (SDNs) all need to pre-determine the
probability q. The objective of entanglement routing is to maximize end-to-end paths and reserve certain resources on the paths, such as
the number of ebits delivered for all S-D pairs in each time slot. Each [3, 4, 12, 46], which share similarity with entanglement routing. The
ebit must be delivered by a long-distance quantum entanglement, main difference is that in a quantum network, though the topology
built by a path of successful quantum links from the source to the (nodes and channels connecting them) is relatively stable, reserved
destination. Each S-D pair may share multiple ebits. The number of paths for an S-D pair are not reliable because links may arbitrarily
ebits for one S-D pair in one time slot is also called the throughput fail. Hence, more robust algorithms are required. Besides, to build a
between the S-D pair. The objective can then be set to maximizing long-distance entanglement along a path, all hops of the path should
the overall throughput in the network. have one or more success quantum link at the same time. Hence,
This objective does not consider fairness among different S-D time is divided into slots and phases for synchronization. Due to
pairs, but we show the proposed algorithms achieve a certain level the two differences above, the algorithm of entanglement routing is
of fairness as in ğ 5. In addition, in ğ 6 we propose a simple extension very different from that of circuit-switching, in the following two
to our designs to achieve better fairness among S-D pairs. novel designs: 1) multiple paths are selected in P2, based on global
SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA Shouqian Shi and Chen Qian

C D
and stable network topology; and 2) path recovery on P4, based on A C D B A B
local and probabilistic link states. The recently proposed multipath
routing for quantum entanglements [35] is a circuit-switching style
protocol and will be compared to our work in Sec. 5. C’ D’
(a) One path with width two can deliver
(b) Two disjoint paths may easily fail
one qubit even under three failed links
with two unsuccessful links
4 ENTANGLEMENT ROUTING ALGORITHMS
The proposed entanglement routing algorithms utilize the unique Figure 4: A wide path (subfigure a) is more reliable than disjoint
properties of quantum networks that have not been explored in paths (subfigure b) using the same resource
classic network routing. Compared to recent quantum network 2.5 3-path 1.75 3-path
1.50
studies [14, 23, 35, 48], the proposed protocols follow a standard 2.0 2-path
1.25
2-path
1-path 1.00 1-path

EXT

EXT
1.5
protocol-design approach and use more realistic network models: 0.75
1.0 0.50
arbitrary network topologies, multiple concurrent S-D pairs to 0.5 0.25
compete for links, link state exchanges, and limited qubit capacity 0.00
2 4 6 8 10 2 4 6 8 10
of each node. Number of hops Number of hops

4.1 Main ideas Figure 5: EXT, p = 0.9 Figure 6: EXT, p = 0.6

Our design is based on the following three innovative ideas to W
Õ
utilize the unique features of a quantum network: Et = q h · i · Phi (2)
1. Path computation based on global topology and path re- i =1
covery based on local link states. The quantum network graph We show some numerical results. For simplicity, we set p1, p2,
G = ⟨V , E, C⟩ is relatively stable and hence can be known to every · · · , ph = p, and let p be 0.9 or 0.6. We vary the W from 1 to 3 and the
node. However, the link states are highly dynamic and probabilis- h from 1 to 10, and the results of the EXT of a W -path are shown
tic in each time slot. The frequent link state changes cannot be in Figures 5 and 6. It is obvious that a W -path has a significant
propagated throughout the whole network, especially when the improvement of EXT over a 1-path, for more than a factor of W .
entanglements decay quickly. Hence, nodes select and agree on the 3. Offline computation versus contention-aware online path
same list of paths based on global topology information in P2, and selection. In different time slots, the S-D pairs may be different.
try to recover from link failures based on local link states in P4. We propose two approaches to select paths for S-D pairs in each
2. Wide paths are preferred. Recall that on a W -path, each time slot. The first approach utilizes offline computation, which hap-
edge has at least W parallel channels. Fig. 4(a) shows an example pens before any time slot, e.g., during system initialization. Multiple
of a 2-path from A to B. Compared to two disjoint paths shown paths for each potential S-D pair are pre-computed and stored by
in Fig. 4(b), which cost the same amount of qubits and channels, all nodes as common knowledge. In P2 of each time slot, nodes
the wide path is more reliable because it only fails when two links select the pre-computed paths for current S-D pairs. The contention-
fail simultaneously at a single hop. To achieve high throughput aware online algorithm, however, does not pre-compute the paths
on a path with W > 1, nodes should share a consensus on how to for all S-D pairs. At each time slot, the algorithm finds contention-
perform swapping (place internal links in Fig. 4) instead of making free paths for current S-D pairs. A set of paths are ‘contention-free’
choice randomly. Each channel is assigned a globally unique ID. if the network can simultaneously satisfy the qubit and channel
During P4, each node places an internal link between the link with requirement for all the paths in full width.
the smallest ID to its predecessor and the link with the smallest ID 4.2 Q-PASS: Pre-computed pAth Selection and
to its successor. And it repeats this process until no internal link Segment-based recovery
can be made for this path.
Formally, we may define a routing metric, called the expected 4.2.1 Algorithm overview. We present the algorithm Q-PASS, whose
number of ebits or expected throughput (EXT) Et , to quantify workflow follows the four-phase time slot model with an additional
an end-to-end path on the network topology. For a (W , h)-path offline phase. The core idea of Q-PASS is to pre-compute potential
P, suppose the success rate of a single channel on the i-th hop is ‘good’ paths between all possible S-D pairs based on the network
pi , where i ∈ {1, 2, · · · , h}. We denote the probability of the k-th topology G. Then in each time slot, every node uses an online
hop on the path having exactly i successful links as Q ki , and the algorithm to make qubit-to-channel assignments based on the pre-
computed paths of current S-D pairs and make local swapping de-
probability of each of the first k hops of P has > i successful links
cisions based on local link states. The design includes both offline
as Pki . Then we get the recursive formula set, for i ∈ {1, 2, · · · ,W }
and online algorithms.
and k ∈ {1, 2, · · · , h}:
The offline computation happens at the system initialization and
Q ki = W
i W −i after the network topology changes. The results of an offline phase
i pk (1 − pk )
can be used by many succeeding time slots until a topology change
P 1i = Q 1i
(1) happen. Hence, we may assume the time for an offline period is
W W
Õ Õ sufficiently long. The offline algorithm runs at the network infor-
Pki = Pki −1 · Q kl + Q ki · Pkl −1
mation server, which is honest but curious, with replica servers
l =i l =i +1
for robustness. These servers connect to all quantum nodes via
Further, considering the success probability q of each entanglement classical networks. The output of the offline algorithm is the łcan-
swapping, we get the EXT: didate pathsž for all possible S-D pairs. The candidate paths of each
Concurrent Entanglement Routing for Quantum Networks: Model and Designs SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA

C’ D’ C’ D’ C’ D’

A B A B A B

C D E C D E C D E
(a) Offline algorithm finds two paths for A-B: (b) In P2, ACDEB is reserved as the main path. (c) In P4, D finds that the main path is disconnected.
ACDEB and AC’D’DEB AC’D’DEB does not have enough resource but part of It chooses to route through AC’D’D, and swaps ink
it can be the recovery path DE with link DD’, instead of CD-DE
Figure 7: Example of Q-PASS. Suppose ⟨A, B ⟩ is the only S-D pair.
S-D pair are paths connecting the S-D nodes and with the smallest 50% percent in the next offline phase if the paths happen to be not
values of the selected metric. enough for a pair.
The algorithm of each time slot follows the four-phase time 4.2.3 P2 algorithm of Q-PASS. The P2 algorithm runs on each node
slot model shown in Fig. 2 and runs on each node in a distributed locally. The inputs are all the offline paths P (known before P1) and
and concurrent manner. It should be fast and only use the k-hop the S-D pairs (received in P1) O = {oi }, where oi is an S-D pair
link-state information. P1 and P3 only include standard processes ⟨si , di ⟩. The output is an ordered list of selected paths P ′ , each of
and do not have special algorithmic designs. Q-PASS P2 takes the which connects a single S-D pair in O. According to the output
candidate paths from the offline algorithm and the S-D pairs as path list, each node performs the local qubit-to-channel assignment
the input. It computes a number of selected paths for the S-D pairs and tries to establish entanglements on the bound channels with
and its local qubit-to-channel assignment. Note that the inputs neighbor nodes to build quantum links on these paths. Since P and
are globally consistent on all nodes. Hence, the selected paths are O are globally known for all nodes, the output P ′ is also consistent
also consistent on different nodes. The assignment will produce on different nodes, similar to the global consistency of classical
a number of successful links in P2. And in P3, nodes exchange link-state routing.
the link states with their k-hop neighbors. Q-PASS P4 uses the The algorithm consists of two steps. Step 1) The paths computed
selected paths and link state information as the input to compute from the offline algorithm for all S-D pairs are retrieved and put
the swapping decisions (i.e., internal links). After P4, possible long- into a priority queue, ordered by the selected routing cost metric.
distance entanglement can be built for S-D pairs. We present the Then from the path with the lowest routing cost to the highest,
algorithms in detail. channels and qubits taken by the path are reserved exclusively. If a
4.2.2 Offline path computation. The offline algorithm should find path has width w by the offline algorithm, but currently available
multiple paths for each S-D pair to provide multiple candidate paths resource can only support width 0 6 w ′ < w, then the path is
in P2 of each time slot. We use Yen’s algorithm [57] to get multiple reinserted to the queue with an updated metric calculated from w ′ .
paths for each pair. Note that the results of Yen’s algorithm are not This process ends until no paths can be fully satisfied. The paths
contention-free: the paths may overlap at nodes or channels, and selected in Step 1 are called major paths. Step 2) After Step 1, the
in a single time slot, the network may not have enough qubits or queue contains all unsatisfiable paths in the ascending order of the
channels to satisfy all the candidate paths for an S-D pair. routing metric. Each unsatisfiable path, however, may contain one
Yen’s algorithm implicitly requires a selection of the routing met- or more satisfiable segments or ‘partial paths’. The partial paths can
ric. As shown in Equ. 1, computing the proposed routing metric EXT be used to recover link failures for the major paths, and thus are
involves recursions, which may be prohibitively slow for multi-path called recovery paths. The qubits and channels for recovery paths
computation for all possible S-D pairs. Hence, we propose three are reserved in the order of its priority in the queue.
routing metrics, which are suboptimal in overall throughput but For example in Fig. 7, ACDEB and AC ′D ′DEB are two candi-
faster to compute. 1) Sum of node distances (SumDist). SumDist date paths. In Step 1, ACDEB and AC ′D ′DEB are put into a priority
is computed as ΣLi , where Li is the length of any channel on the queue, and ACDEB is more prioritized than the other and is reserved
i-th hop of the path. As the success rate of a channel decreases as the major path. Since D, E, and B do not have enough resources
exponentially with the physical distance L, SumDist can partially for AC ′D ′DEB, Step 1 stops. In Step 2, AC ′D ′D is reserved as a re-
reflect the difficulty of a path. 2) Creation Rate (CR). CR is com- covery path. When the two steps finish, all nodes know the same set
puted as Σ1/pi , where pi is the success rate of any channel on the of selected major paths and recovery paths because they share the
i-th hop of the path. Compared to SumDist, CR further considers same set of inputs: network topology, S-D pairs, and offline paths.
the path width. 3) Bottleneck capacity (BotCap). From Figures 5 Hence, each node assigns its qubits to the corresponding channels
and 6, the path width W has a greater impact on the path qual- and try to generate quantum links together with the neighbors. For
ity. The BotCap metric is −W , prefers wider paths over narrower example, node A in Fig. 7(b) will assign one qubit to the channel to
paths, and uses the CR to break ties for paths with the same width. C and another to the channel to C ′ , and try to entangle with C and
We consider the routing metric as a design parameter, and their C ′ via channel AC and AC ′ respectively. The pseudocode for the
efficiency is compared in ğ 5. P2 algorithm is shown in Alg. 1.
For each possible S-D pair, the server running the offline algo- 4.2.4 P4 algorithm of Q-PASS. If the entanglement attempts in P2
rithm will use Yen’s algorithm to get N = 25 paths (offline paths) always succeed, each node just performs entanglement swapping
for the pair and tell each node in the network about the resulting to connect the links on the major paths, and the whole paths will
paths. An example is shown in Fig. 7(a), the offline algorithm finds be successful. In practice, however, link failures happen at a high
ACDEB and AC ′D ′DEB as two candidate paths. N will grow by probability and are not predictable. The P4 algorithm focuses on
SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA Shouqian Shi and Chen Qian

the recovery of broken major paths based on the recovery paths B’ B B’ B

𝑀25 𝑀25
established in P2. The inputs of P4 algorithm are: 1) S-D pairs from
P1, 2) a major path list and a recovery path list from P2, and 3) the … …
𝑀2 𝑀2
k-hop link states of this node from P3.
𝑀1 𝑀1

Algorithm 1: Adaptive resource allocation A A’ A A’

Input :G = ⟨V , E, C ⟩, O , P
// O : list of S-D pairs (a) A-B and A’-B’ are two S-D pairs. Q-PASS (b) Q-CAST also reserves all qubits on
offline algorithm outputs 25 paths for A-B: 𝑴𝒊 for path 𝑨’ 𝑴𝟏 𝑴𝟐 ⋯ 𝑴𝟐𝟓 𝑩’. However, in
// P : mapping from any S-D pair to its offline paths 𝑨𝑴𝒊 𝑩, 𝒊 ∈ {𝟏, 𝟐, … , 𝟐𝟓} . However, path the residual graph, Q-CAST finds two
Output : ⟨LC , L P ⟩ 𝑨’ 𝑴𝟏 𝑴𝟐 ⋯ 𝑴𝟐𝟓 𝑩’ is better by metric than more paths for A-B: 𝑨𝑩′ 𝑩 and AA’B. These
𝑨𝑴𝒊 𝑩 and reserves all qubits on 𝑴𝒊 . The paths are worse than 𝑨𝑴𝒊 𝑩 by metric,
// LC : list of channels to assign qubits pair A-B fails due to no more offline paths. thus not recorded in the offline Q-PASS.
// L P : ordered list of selected paths
Figure 8: Comparison of Q-PASS and Q-CAST
1 LC ← œ
2 LP ← œ ACDEB is divided into two segments ACD and DEB, such that all
3 TQ ← a table to map a node x to its qubit capacity Q x nodes on a single segment know this segment is successful or not. If
4 construct TQ from current topology not, they will try to use a recovery path. In this example, A, C, and
5 W ←œ D know link C-D fails. Hence, the recovery path AC ′D ′D is taken
// empty table to map a path p to its width w p by D. The distributed recovery path selection is consistent among
6 q←œ all nodes because recovery paths are found from local link states
// empty priority queue of paths, sorted by routing known to all involved nodes, and the recovery paths are ordered
metric deterministically via the specified routing metric.
7 for o ∈ O do
8 for p ∈ P [o] do
4.3 Q-CAST: Contention-free pAth Selection at
9 TW [p] ← W idt h(p, TQ ) runTime
10 m ← routing metric of p with width W[p] The offline algorithm in Q-PASS has two fundamental disadvan-
11 q .enqueue(p, m) tages. 1) It has to compute candidate paths for n(n − 1)/2 pairs
because it does not know the runtime S-D pairs. 2) The candidate
12 while q is not empty do
paths exhibit a low utilization rate due to severe resource contention
13 p ← q .dequeue()
among them. Q-CAST does not require any offline computation
14 if W idt h(p, TQ ) < widt h[p] then
and always finds the paths if only paths exist in the residual graph.
// The width of p has changed
For example in Fig. 8(a), AB and A′B ′ are two S-D pairs. The of-
15 Update widt h[p] and re-insert p to q
fline algorithm of Q-PASS finds 25 paths for AB, passing by nodes
16 continue
M 1, · · · , M 25 . But a single path A′, M 1, · · · , M 25, B ′ takes all avail-
17 if W idt h(p, TQ ) = 0 then
able qubits on Mi , and thus in the residual graph, all 25 candidate
// Even the best path is unsatisfiable
paths of AB fail, though paths AA′B and AB ′B exist outside the
18 break
offline paths, which are correctly found and reserved by Q-CAST
19 L P ← L P + ⟨p, widt h[p]⟩ online algorithm as shown in Fig. 8(b). Due to unpredictable com-
20 for ⟨n1, n2⟩ ∈ edges of p do binations of S-D pairs and the resulting residual graphs, it is hard
21 TQ [n1] ← TQ [n1] − widt h[p] to pre-calculate and store the paths for all S-D pair combinations.
22 TQ [n2] ← TQ [n2] − widt h[p] 4.3.1 Algorithm overview. Q-CAST does not require any offline
23 LC ← LC + widt h[p] unbound channels on ⟨n1, n2⟩ computation and follows the four-phase model in Fig. 2. Q-CAST
24 par t ial ← L P + (q as List) P1 and P3 are standard procedures similar to those of Q-PASS. The
25 for p ∈ par t ial do inputs of Q-CAST P2 are the network topology and the S-D pairs. In
26 Update TQ and LC as line 21-23, only on available edges P2, Q-CAST selects major paths for each S-D pair, without resource
contention. Besides, contention-free recovery paths are also selected
in P2. P4 takes the major paths and recovery paths from P2 and the
We propose segment-based path recovery for P4. On each node, link states from P3 to compute the swapping decisions.
each major path given by P2 ⟨(v 0 , v 1 , · · · , vh ), W ⟩ is divided into 4.3.2 P2 Algorithm of Q-CAST. The core task for Q-CAST P2 is
⌈h/(k + 1)⌉ segments, each with width W : (v 0 , v 1 , · · · , vk +1 ), (vk +1 , to find multiple paths based on the knowledge of S-D pairs, and
v k +2 , · · · , v 2k +2 ), · · · , (v ⌈h/(k +1)−1⌉(k +1) , · · · , vh−1 , vh ). The length of the paths should be contention-free on qubits and channels. Yen’s
the segments is set to k + 1 such that each node knows the states algorithm [57] does not satisfy the requirements because its output
of all links on the segment containing it, via the k-hop link states paths are highly overlapped. Note, Q-PASS uses Yen’s algorithm to
received in P3. Then for segment (vi 0 , vi 1 , · · · , vi k +1 ), each node find offline paths because the resulting overlapped path naturally
on it finds paths connecting the two ends, vi 0 and vi k +1 , using provides small detours, serving as recovery paths for major paths.
successful links in the k-hop neighborhood. We propose to search multiple contention-free paths for online S-D
An example is shown in Fig. 7. Assume k = 1, and thus each pairs using a greedy algorithm, which runs as follows. Step 1) For
node knows the link states of its 1-hop neighbors. The major path every S-D pair, it uses the Extended Dijkstra’s algorithm (described
Concurrent Entanglement Routing for Quantum Networks: Model and Designs SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA

later) to find the best path in terms of the routing metric EXT Algorithm 2: The Extended Dijkstra’s algorithm
(Equation2) between this pair. Step 2) Among the best paths of all S- Input: G = ⟨V , E, C ⟩, e, ⟨sr c , dst ⟩
D pairs, it further selects the path with the highest EXT and reserve Output: The best path ⟨p, W ⟩
the resources (qubits and channels) of this path, and the network // Initialize empty states
topology is updated to the residual graph by removing the reserved 1 E ← an array of nelements, all set to − ∞
resources. Steps 1) and 2) are repeated with the residual graph, 2 pr ev ← an array of n elements, all set to null
until no more path can be found, or the number of paths exceeds 3 visit ed ← an array of n elements, all set to false
200 ś a value limiting the number of paths to avoid unnecessary 4 widt h ← an array of n elements, all set to 0
computation. We call this algorithm as Greedy EDA (G-EDA). 5 q ← fibonacci-heap, highest E[·] first
The above process aims to maximize the network throughput // Initialize states of sr c
but does not consider fairness among S-D pairs. We will discuss 6 E[sr c] ← +∞
how to balance throughput and fairness in ğ 6, and this could also 7 widt h[sr c] ← +∞
be a future research topic. 8 q .enqueue(sr c)
The optimal routing metric. To find the optimal path under // Track the best path until dst
the EXT metric in a quantum network, the classical Dijkstra’s algo- 9 while q is not empty do
rithm fails because it only finds the shortest path when the routing // Get the current best end node
metric is ‘additive’. Here, additive means the sum of the costs of all 10 u ← q .dequeue()
edges on the path is exactly the cost of the whole path. Obviously, 11 if visit ed[u] then continue
the EXT Et computed by Equation 2 is not additive. We propose the 12 else visit ed[u] ← true
Extended Dijkstra’s algorithm (EDA) to find the best path between 13 if u = dst then
any S-D pair for any non-additive but monotonic routing metric. 14 ⟨p, W ⟩ ← Construct path via pr ev and widt h
The resulting path gives the maximum evaluation value among 15 return ⟨p, W ⟩
all possible paths between the S-D pair, with respect to a routing // Expand one hop based on u
metric function e. The input of e is a path ⟨p,W ⟩, and the output is 16 for v ∈ neighbors of u do
the path quality evaluation value. 17 if visit ed [v] then continue
Similar to the original Dijkstra algorithm, EDA also constructs an 18 ⟨p, W ⟩ ← Construct path via pr ev and widt h
optimal spanning tree rooted at the source node s. At the beginning, 19 E ′ ← e(p, W )
the visited set only includes s. The evaluation value from s to an 20 if E[v] < E ′ then
unvisited node x is set as 0 or the evaluation value e(s, x) of the 21 E[v] ← E
edge (s, x) if s and x are neighbors. Each time, the node y with the 22 pr ev[v] ← u
maximum evaluation value to s is added to the visited set and the 23 widt h[v] ← W
evaluation values from s to any other node x are updated if x and y 24 q .reorder(v)
are neighbors. The algorithm stops when the destination is visited.
The pseudocode of EDA is shown in Alg. 2. H
F G
We skip the proof of the correctness of EDA due to space limit. G H
I
F
Its correctness rely on a fact that the evaluation function e of a I
path ⟨p,W ⟩ should monotonically decrease when extending p to a A C D E B
longer path p ′ by adding another node at the end of p. Since we A C D E B (b) In P4, the major path disconnects at AC
and EB. Hence recovery paths AFC and EIB are
use Et as the evaluation function, we explain the monotonicity (a) In P2, The major path ACDEB and taken. G and H still swap along DGHB, but D
three recovery paths (in the order of swaps CD and DE, in favor of a smaller detour
of Et without a strict proof. As the W -path p grows, W may stay being founded): AFC, EIB, and DGHB from the major path for higher ETX
unchanged or decrease because the new edge may be narrower
Figure 9: Example of Q-CAST recovery algorithm
than W . In addition, adding one more hop means more hops to be
transmitted. Neither of the above can increase Et . ignored because it is unlikely to be a good path. The value of hm
Different from the original Dijkstra algorithm, updating the path can be determined at system initialization. For a new network G,
by adding one hop may cause a re-evaluation of the entire path, 100 pairs of nodes are randomly selected. Then, multipath routing
rather than simply adding the cost of a link. To avoid expensive is performed via G-EDA for each pair with hm = inf. Then hm
recalculation for path updates, one optimization can be applied is set to equal the largest hopcount of the resulting paths whose
when using Et as the evaluation function. If a (W , h)-path p grows Et > 1.
by one hop with width > W , then the width of the new path p ′ stays Recovery paths. After finding the major paths via G-EDA, the
unchanged to be W . Hence, in the calculation of Et (p ′ ) = qh+1 · remaining qubits and channels can be utilized to construct recovery
ÍW i i
i=1 i · Ph+1 , the original values Ph when calculating E t (p) can be
paths, each of which ends at two nodes (denote as switch nodes) on
re-used, which significantly reduce the complexity by performing a single major path. The switch nodes should be no more than k
just one iteration. hops away on a major path, where k is the link state range, because
Bound the path length. We set the upper-bound threshold in P4, the two nodes should ensure consistent swapping decisions.
hm for the hopcount of major paths to ensure bounded searching The recovery paths are found as following. For every node x on
in EDA. During EDA, any path with hopcount larger than hm is a major path, we use EDA to find 6 R recovery paths between x
and y in the residual graph, where y is the 1-hop ahead node on the
SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA Shouqian Shi and Chen Qian

F G H
F G H

A C D E B
A C D E B
(b) In P4, the major path disconnects at CD and
(a) In P2. Major path ACDEB and two EB. Recovery paths AFE and DGHB are taken,
recovery paths (in the order of being because 𝑨𝑪, 𝑪𝑫, 𝑫𝑬, 𝑬𝑩 ⨁ 𝑨𝑭, 𝑭𝑬, 𝑨𝑪, 𝑪𝑫, 𝑫𝑬
founded): AFE and DGHB ⨁ 𝑫𝑮, 𝑮𝑯, 𝑯𝑩, 𝑫𝑬, 𝑬𝑩 connects A and B.

Figure 10: Example of Q-CAST recovery via exclusive-or

major path, and R is a small constant parameter. When all nodes
are processed, the algorithm will iterates further for the recovery
Figure 11: Visualized network Figure 12: Visualized path selec-
paths that covers l hops on the major path, for l = 2, 3, · · · , k. In with qubits and channels tion and resource
Fig. 9(a), the major path is ACDEB and three recovery paths are
found. Cost of routing metric evaluation. The time cost to calculate
Every node will assign its qubits based on the reserved major Et for a (W , h)-path according to Equation 1 is O(hW ), and the
paths and recovery paths, without qubit/channel contention. space cost is O(W ).
4.3.3 P4 Algorithm of Q-CAST. In P4, each node knows the ma- Cost of Q-PASS P2. The space cost is O (mKm hm + n) and the
jor paths, the recovery paths, and the k-hop link states. It then time cost is O (mKm (hm + log(mKm ))).
makes the swapping decisions locally. The challenges for Q-PASS Cost of EDA. The space cost is O(n). The time cost for EDA is
P4 still present for Q-CAST P4: probabilistic link failures and non- O (n log n + |E |(hm Wm )).
interactive communication betweenÉ nodes. 5 PERFORMANCE EVALUATION
We propose an exclusive-or (xor, ) based algorithm to recover 5.1 Simulator Implementation
from potential link failures. We define the xor operator of two set We implement the proposed network models and algorithms on
É
of edges E 1, E 2 : E 1 E 2 = E 1 ∪ E 2 \ (E 1 ∩ E 2 ). As both ends (switch a custom-built time-based simulator, with additional supports for
nodes) of a recovery path pr are on a single major path, a segment the topology generation, statistics, and network visualization. We
of the major path pm is covered by pr , where pm and pr form a do not use packet-based simulation because quantum networks do
loop in the graph, called a recovery loop. Then, the link recovery not use packet switching. As shown in Figures 11 and 12, the visu-
algorithm works as following. The major path list is traversed alization tool shows the network topology, current qubit/channel
from beginning to end. Each visited major path p is treated as W occupation, and existing quantum links at simulation runtime, for
separated 1-paths, where W is the width of p. For each separated protocol analysis and demonstration. The source code repository
1-path, the set E collects the successful edges of it. K recovery paths of the simulator can be found on this link [1].
are found, and the edges of the recovery loops of the recovery We do not assume any specific topology and randomly generate
paths are collected as Ep1 , Ep2 , · · · , EpK , such that the S-D pair is quantum networks for simulations. We set the area A holding quan-
É É É É
connected in the graph ⟨V , E Ep1 Ep2 ··· EpK ⟩ , where tum networks is a 100K units by 100K units square, each unit may
V is the set of nodes on the major path and the K recovery paths. To be considered as 1km. The network generation algorithm requires
break the tie, shorter recovery paths are preferred because shorter three input parameters: the number of nodes n, the average number
paths are more likely to succeed after swapping. The Q-CAST of neighbors Ed , and the average success rate of all channels Ep .
recovery algorithm is different from that of Q-PASS because each Nodes are randomly placed and the distance of any two nodes is
recovery path in Q-CAST is dedicated to a single major path, and √
at least 6 50/ n units. The edges are generated by the Waxman
they are contention-free. model [55] that was used for Internet topologies [29].
As an example, in Fig. 9, the major path disconnects at AC and After the topology generation, a binary search on the model
EB. Nodes F, G, and H swap along the recovery path no matter the parameter α is further carried out to make the average channel
recovery path is used or not. As switch nodes, A and C recover the success rate to be Ep ± 0.01. The number of qubits Q for each node
broken edge AC by the recovery path AFC. Both D and E know is independently uniformly picked from 10 to 14. The edge width
the two recovery paths covering EB, namely DGHB and EI B. The W is independently uniformly generated from 3 to 7, for each edge.
shorter one EI B is used. D still swaps qubits on the major path and We pick the range for Q and W based on our conjecture of a well-
E switches to the recovery path. functioning quantum network. Our designs should work on wider
As another example, in Fig. 10, the major path disconnects at CD ranges, which we cannot cover due to enormous possibilities.
and EB. Recovery paths AF E and DGHB are taken, because the xor
of the major path ACDEB and two recovery loops AF EDCA and
5.2 Methodology
We evaluate the throughput, scalability, and fairness of the pro-
DGHBED connects A and B. Note the edge DE appears 3 times in
posed entanglement routing algorithms. To gain insight into the
the xor and is used łreverselyž on the final path.
performance metrics and to provide a reference for future research,
4.4 Time and space costs we show more simulation statistics: the resource efficiency towards
We denote the number of S-D pairs as m, and the maximum width high throughput, the contribution of recovery paths for both al-
of paths as Wm , which is determined by node capacities and edge gorithms. Each data shown in the section is the average from 10
widths. We denote the maximum number of paths as Km in EDA. different network topologies.
The number of nodes is n. We summarize the results here and some We let the number of nodes n vary in set {50, 100, 200, 400, 800},
details can be found in the Appendix. average channel success rate Ep vary in {0.6, 0.3, 0.1}, internal link
Concurrent Entanglement Routing for Quantum Networks: Model and Designs SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA

1.0 1.0 1.0

Throughput (eps)
SLMP SLMP 12
0.8 SLMP 0.8 Greedy 0.8 CR
BotCap Q-CAST Greedy 10
0.6 Greedy 0.6 0.6
CR Q-CAST 8
SumDist Q-CAST
0.4 MultiMetric 0.4 MultiMetric
0.4 MultiMetric
6 Q-PASS
SumDist SumDist
0.2 Q-CAST
0.2 BotCap 0.2 BotCap 4
Greedy
CR SLMP
0.0 0.0 0.0 2
0 5 10 15 20 25 0 5 10 0 5 10 15 20 25 0 3 6 ∞
Throughput (eps) Throughput (eps) Throughput (eps) Link state broadcast range k

Figure 13: CDF of throughput un- Figure 14: CDF of throughput, Figure 15: CDF of throughput, n = Figure 16: Throughput vs. LS
der the reference setting Ep = 0.3 400 sharing ranges
30 14
17.5 Q-CAST Q-CAST 12 Q-CAST

Throughput (eps)

Throughput (eps)
Q-CAST
Throughput (eps)
Throughput (eps)

25 12
15.0
Q-PASS Q-PASS Q-PASS
Q-PASS
10
20 Greedy Greedy 10 Greedy Greedy
12.5 8
15
SLMP SLMP
8 SLMP SLMP
10.0 6
6
10 7.5 4
4
5 5.0 2
2
0 2.5 0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.800
0.8250.850
0.875
0.900
0.925
0.950
0.975
1.000 50 100 200 400 800 2 4 6 8 10
Average channel success rate Swapping success rate |V| # S-D pairs in one time slot

Figure 17: Throughput vs. Figure 18: Throughput vs. Figure 19: Throughput vs. net- Figure 20: Throughput vs. # of
channel success rates swapping success rates work sizes S-D pairs

success rate q vary in {0.8, 0.9, 1.0}, link state range k vary in {0, unreliable recovery paths may be selected. This would occupy the
3, 6, ∞}, average degree Ed vary in {3, 4, 6}, and the number of routing resource which could have been allocated to other shorter
S-D pairs m vary from 1 to 10. To control variable, we show the and more reliable recovery paths.
results under the reference setting n = 100, Ep = 0.6, q = 0.9, Vary link success rates. Figures 17 and 18 show the average
k = 3, Ed = 6, m = 10, unless explicitly changed to observe the data throughput of Q-PASS, Q-CAST, Greedy, and SLMP on different
trend. For each setting of (n, Ep , q, k, Ed , m), 10 random networks quantum device abilities by varying the average channel success
are generated, and we simulate 1000 independent time slots on each rate and swapping (internal link) success rate. When the channel
of the networks. success rate p or the swapping success rate q is small, the overall
We compare Q-PASS and Q-CAST with two existing routing algo- throughput will be degraded. A robust routing algorithm should still
rithms that have been used in quantum network studies: single-link perform well on low ability networks. From the figures, the swap-
multipath routing (SLMP) [35] (a circuit-switching style protocol) ping success rate also has big impact on the average throughput,
and greedy routing [23] (a distributed protocol). because the link failure in the P2 can be mitigated by the recovery
algorithms in P4, but there is no circumvention for swapping errors.
5.3 Evaluation results And the Q-CAST performs the best among the four algorithms.
Throughput. Figures 13 to 15 show the CDF of throughputs for Scalability. We evaluate the scalability of routing algorithms
Q-PASS, Q-CAST, Greedy, and SLMP, under the reference setting. on two dimensions: the size of the network n and the number
The throughput results are calculated in terms of ebits per time slot of concurrent S-D pairs m. A larger network means the average
(eps). The BotCap, CR, and SumDist are the routing metrics for distance of S-D pairs is longer; and more concurrent S-D pairs in one
the Q-PASS, and they are shown separately for better comparison. time slot introduce higher level of resource contention. Figures 19
Despite the multipath routing, SLMP shows the lowest throughput and 20 show the average throughput on the two dimensions. All
because of the unreliability of a single channel/link. It fails to deliver algorithms exhibit a logarithmic throughput decrease with the
any ebits in >10 percent of the time slots, and for 90 percent of the number of nodes in the network. Q-CAST outperforms others on
time slots, the total throughput between 10 S-D pairs are less than 5. all network sizes, and the throughput of Q-CAST is as high as
The Greedy enjoys a high throughput, and for more than 90 percent 7.5eps when the network contains 800 nodes. The reason of lower
of the time, it delivers more than 15 ebits for 10 random S-D pairs. throughput in larger networks is because the average path length
For Q-PASS, all the three metrics of it exhibit similar throughput, is longer for the S-D pairs. Longer paths are more likely to fail
and the CR metric gives the highest throughput among all metrics, in quantum networks. Besides, the throughput of all algorithms
which delivers about 2 eps more than the Greedy. Q-CAST shows grow sub-linearly with the number of S-D pairs, due to resource
great advantages over all other algorithms and outperforms the CR contentions. Q-CAST outperforms others on most settings, and the
about 5 eps. Q-CAST is also the most reliable because it seldom advantage of Q-CAST over other algorithms grows rapidly with
delivers less than 5 eps. Since CR is slightly better than other metrics, the number of S-D pairs. It is because Q-CAST actively resolves the
we use CR to represent Q-PASS in the following results. resource contentions for the S-D pairs.
Vary link state range. In P3, each node shares its link states Fairness. Though we aim to maximize the throughput in the
with its k-hop neighbors, and hence, k influences the path recovery current designs, the fairness among the S-D pairs is evaluated.
performance. Fig. 16 shows the average throughput on different k. Fig. 21 shows the average number of successful S-D pairs under
The Greedy algorithm does not rely on k and is shown for reference. different numbers of concurrent requests. For a time slot, an S-
k contributes little to the overall performance because most path D pair is successful (epair) when they establish at least one ebits
failures are just one hop vi -vi+1 , which can be recovered by vi and after P4. Q-CAST outperforms others and all algorithms grow sub-
vi+1 with their own link states. k = 3 is sufficient for Q-CAST, and linearly. Fig. 22 shows the CDF of the number of paths allocated
larger k slightly degrades the throughput because longer and more
SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA Shouqian Shi and Chen Qian

5 1.0 1.0 30000

# occupied channels
Q-CAST
# succ S-D pairs

Q-PASS 0.8 0.8 25000

4 Greedy 20000 Q-CAST\R
3 SLMP 0.6 0.6

CDF

CDF
Q-PASS
SLMP Q-PASS\R
15000
Q-CAST
2 0.4 Greedy 0.4 Q-CAST 10000 Q-PASS\R

1 0.2 Q-CAST
0.2 Q-PASS
5000
Q-PASS Q-CAST\R
0.0 0.0 0
2 4 6 8 10 0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
# S-D pairs in one time slot Total width of allocated major paths Throughput (eps) Throughput (eps)

Figure 21: # of successful con- Figure 22: CDF of # of major paths Figure 23: Contribution of re- Figure 24: Overhead of recov-
current S-D pairs covery paths ery paths
1.0 1.0
to every S-D pair. A W -path is counted as W separate paths. As a 1 S-D pair 6 S-D pairs
|V| = 50
0.8 0.8
2 S-D pairs 7 S-D pairs
3 S-D pairs 8 S-D pairs |V| = 100
baseline requirement, any S-D pair should be allocated at least one 0.6
4 S-D pairs
5 S-D pairs
9 S-D pairs
10 S-D pairs 0.6
|V| = 200

CDF

CDF
|V| = 400
major path, which is fulfilled by all algorithms. The SLMP is the 0.4 0.4 |V| = 800

fairest. The Q-CAST has a turning point on the CDF figure, which 0.2 0.2

means 40 percent of S-D pairs are allocated less than 9 paths, and 0.0 0.0
0 1 2 3 4 5 6 0 25 50 75 100 125 150 175
the other pairs are allocated 10 to 14 paths, which is very fair. The Width of recovery path # recovery paths per major path

Q-PASS is the most biased algorithm. Figure 25: CDF of the width of Figure 26: CDF of # of recovery
Recovery paths. We evaluate the contribution of recovery recovery paths paths on a single major path
paths to the overall throughput for both Q-PASS and Q-CAST, by factor such as 2, and the related routing metric is multiplied with
comparing their throughput with that of their recovery path-free 1.12 in T + 2. Eventually, this pair will succeed.
versions Q-PAST\R and Q-CAST\R. The results are shown in Fig. 23. Prioritized routing. Both Q-PASS and Q-CAST are extendable
The recovery paths contribute about 0.5eps to Q-PASS and 1eps to to support simple prioritized routing. Suppose S-D pairs are in
Q-CAST. We further show the average number of occupied channels different priority classes, identified by the number 1, 2, · · · , 10, and
in one time slot for Q-CAST, Q-CAST\R, and Q-PASS in Fig. 24, the priority is ‘hard’ ś a single S-D pair in priority class c is far more
where the x-axis shows the throughput of each case. Q-PASS is valuable than all S-D pairs in priority class c − 1 and lower. In P2 of
not shown in this figure because it takes way more channels in Q-PASS and Q-CAST, the offline paths (only Q-PASS) and online
the recovery paths and the results are not in this range of y-axis. paths (both algorithms) of the highest priority S-D pair are selected
Q-PASS\R takes times fewer channels compared with Q-PASS, and until no more path is available. More paths are then selected in
Q-CAST\R saves 25% channels from the 400 channels taken by the residual graph for S-D pairs in lower priority classes. The P4
Q-CAST. of Q-CAST is not modified because the selected paths have no
As the recovery paths are contention-free for Q-CAST, more contention. In P4 of Q-PASS, the paths of the highest priority S-D
interesting statistics are collected on Q-CAST recovery paths. The pair are recovered first.
CDF of the width of recovery paths is shown in Fig. 25. The recovery
paths can be wider when the number of S-D pairs is small, because 7 CONCLUSION AND FUTURE WORK
of the low resource contention between S-D pairs. For most cases, This work presents a new entanglement routing model of quan-
the widths of recovery paths for a single S-D pair are larger than tum networks that reflects the differences compared to classical
those of the 10 concurrent S-D pairs by 2. Besides, the CDF of the networks and new entanglement routing algorithms that utilize the
total number of recovery paths of a single major path is shown in unique properties of quantum networks. The proposed algorithm
Fig. 26. In larger networks, the major paths are longer, and more Q-CAST increases network throughput by a big margin compared
recovery paths can be found. to other methods. We expect more future research will be conducted
Summary of evaluations. Q-CAST exhibits much higher through- on the entanglement routing problem and could contribute to the
put, robustness, and scalability than other routing algorithms. Q- eventual success of quantum networks.
PASS also shows good throughput and the metric CR provides the There could be a large amount of possible future work on the
highest throughput for Q-PASS. If the minimum resource utiliza- topic of routing in quantum networks. Among them, we identify
tion is a concern for some quantum networks, recovery paths for three possible future research topics that are directly related to this
both algorithms can be disabled for better efficiency. Q-CAST\R is work: 1) properly find offline paths such that generated paths are
a good balance between throughput and resource efficiency. resilient to the runtime resource contention, 2) find an efficient
algorithm to correctly select the recovery loops for Q-CAST P4,
6 DISCUSSION and 3) make use of the entangled but not used pairs in the previous
Better fairness. The algorithms proposed in this paper aim to
time slot, rather than resetting the whole network at the beginning
maximize throughput, and each time slot is considered totally sepa-
of every time slot.
rately. A simple extension, however, is available to both Q-PASS and
This work does not raise any ethical issues.
Q-CAST to provide better fairness while maintaining high through-
put. For any S-D pair that has failed to share an ebit in a slot T , the 8 ACKNOWLEDGEMENT
pair and the failing streak ⟨(s, d), 1⟩ are broadcast to all nodes in P1 This work was partially supported by National Science Founda-
in the slot T + 1. The routing metric of all paths connecting this tion Grants 1717948, 1750704, and 1932447. We thank Liang Jiang,
S-D pair is multiplied with a factor such as 1.1, which means their Rui Li, Peter Young, our shepherd Hongqiang Harry Liu, and the
paths are slightly over-evaluated, and thus are more likely to be anonymous reviewers for their suggestions and comments.
selected. If the pair still fails, the failing streak increases to a higher
Concurrent Entanglement Routing for Quantum Networks: Model and Designs SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA

REFERENCES Modeling, Analysis and Simulation of Computer and Telecommunications Systems.

[1] 2019. Source Code of the Quantum Routing Simulations. https://github . com/ [30] Rodney Van Meter, Takahiko Satoh, Thaddeus D. Ladd, William J. Munro, and
QianLabUCSC/QuantumRouting. Kae Nemoto. 2013. Path Selection for Quantum Repeater Networks. Networking
[2] Ian F Akyildiz, Weilian Su, Yogesh Sankarasubramaniam, and Erdal Cayirci. 2002. Science (2013).
Wireless sensor networks: a survey. Computer Networks (2002). [31] David L Moehring, Peter Maunz, Steve Olmschenk, Kelly C Younge, Dzmitry N
[3] Mohammad Al-Fares, Sivasankar Radhakrishnan, Barath Raghavan, Nelson Matsukevich, L-M Duan, and Christopher Monroe. 2007. Entanglement of single-
Huang, and Amin Vahdat. 2010. Hedera: dynamic flow scheduling for data atom quantum bits at a distance. Nature (2007).
center networks. In Proceedings of USENIX NSDI. [32] J. Moy. 1998. OSPF Version 2. RFC 2328.
[4] James Aspnes, Yossi Azar, Amos Fiat, Serge Plotkin, and Orli Waarts. 1993. On- [33] S Olmschenk, DN Matsukevich, P Maunz, D Hayes, L-M Duan, and C Monroe.
line load balancing with applications to machine scheduling and virtual circuit 2009. Quantum teleportation between distant matter qubits. Science (2009).
routing. In Proceedings of the twenty-fifth annual ACM symposium on Theory of [34] Jian-Wei Pan, Dik Bouwmeester, Harald Weinfurter, and Anton Zeilinger. 1998.
computing. Experimental entanglement swapping: entangling photons that never interacted.
[5] Charles H Bennett and Gilles Brassard. 1984. Quantum Cryptography: Public Physical Review Letters (1998).
Key Distribution and Coin Tossing. In Proceedings of the International Conference [35] Mihir Pant, Hari Krovi, Don Towsley, Leandros Tassiulas, Liang Jiang, Prith-
on Computers, Systems and Signal Processing. wish Basu, Dirk Englund, and Saikat Guha. 2019. Routing Entanglement in the
[6] Hannes Bernien, Bas Hensen, Wolfgang Pfaff, Gerwin Koolstra, Machiel S Blok, Quantum Internet. npj Quantum Information (2019).
Lucio Robledo, TH Taminiau, Matthew Markham, Daniel J Twitchen, Lilian [36] James L Park. 1970. The concept of transition in quantum mechanics. Foundations
Childress, and R. Hanson. 2013. Heralded entanglement between solid-state of Physics (1970).
qubits separated by three metres. Nature (2013). [37] Momtchil Peev, Christoph Pacher, Romain Alléaume, Claudio Barreiro, Jan Bouda,
[7] Anne Broadbent, Joseph Fitzsimons, and Elham Kashefi. 2009. Universal blind W Boxleitner, Thierry Debuisschert, Eleni Diamanti, M Dianati, JF Dynes, S Fasel,
quantum computation. In 2009 50th Annual IEEE Symposium on Foundations of S Fossier, M Fürst, J-D Gautier, O Gay, N Gisin, P Grangier, A Happe, Y Hasani,
Computer Science. M Hentschel, H Hübel, G Humer, T Länger, M Legré, R Lieger, J Lodewyck, T
[8] Angela Sara Cacciapuoti, Marcello Caleffi, Francesco Tafuri, Francesco Saverio Lorünser, N Lütkenhaus, A Marhold, T Matyus, O Maurhart, L Monat, S Nauerth,
Cataliotti, Stefano Gherardini, and Giuseppe Bianchi. 2019. Quantum internet: J-B Page, A Poppe, E Querasser, G Ribordy, S Robyr, L Salvail, A W Sharpe,
networking challenges in distributed quantum computing. IEEE Network (2019). A J Shields, D Stucki, M Suda, C Tamas, T Themel, R T Thew, Y Thoma, A
[9] M. Caleffi. 2017. Optimal Routing for Quantum Networks. IEEE Access (2017). Treiber, P Trinkler, R Tualle-Brouri, F Vannel, N Walenta, H Weier, H Weinfurter,
[10] Marcello Caleffi, Angela Sara Cacciapuoti, and Giuseppe Bianchi. 2018. Quan- I Wimberger, Z L Yuan, H Zbinden, and A Zeilinger. 2009. The SECOQC quantum
tum Internet: From Communication to Distributed Computing! (Proceedings of key distribution network in Vienna. New Journal of Physics (2009).
NANOCOM ’18). [38] Charles E. Perkins and Elizabeth M. Royer. 1999. Ad-hoc On-demand Distance
[11] Lilian Childress and Ronald Hanson. 2013. Diamond NV centers for quantum Vector Routing. In IEEE WORKSHOP ON MOBILE COMPUTING SYSTEMS AND
computing and quantum networks. MRS bulletin 38, 2 (2013), 134ś138. APPLICATIONS. 90ś100.
[12] Richard Cole, Bruce M Maggs, Friedhelm Meyer auf der Heide, Michael Mitzen- [39] Stefano Pirandola. 2019. End-to-end capacities of a quantum communication
macher, Andréa W Richa, Klaus Schröder, Ramesh K Sitaraman, and Berthold network. Commun. Phys 2 (2019), 51.
Vöcking. 1998. Randomized protocols for low-congestion circuit routing in mul- [40] Stefano Pirandola, Ulrik L Andersen, Leonardo Banchi, Mario Berta, Darius
tistage interconnection networks. In Proceedings of the thirtieth annual ACM Bunandar, Roger Colbeck, Dirk Englund, Tobias Gehring, Cosmo Lupo, Carlo
symposium on Theory of computing. Ottaviani, J. Pereira, M. Razavi, J. S. Shaari, M. Tomamichel, V. C. Usenko, G.
[13] Axel Dahlberg, Matthew Skrzypczyk, Tim Coopmans, Leon Wubben, Filip Roz- Vallone, P. Villoresi, and P. Wallden. 2019. Advances in Quantum Cryptography.
pedek, Matteo Pompili, Arian Stolk, Przemyslaw Pawelczak, Robert Knegjens, arXiv:1906.01645 [quant-ph]
Julio de Oliveira Filho, Ronald Hanson, and Stephanie Wehner. 2019. A link layer [41] Stefano Pirandola, Raul García-Patrón, Samuel L Braunstein, and Seth Lloyd.
protocol for quantum networks. In Proceedings of ACM SIGCOMM. 2009. Direct and reverse secret-key capacities of a quantum channel. Physical
[14] S. Das, S. Khatri, and J. P. Dowling. 2018. Robust quantum network architectures review letters 102, 5 (2009), 050503.
and topologies for entanglement distribution. Phys. Rev. A (2018). [42] Stefano Pirandola, Riccardo Laurenza, Carlo Ottaviani, and Leonardo Banchi.
[15] Vasil S Denchev and Gopal Pandurangan. 2008. Distributed quantum computing: 2017. Fundamental limits of repeaterless quantum communications. Nature
A new frontier in distributed systems or science fiction? ACM SIGACT News communications (2017).
(2008). [43] Chen Qian and Simon Lam. 2011. Greedy distance vector routing. In Proc. of IEEE
[16] Whitfield Diffie and Martin Hellman. 1976. New directions in cryptography. IEEE ICDCS.
transactions on Information Theory (1976). [44] Mark Riebe, H Häffner, CF Roos, W Hänsel, J Benhelm, GPT Lancaster, TW
[17] Artur K Ekert. 1991. Quantum cryptography based on Bell’s theorem. Physical Körber, C Becher, F Schmidt-Kaler, DFV James, et al. 2004. Deterministic quantum
review letters (1991). teleportation with atoms. Nature 429, 6993 (2004).
[18] Chip Elliot. 2002. Building the quantum network. New Journal of Physics (2002). [45] Ronald L Rivest, Adi Shamir, and Leonard Adleman. 1978. A method for obtaining
[19] Xiaoyan Hong, Kaixin Xu, and Mario Gerla. 2002. Scalable routing protocols for digital signatures and public-key cryptosystems. Commun. ACM (1978).
mobile ad hoc networks. IEEE network (2002). [46] Reza Rooholamini, Vladimir Cherkassky, and Mark Garver. 1997. Finding the
[20] Peter C. Humphreys, Norbert Kalb, Jaco P. J. Morits, Raymond N. Schouten, Right ATM Switch for the Market. IEEE Computer (1997).
Raymond F. L. Vermeulen, Daniel J. Twitchen, Matthew Markham, and Ronald [47] M. Sasaki, M. Fujiwara, H. Ishizuka, W. Klaus, K. Wakui, M. Takeoka, S. Miki, T.
Hanson. 2018. Deterministic delivery of remote entanglement on a quantum Yamashita, Z. Wang, A. Tanaka, K. Yoshino, Y. Nambu, S. Takahashi, A. Tajima,
network. Nature (2018). A. Tomita, T. Domeki, T. Hasegawa, Y. Sakai, H. Kobayashi, T. Asai, K. Shimizu, T.
[21] David B. Johnson and David A. Maltz. 1996. Dynamic Source Routing in Ad Hoc Tokura, T. Tsurumaru, M. Matsui, T. Honjo, K. Tamaki, H. Takesue, Y. Tokura, J. F.
Wireless Networks. In Mobile Computing. Kluwer Academic Publishers, 153ś181. Dynes, A. R. Dixon, A. W. Sharpe, Z. L. Yuan, A. J. Shields, S. Uchikoga, M. Legré,
[22] Brad Karp and Hsiang-Tsung Kung. 2000. GPSR: Greedy perimeter stateless S. Robyr, P. Trinkler, L. Monat, J.-B. Page, G. Ribordy, A. Poppe, A. Allacher, O.
routing for wireless networks. In Proceedings of the 6th annual international Maurhart, T. Länger, M. Peev, and A. Zeilinger. 2012. Field test of quantum key
conference on Mobile computing and networking. distribution in the Tokyo QKD Network. Optics Express (2012).
[23] Axel Dahlberg Kaushik Chakraborty, Filip Rozpedek and Stephanie Wehner. 2019. [48] Eddie Schoute, Laura Mancinska, Tanvirul Islam, Iordanis Kerenidis, and
Distributed Routing in a Quantum Internet. arXiv:1907.11630 (2019). Stephanie Wehner. 2016. Shortcuts to quantum network routing. arXiv preprint
[24] H Jeff Kimble. 2008. The quantum internet. Nature (2008). arXiv:1610.05238 (2016).
[25] Peter Komar, Eric M Kessler, Michael Bishof, Liang Jiang, Anders S Sùrensen, Jun [49] Peter W Shor. 1994. Algorithms for quantum computation: discrete logarithms
Ye, and Mikhail D Lukin. 2014. A quantum network of clocks. Nature Physics and factoring. In Proceedings 35th annual symposium on foundations of computer
(2014). science.
[26] Simon S. Lam and Chen Qian. 2011. Geographic Routing in d -dimensional Spaces [50] A. Singla, C.-Y. Hong, L. Popa, and P. B. Godfrey. 2012. Jellyfish: Networking
with Guaranteed Delivery and Low Stretch. In Proceedings of ACM SIGMETRICS. Data Centers Randomly. In Proc. of USENIX NSDI.
[27] Yuan Lee, Eric Bersin, Axel Dahlberg, Stephanie Wehner, and Dirk Englund. [51] Pirandola Stefano and Braunstein Samuel Leon. 2016. Unite to build a quantum
2020. A Quantum Router Architecture for High-Fidelity Entanglement Flows in internet. Nature 532 (2016), 169ś171.
Multi-User Quantum Networks. arXiv:2005.01852 [quant-ph] [52] Masahiro Takeoka, Saikat Guha, and Mark M Wilde. 2014. Fundamental rate-loss
[28] David Luong, Liang Jiang, Jungsang Kim, and Norbert Lütkenhaus. 2016. Over- tradeoff for optical quantum key distribution. Nature communications (2014).
coming lossy channel bounds using a single quantum repeater node. Applied [53] Rodney Van Meter and Joe Touch. 2013. Designing quantum repeater networks.
Physics B 122, 4 (Apr 2016). https://doi . org/10 . 1007/s00340-016-6373-4 IEEE Communications Magazine (2013).
[29] Alberto Medina, Anukool Lakhina, Ibrahim Matta, , and John Byers. 2001. BRITE: [54] Gayane Vardoyan, Saikat Guha, Philippe Nain, and Don Towsley. 2019. On the
An Approach to Universal Topology Generation. In International Workshop on Stochastic Analysis of a Quantum Entanglement Switch. In ACM SIGMETRICS
Performance Evaluation Review.
SIGCOMM ’20, August 10ś14, 2020, Virtual Event, NY, USA Shouqian Shi and Chen Qian

[55] Bernard M Waxman. 1988. Routing of multipoint connections. IEEE journal on path with W = 2, which can be found via G-EDA. The max-flow
selected areas in communications (1988). algorithm, however, gives the green path, the blue path, and the
[56] Stephanie Wehner, David Elkouss, and Ronald Hanson. 2018. Quantum internet:
A vision for the road ahead. Science (2018). red path ś all paths are single ś which is the sub-optimal solution.
[57] Jin Y Yen. 1971. Finding the k -shortest loopless paths in a network. Manegement Example 3. Due to the enormous search space, we failed to find
Science (1971).
[58] Juan Yin, Yuan Cao, Yu-Huai Li, Sheng-Kai Liao, Zhang, Ji-Gang Ren, Wen-Qi Cai,
the optimal strategy via brute-force even in a 10-node network. Sup-
Wei-Yue Liu, Bo Li, Hui Dai, Guang-Bing Li, Qi-Ming Lu, Yun-Hong Gong, Yu Xu, pose |V | = 10, every node has 15 qubits and 6 edges, and each edge
Shuang-Lin Li, Feng-Zhi Li, Ya-Yun Yin, Zi-Qing Jiang, Ming Li, Jian-Jun Jia, Ge is composed of 5 quantum channels. In the brute-force searching,
Ren, Dong He, Yi-Lin Zhou, Xiao-Xiang Zhang, Na Wang, Xiang Chang, Zhen-
Cai Zhu, Nai-Le Liu, Yu-Ao Chen, Chao-Yang Lu, Rong Shu, Cheng-Zhi Peng, we do not assume the P2 and P4 are carried out based on ‘paths’,
Jian-Yu Wang, and Jian-Wei Pan. 2017. Satellite-based entanglement distribution but just try all possible assignments of qubits to channels, perform
over 1200 kilometers. Science (2017). the swapping, calculate the Et between the given S-D pair, and
[59] Ye Yu and Chen Qian. 2014. Space Shuffle: A Scalable, Flexible, and High-
Bandwidth Data Center Network. In Proceedings of IEEE ICNP. record the highest result. For any S-D pair, the search space for
P2 is ∼ 2.3 · 10364 . Even worse, the entanglement swapping in P4
A APPENDIX depends on local states, which is prohibitively hard to enumerate
all possible swapping combinations.
Appendices are supporting material that has not been peer-reviewed.

A.1 Finding the optimal path selection for A.2 Time and space cost analysis
To avoid unbound computation and space cost in P2, we set the
Q-CAST
maximum number of multipath Km = 200. We set the maximum
path hopcount according to the network itself. For any input G,
C D
100 S-D pairs are randomly selected, and then multipath routing is
performed via G-EDA between each S-D pair. The largest hopcount
A B of selected paths whose Et > 1 is the maximum hopcount hm of all
selected paths. We denote the number of nodes as n, the number of
S-D pairs as m, and the maximum width of paths as Wm , which is
determined by node capacities and edge widths.
E F A.2.1 Cost of routing metric evaluation. The calculation of Et can
Figure 27: Counterexample for two possible algorithms be performed by following the recursive formula set 1. For an h-hop
We summarize the hardness of the contention-free path selection path with width W , the calculation of Et goes as following. Iterate
problem without classifying it into a certain complexity class, and on k, from 1 to h and further iterate on i, from W to 1: calculate
show its hardness in three examples. On one hand, because of Q ki , W Ql , P i , W P l , and Pki . Five W -element arrays
Í Í
l =i k k −1 l =i+1 k −1
the resource constraints (qubits/channels), path selection depends are allocated to store the values. After that, Et = qh · W
Í i
i=1 i · Ph is
highly on the link states and hence the search space is much more calculated in W + h time. Hence, the time cost is O(hW ), and the
than the classical algorithms which only depends on the weighted space cost is O(W ).
graph while edges and nodes have unlimited capacity; on the other
hand, Et is non-linear, which invalids many existing proofs based A.2.2 Cost of P2 algorithm of Q-PASS. The initialization costs O(n)
on the linear additivity of the routing metric and thus degrades the time. The double-for loop costs O(mKm (hm + hm + log(mKm )))
efficiency of classical algorithms. time. The while loop costs O(mKm (hm + log(mKm ) + hm )) time.
Example 1. Despite its good performance (shown in ğ 5), we Hence, the overall time cost is
prove G-EDA is not the optimal. An example graph is shown in O(mKm (hm + log(mKm ))).
Fig. 27 3 . Suppose all the edges have width 3, all channels have Each of the LC , L P , and q costs O(mKm hm ) space, the TQ costs
creation rate p = 0.99, the swapping success rate q = 1, s and d O(n) space, the width costs O(mKm ) space. Hence, the overall space
have qubit capacity 6, and all other nodes have capacity 3. Then cost is O(mKm hm + n).
the optimal contention-free paths are the blue path plus the green A.2.3 Cost of EDA. For a classical network ⟨V , E⟩, the Dijkstra’s
path. But the G-EDA will output only the red path. The reason of algorithm costs O(n log n + |E|) time because the dequeue operation
the failure of G-EDA is it falls in a local minimum and fails to give costs O(log n) time, the reorder operation of the Fibonacci heap
the max-flow ś the width of the red path is 3, as opposed to 6 for costs O(1) time, and each edge is visited at most once. Similarly, the
the blue path plus the green path. time cost for EDA is O(n log n +hmWm + |E|(hmWm )) = O(n log n +
Example 2. Though the classical max-flow algorithm gives the |E|(hmWm )). The space cost is O(n).
optimal solution, it performs worse than G-EDA in some other cases.
Consider the same topology in Fig. 27 with changed parameters.
Suppose all blue and green edges have width 1, red edges have width
2, all channels have creation rate p = 0.6, the swapping success
rate q = 1, s and d have qubit capacity 3, and all other nodes have
capacity 2. From Fig. 6, we know when p = 0.6, one (2, 3)-path is
4 This
better than three (1, 4)-paths. Hence, the optimal solution is the red number is got via a recursive algorithm instead of mathematical derivation.
Consider the number of unique combinations of 15 indistinguishable balls put into 6
3 Red path: (s , A, B, d). Green path: (s, C , A, E, d ). Blue path: (s , D, B, F , d). different buckets, each with capacity 5.