On Optimal Deadlock Detection Scheduling

Yibei Ling

, Shigang Chen

, Cho-Yu Jason Chiang

Applied Research Laboratories, Telcordia Technologies,

lingy,[email protected]

Department of Computer & Information Science & Engineering

University of Florida
[email protected]
February 13, 2006
Abstract
Deadlock detection scheduling is an important, yet often overlooked problem that
can signicantly aect the overall performance of deadlock handling. Excessive initia-
tion of deadlock detection increases overall message usage, resulting in degraded system
performance in the absence of deadlocks; while insucient initiation of deadlock detec-
tion increases the deadlock persistence time, resulting in an increased deadlock resolu-
tion cost in the presence of deadlocks. The investigation of this performance tradeo,
however, is missing in the literature. This paper studies the impact of deadlock de-
tection scheduling on the overall performance of deadlock handling. In particular, we
show that there exists an optimal deadlock detection frequency that yields the mini-
mum long-run mean average cost, which is determined by the message complexities of
the deadlock detection and resolution algorithms being used, as well as the rate of dead-
lock formation, denoted as . For the best known deadlock detection and resolution
algorithms, we show that the asymptotically optimal frequency of deadlock detection
scheduling that minimizes the overall message overhead is O((n)
1/3
), when the total
number n of processes is suciently large. Furthermore, we show that in general fully
distributed (uncoordinated) deadlock detection scheduling cannot be performed as ef-
ciently as centralized (coordinated) deadlock detection scheduling.
1
Keywords: Deadlock detection scheduling, Deadlock formation rate, Deadlock persistence
time
1
The material in this paper was presented in part at the Twenty-Fourth Annual ACM SIGACT-SIGOPS
Symposium on Principles of Distributed Computing, Las Vegas, Nevada, July 17-20, 2005
1
1 Introduction
The distributed deadlock problem [8, 20, 16, 26, 11, 14] arises from resource contention intro-
duced by concurrent processes in distributed computational environments. It has received
a great deal of attention in dierent areas such as distributed computing theory [22, 26, 9],
distributed database [17, 14, 8, 10, 11], and parallel and distributed simulation [2, 28, 21]. A
deadlock is a persistent and circular-wait condition, where each process involved in a dead-
lock waits indenitely for resources held by other processes while holding resources needed
by others. As a result, none of the processes waiting for needed resources can continue
computation any further without obtaining the waited-for resources. A deadlock has an
adverse performance eect that osets the advantages of resource sharing and processing
concurrency.
There are three common strategies of dealing with the deadlock problem: deadlock pre-
vention, deadlock avoidance, and deadlock detection and resolution. It is a long-held con-
sensus that both deadlock prevention and deadlock avoidance strategies are conservative
and less feasible in handling the deadlock problem in general, whereas the deadlock de-
tection/resolution strategy is widely accepted as an optimistic and feasible solution to the
deadlock problem, because of its exclusion of the unrealistic assumption about resource allo-
cation requirements of the processes [10, 16, 26, 7, 27]. The central idea behind the deadlock
detection and resolution strategy is that it does not preclude the possibility of deadlock
occurring but leaves the burden of minimizing the adverse impact of deadlock to deadlock
detection and resolution mechanisms. Under this scheme, the presence of deadlocks is de-
tected by a periodic initiation of a deadlock detection algorithm and then resolved by a
2
deadlock resolution algorithm [31, 27, 7].
Despite signicant performance improvement in the past, deadlock detection remains a
costly operation [26, 11, 19]. It requires dynamical maintenance of wait-for-graph (WFG)
that reects the runtime wait-for dependency among distributed processes, and performs a
graph analysis to detect the presence of deadlocks. There is a substantial tradeo between
the cost of deadlock detection and that of deadlock resolution [26, 16, 23]. An initiation of
deadlock detection consumes runtime system and network resources which are basically pure
overheads when no deadlock is present [26, 19]. Excessive initiation of deadlock detection
would reduce the deadlock resolution cost but result in system performance degradation in
the absence of deadlock, while infrequent deadlock detection would be accompanied by the
increased deadlock size, resulting in an increased deadlock resolution cost in the presence of
deadlocks [23, 16, 15, 1]. It is evident that deadlock detection scheduling is one of key factors
aecting the overall system performance of deadlock handling. Nevertheless, to the best of
our knowledge, this subject is generally missing in the literature.
This paper investigates the optimal deadlock detection scheduling. We study how to best
schedule deadlock detections so as to minimize the long-run mean average cost of deadlock
handling. We formulate this problem by introducing a generic cost model (utility metric)
and use this cost model to establish a connection between deadlock detection and deadlock
resolution costs, in relation to the rate of deadlock formation. We show that there exists
a unique optimal deadlock detection frequency that yields the minimum long-run mean
average cost. Moreover, our result indicates that the asymptotically optimal frequency of
deadlock detection that minimizes the message overhead is O((n)
1/3
), when the number n
of processes in the system is suciently large. In addition, we prove that a fully distributed
3
(uncoordinated) detection scheduling can not be performed as eciently as its centralized
counterpart (coordinate scheduling).
The rest of this paper is organized as follows. Section 2 contains a brief summary of the
distributed deadlock detection and resolution algorithms. Section 3 gives the notions and
denitions. Section 4 provides the detailed mathematical analysis and proves the existence
and uniqueness of an optimal detection frequency. The determination of the optimal deadlock
detection frequency, its asymptotic relation with the number of processes in a distributed
system, and the impact of random detection scheduling upon the long-run mean average
cost of deadlock handling, are presented. In Section 5, the main contribution of this paper
is highlighted and the possible future work is discussed.
2 Background
In this section we provide a brief summary of worst-case analysis of existing distributed
detection algorithms of generalized deadlocks and deadlock resolution algorithms since some
results will be used later on. We also touch on Grays simulation model [8] as well as Masseys
formulation [20].
We restrict our discussion to distributed detection and resolution algorithms. The refer-
ences [10, 12, 13, 11, 14, 16] provide excellent gateways to the state of the art in this area
for the generalized resource request model. In the following, we give a brief summary of the
worst-case performance of the existing distributed detection algorithms.
Table 1 summarizes the worst-case complexities of distributed deadlock detection algo-
rithms [3, 30, 12, 14], where n is the total number of processes, e the number of edges, d the
diameter, and l the number of sink nodes of the WFG. The distributed detection algorithm
4
Criterion Bracha- Wang [30] Kshemkalyani & Singhal
Toueg [3] et al. [12] [14]
Phases 2 2 1 1
Delay 4d 3d + 1 2d + 2 2d
Message 4e 6e 4e 2n + 2l 2e
Table 1: Distributed Deadlock Detection Algorithms
for generalized deadlocks by Kshemkalyni and Singhal [14] is the clear winner among the
algorithms listed in Table 1. Their algorithm has achieved a message complexity of 2e and
a time complexity of 2d, which are believed to be optimal. Since e = n(n 1) and d = n
in the worst-case analysis, the worst-case message complexity and time complexity thus can
be written as 2n
2
and 2n, respectively.
Although deadlock detection and deadlock resolution are often discussed separately, the
latter is as important as the former [10, 26, 12, 7, 27, 32, 16]. The primary issue of deadlock
resolution [15, 16, 17] is to selectively abort a subset of processes involved in the deadlock so as
to minimize the overall abortion cost [19, 26, 27, 7]. This is often referred to as the minimum
abort set problem. These victim (aborted) processes must cancel all pending requests and
release all the acquired resources in order to avoid false deadlock detection and resolution
[26, 12, 7]. The abortion cost thus includes (1) the sending of cancel messages to those
resources, and (2) the sending of reply messages to all the waiting processes that are currently
being blocked for the resources held by the aborted processes. One noteworthy point is
that these waiting processes could be either transitively blocked or deadlocked processes.
To further reduce the abortion cost, checkpointing is sometimes introduced to prevent the
victim processes from being rolled back from scratch [18].
In addition, it is possible that more than two processes can independently detect the same
5
deadlock. If each process that detects a deadlock resolves it, then the deadlock resolution
will be highly inecient and will result in subsequent false deadlock detection and deadlock
resolution [26, 7, 13, 15]. Therefore, only one process should be selected for resolving a
deadlock, which in turn requires that the initiations of deadlock resolution algorithm in
dierent sites be coordinated. Such a coordination for safe deadlock resolution comes at an
additional communication cost in message exchange [7].
Generally, deadlock resolution cost is measured either in terms of time complexity [6, 17, 27],
or in terms of message complexity [15, 16, 7]. The complexity of resolution algorithms is
summarized in Table 2, where n is the total number of processes, m the number of processes
having the priorities greater than deadlocked processes, N
r
the number of resources, and n
D
the size of a deadlock. Note that the message complexities are not given in [17, 27].
Complexity Lin Terekhov & Mendivil
& Chen [17] & Camp [27] et al. [7]
Time O(n
3
D
) O(n
3
N
r
) O(mn
D
)
Message O(mn
2
D
)
Table 2: Distributed Deadlock Resolution Algorithms
By transforming the problem of deadlock resolution into a minimum vertex cut problem,
Lin & Chens algorithm [5] can identify an optimal set of victim processes to be aborted, with
the properly selected abortion cost to avoid the starvation and livelock problems. The main
feature of Terekhov & Camps algorithm is to take the number of resources into account. The
deadlock resolution algorithm proposed by Mendivil et al. [7] uses a probe-based approach,
with a focus on the safety aspect of deadlock resolution. The novelty of this algorithm
is to use an additional round of message exchanges to gather the information needed for
ecient resolution after deadlocks are detected. The algorithm uses special message known
6
as probes to travel in the opposite direction of the edges in AWFG (asynchronous wait-
for graph), and then chooses the lowest priority process of each detected cycle as a victim
process to be aborted, hence avoiding the livelock and starvation problems. This deadlock
resolution algorithm [7] excels in the use of formal methods to prove the correctness and in
its ne-granular analysis of the algorithm complexities. In particular, its message complexity
is of O(mn
2
D
). The worst-case message complexity can also be written as O(n
3
) because the
eventual deadlock size, n
D
, is bounded by the total number of processes in the distributed
system, that is, m = O(n) and n
D
= O(n).
The past research has been primarily aimed at minimizing the complexities (costs) of
the deadlock detection and resolution algorithms. Although deadlock detection scheduling
(particularly how frequently deadlock detection should be performed) has signicant impact
on the overall performance of deadlock handling in practice, it is not explicitly studied but
rather implicitly reected in the description of deadlock detection algorithms, without a clear
guideline. For instance, in [10, 26, 14, 16, 19, 4, 10, 5], the authors stated that a deadlock
detection is initiated when a deadlock is suspected. Other works [23, 11] suggested that it
would be highly inecient if deadlock detection is performed whenever a process/transaction
becomes blocked.
The performance of deadlock handling not only depends on the per-detection cost of the
deadlock detection algorithm, but also on how frequently the deadlock detection algorithm is
executed [11, 23, 19]. The choice of deadlock detection frequency presents a tradeo between
deadlock detection cost and deadlock resolution cost [10, 26, 23, 16, 11]. Park et al. [23]
pointed out that the reduction of deadlock resolution cost can be achieved at the expense
of deadlock detection cost. Krivokapie et al. [11] showed in their simulation study that the
7
path-pushing algorithm (one type of deadlock detection algorithm) is highly sensitive to the
frequency of deadlock detection. Gray et al. [8] showed that the probability of a transaction
waiting for a lock request is rare. They used a straw-man analysis in their simulation
model that agreed well with the observation on several data management systems. Massey
[20] formulated a probabilistic model that gave an analytic justication for the simulation
results reported in [8], showing that the probability of deadlock grows linearly with respect to
the number of transactions and grows in the fourth power of the average number of resources
required by transactions.
To our best knowledge, only a few papers [8, 16, 27, 5, 26, 19, 6] mentioned about deadlock
detection scheduling but under a dierent context from this paper. The idea of relating
deadlock recovery cost to deadlock persistence time, and identifying an optimal deadlock
detection frequency that minimizes the long-run mean average cost from the perspective of
deadlock handling, has not been considered before.
3 Deadlock Persistence time and Deadlock Recovery
Cost
In this section, we rst give the following denitions in order to simplify problem formulation.
Denition 1 A deadlock refers to a circular-wait condition where a set of processes waits
indenitely for resource from each other. A blocked process (a process in a deadlock) refers
to the process that waits indenitely on other processes to progress. Deadlock size refers to
the total number of blocked processes involved in the deadlock.
Blocked processes can be decomposed into two categories: deadlocked and transitively blocked
processes [16]. Deadlocked processes belong to a cycle in the WFG, while a transitively
8
blocked process refers to one that waits for the resources held by other processes but does
not belong to any cycle in the WFG.
Denition 2 Two deadlocks are said to be independent of each other if they dont share any
deadlocked process.
The independence of deadlock occurrence can be justied by the wide acceptance of
large-scale distributed systems and adoption of ne-granularity locking mechanism such as
semantic locking [24, 11] and record-granularity locking [24]. After decades of research and
development, large-scale distributed systems allow resource sharing among hundreds or even
thousands of sites across a network [24, 11]. The ne-granular locking mechanisms enable a
higher degree of parallelism. Large-scale resource distribution and ne-granularity of locking
make deadlocks likely to form independently.
Now we are in a position to introduce the notion of deadlock persistence time which serves
as a basis for our problem formulation. Let S = S
1
, S
2
, be the time instants at which
independent deadlocks initially occur, i.e., the ith deadlock forms at time S
i
.
Denition 3 The persistence time of the ith deadlock with respect to time t, denoted by
t
p
(t, S
i
), is
t
p
(t, S
i
) =
_
t S
i
, t > S
i
;
0, t S
i
The function t
p
(t, S
i
) represents the time interval between the present time and the time at
which the deadlock is initially formed. It grows linearly until the deadlock is resolved. The
notion of deadlock persistence time in spirit is similar to that of deadlock latency or deadlock
duration in [16, 15].
9
Once a deadlock is formed, other processes requesting resources currently held by the
blocked processes in the deadlock (including deadlocked and transitively blocked processes)
will be blocked forever unless the deadlock is resolved. As a result, each deadlock acts as
an attractor to trap more processes into it. As the deadlock persistence time increases, the
size of the deadlock (the total number of processes involved in the deadlock) keeps growing
[26, 9, 16, 15], which in turn increases the deadlock resolution cost.
(a) time = 1 (b) time = 2
x
x
y
z
y
z
Figure 1: Increasing Deadlock Size with Deadlock Persistence Time
This dependency of deadlock resolution cost upon deadlock persistence time can be illus-
trated in the example in Fig(1). At time=1, there are three circularly deadlocked processes
and two transitively blocked processes. At time=2, there are seven circularly deadlocked
processes. The graphs (a) and (b) in Fig(1) represent two snapshots in the wait-for graph,
showing that the deadlock size (including both deadlocked and transitively-blocked pro-
cesses) grows with the deadlock persistence time. Intuitively, a deadlock resolution algorithm
will have to explore the entire deadlock in order to identify the least costly set of victim pro-
cesses to be aborted. The intrinsic dependency of deadlock size (and thus deadlock resolution
cost) upon deadlock persistence time was observed by Singhal et al. [26, 13, 29], Lee [16, 15],
10
Krivokapic et al. [11], Lin et al. [17], and Park et al. [23].
Throughout this paper, we use n to denote the total number of processes in a distributed
system and n
D
(.) to denote the size of a deadlock. Consider an arbitrary deadlock. Its size
is a function of deadlock persistence time t
p
, denoted as n
D
(t
p
). The deadlock size n
D
(t
p
) by
nature is a discrete staircase function that jumps by one whenever a new process becomes
transitively blocked by the deadlocked processes. To facilitate our mathematical analysis,
we will treat n
D
(t
p
) instead as a continuous, increasing function, which is an approximation
of the staircase one.
The deadlock size function n
D
(t
p
) has the following mathematical properties. (1) n
D
(0) =
0, (2) monotonicity: n

D
(t
p
) > 0, t
p
0, and (3) bounded: n
D
() n, where n

D
(t
p
) is the
derivative of n
D
(t
p
). The rst property refers to the initial deadlock size at t
p
= 0 is zero.
The second property reects the fact that the number of blocked processes in the deadlock
increases monotonically with deadlock persistence time t
p
, and the third property indicates
that the eventual deadlock size is bounded by the total number of distributed processes. For
the sake of easy presentation, we drop the subscript p hereafter.
Now lets revisit the message complexity achieved by the deadlock resolution algorithm
proposed by Mendivil et al. [7], which is O(mn
2
D
) = O(nn
2
D
), where m is the number of
deadlocked processes having priority values greater than those of the deadlocked processes.
Notice that the deadlock size, n
D
, is a function of deadlock persistence time. To make this
dependency concrete, the message overhead can be written as cnn
2
D
(t) for some constant c.
This result will be used later to derive the optimal frequency of deadlock detection scheduling.
11
4 Mathematical Formulation
In this section, we begin with a generic cost model that accounts for both deadlock detection
and deadlock resolution, which is independent of deadlock detection/resolution algorithms
being used. We then prove the existence and the uniqueness of an optimal deadlock detection
frequency that minimizes the long-run mean average cost in terms of the message complexities
of the best known deadlock detection/resolution algorithms.
In this paper we choose the message complexity as the performance metric for measuring
the detection/resolution cost. The reason for choosing message complexity is that communi-
cation overhead is generally a dominant factor that aects the overall system performance in
a distributed system [26, 10, 13, 14], as compared with processing speed and storage space.
Note that the worst-case message complexity can normally be expressed as a polynomial
of n. Per deadlock detection cost is denoted as C
D
. The resolution cost for a deadlock
is denoted as C
R
(t), which is a function of the deadlock persistence time t. In general,
the resolution cost is a polynomial of n
D
(t). For example, the deadlock resolution cost for
Mendivils algorithm [7] is cnn
2
D
(t). Because n
D
(t) is a monotonically increasing function of
deadlock persistence time. C
R
(t) is also monotonically increasing with deadlock persistence
time. We assume that deadlock formation follows a Poisson process for two reasons: First,
the Poisson process is widely used to approximate a sequence of events that occur randomly
and independently. Second, it is due to mathematical tractability of the Poisson process,
which allows us to characterize the essential aspects of complicated processes while making
the problem analytically tractable.
The following theorem presents the long-run mean average cost of deadlock handling in
12
connection with the rate of deadlock formation and the frequency of deadlock detection.
Theorem 1 Suppose deadlock formation follows a Poisson process with rate . The long-run
mean average cost of deadlock handling, denoted by C(T), is
C(T) =
C
D
T
+

_
T
0
C
R
(t)dt
T
, (1)
where the frequency of deadlock detection scheduling is 1/T.
Proof: Let X
i
, i 1 be the interarrival times of independent deadlock formations, where
random variables X
i
, i 1 are independent and exponentially distributed with mean 1/.
Dene S
0
= 0 and S
n
=
n

i=1
X
i
, where S
n
represents the time instant at which the nth
independent deadlock occurs.
Let N(t) = supn : S
n
t represent the number of deadlock occurrences within the
time interval (0, t]. The long-run mean average cost is
lim
t
E(random cost in (0, t])
t
, (2)
where E is the expectation function. In order to associate this cost with the deadlock detec-
tion frequency (1/T), we partition the time interval (0, t] into non-overlapping subintervals of
length T. Let
k
(T) be the cost of deadlock handling on the subinterval ((k1)T, kT], k > 0.

k
(T) is a random variable. According to the stationary and independent increments of Pois-
son process [25], E(
i
(T)) = E(
j
(T)), i ,= j. The long-run mean average cost becomes
C(T) = lim
t
E(random cost in (0, t])
t
= lim
t
E(

t
T

k=0

k
(T))
t
= lim
t
E(
t
T
|
1
(T))
t
=
E(
1
(T))
T
, (3)
where x| is the oor function in x.
13
The cost (T) on interval (0, T] is the sum of a deadlock detection cost C
D
and a deadlock
resolution cost for those deadlocks independently formed within the interval (0, T]. For the
ith deadlock formed at time S
i
T, the resolution cost C
R
(T S
i
) is a function of the
deadlock persistence time T S
i
. Hence, the accrued total cost over (0, T] is
(T) = C
D
+
N(T)

i=1
C
R
(T S
i
)I
{N(T)>0}
, (4)
where I

is the indicator function whose value is 1 (or 0) if predicate is true (or false).
Among that, the deadlock resolution cost on interval (0, T] is
N(T)

i=1
C
R
(T S
i
)I
{N(T)>0}
=

i=1
C
R
(T S
i
)I
{S
i
T}
(5)
E
_
C
R
(T S
i
)I
{S
i
T}
_
=
T
_
0
C
R
(T t)f
i
(t)dt (6)
where f
i
(t) is the probability density function of S
i
which follows the gamma distribution
given below:
f
i
(t) =

i
(i 1)!
t
i1
e
t
, t > 0. (7)
Substituting Eq(7) into Eq(6) gives rise to
E
_
C
R
(T S
i
)I
{S
i
T}
_
=
_
T
0
C
R
(T t)

i
(i 1)!
t
i1
e
t
dt. (8)
The expected total resolution cost over the time interval (0, T] is
E(
N(T)

i=1
C
R
(T S
i
)I
{N(T)>0}
) =

i=1
_
T
0
C
R
(T t)

i
t
i1
(i 1)!
e
t
dt
=
_
T
0
C
R
(T t)e
t
_

i=1
(t)
i1
(i 1)!
_
dt =
_
T
0
C
R
(T t)dt =
_
T
0
C
R
(t)dt. (9)
14
Combining Eqs(3), (4), and (9) yields
C(T) =
E(
1
(T))
T
=
C
D
T
+

_
T
0
C
R
(T t)dt
T
=
C
D
T
+

_
T
0
C
R
(t)dt
T
. (10)
Theorem 1 is thus established.
Theorem 1 is mainly concerned with the impact of deadlock detection frequency and
deadlock formation rate on the long-run mean average cost of overall deadlock handling.
It is independent of the choice of deadlock detection/resolution algorithms. The following
corollary is an immediate consequence of Theorem 1.
Corollary 1 The long-run mean average cost of deadlock handling is proportional to the
rate of deadlock formation .
Proof: the proof is straightforward and thus omitted.
Theorem 1 and Corollary 1 state that the overall cost of deadlock handling is closely
associated not only with per-deadlock detection cost, and aggregated resolution cost, but also
with the rate of deadlock formation, . In the following lemma, we will show the existence and
uniqueness of asymptotic optimal frequency of deadlock detection when deadlock resolution
is more expensive than a deadlock detection in terms of message complexity.
Lemma 1 Suppose that the message complexity of deadlock detection is O(n

), and that of
deadlock resolution is O(n

). If < , there exists a unique deadlock detection frequency

1/T

that yields the minimum long-run mean average cost when n is suciently large.
Proof: Dierentiating Eq(1) yields
C

(T) =
C
D
T
2
+
C
R
(T)
T


_
T
0
C
R
(t)dt
T
2
. (11)
15
Dene a function (T) as follows
(T) T
2
C

(T) = C
D
+ TC
R
(T)
_
T
0
C
R
(t)dt. (12)
Notice that C

(T) and (T) share the same sign. Dierentiating (T), we have

(T) = TC

R
(T) (13)
Because C
R
(T) is a monotonically increasing function, C

R
(T) > 0, which means

(T) >
0. Therefore,

(T) is also a monotonically increasing function. C

R
(T) C
R
(t) 0 holds i
T t. For any given 0 < < T, it has
TC
R
(T)
_
T
0
C
R
(t)dt =
_
T
0
(C
R
(T) C
R
(t))dt >
_

0
(C
R
(T) C
R
(t))dt
>
_

0
(C
R
(T) C
R
())dt = (C
R
(T) C
R
()). (14)
Applying Eq(14) to Eq(12), we have
(T) = C
D
+ (TC
R
(T)
_
T
0
C
R
(t)dt) > C
D
+ (C
R
(T) C
R
()) (15)
We further have
(T) > C
D
+ C
R
(T)(1
C
R
()
C
R
(T)
) = C
D
+ C
R
(T) (16)
where = (1 C
R
()/C
R
(T)) and 0 < < 1 since C
R
(T) is monotonically increasing.
Substituting C
D
= c
1
n

and C
R
() = c
2
n

in Eq(16), we obtain
lim
T
(T) > c
1
n

+ c
2
n

(17)
Since < , lim
T
(T) is asymptotically dominated by the term c
2
n

when n is su-
ciently large. Observe that (0) = C
D
< 0, and (T) is monotonically increasing. By the
16
intermediate value theorem, it must be true that there exists a unique T

, 0 < T

< ,
such that
(T) = T
2
C

(T) =

< 0, 0 T < T

= 0, T = T

> 0, T > T

.
It means that C(T) reaches its minimum at and only at T = T

. The existence and the

uniqueness of optimal deadlock detection interval T

= arg
_
min
T>0
C(T)
_
is proved.
To make the idea behind this derivation concrete, we apply the up-to-date results of dead-
lock detection/resolution algorithms. As discussed before, the best-known message complex-
ity of a distributed deadlock detection algorithm is 2n
2
[14] when it is written as a polynomial
of n. The best-known message complexity of a deadlock resolution algorithm is O(nn
2
D
) [7].
Therefore, C
D
= n
2
, and C
R
(t) = cnn
2
D
(t), where c is a positive constant. Because the
deadlock size n
D
(t) is always bounded by n, from (15) we have
() = lim
T
(T) > C
D
+ (C
R
() C
R
()) 2n
2
+ cn
3
. (18)
Note that is a xed value that can be arbitrarily chosen. For a suciently large n, Eq(18)
becomes
() cn
3
> 0 (19)
(0) = C
D
= 2n
2
. Because (T) is monotonically increasing, there exists an optimal
deadlock detection frequency 1/T

such that (T

) and thus C

(T

) are zero, which mini-

mizes the long-run mean average cost C(T) for deadlock handling.
The motivation behind the proof is that the cost per deadlock detection is xed when
the total number of processes in the distributed system is given, while the cost of deadlock
resolution monotonically increases with deadlock persistence time. The resolution cost will
17
eventually outgrow the detection cost if deadlocks persist. As we set the time interval T
between any two consecutive detections longer, the detection cost becomes smaller due to
less frequent executions of the detection algorithm, but the resolution cost becomes larger
due to the growth in deadlock size. This implies that there exists a unique deadlock detection
frequency 1/T

that balances the two costs such that their sum is minimized. The condition
that the asymptotic deadlock resolution cost, C
R
(), is greater than the cost of deadlock
detection, C
D
, constitutes the natural mathematical basis to justify distributed deadlock
detection algorithms.
We are now ready to state the asymptotically optimal frequency for deadlock detection
based on the up-to-date results of distributed deadlock detection and resolution algorithms.
Recall that the best-known message complexity for distributed deadlock detection algorithms
is 2n
2
[14] and that for deadlock resolution algorithms of O(nn
2
D
) [7].
Theorem 2 Suppose the message complexity for distributed deadlock detection is 2n
2
, and
that for distributed deadlock resolution is O(nn
2
D
(t)). Then the asymptotically optimal fre-
quency for scheduling deadlock detections is O((n)
1/3
).
Proof: Assume that the deadlock size function n
D
(t) is both dierentiable and integrable.
2
Then n
D
(t) can be expressed in the form of Maclaurin series as follows:
n
D
(t) =

i=0
n
(i)
D
(0)t
i
i!
=

i=0
c
i
t
i
, (20)
where n
(i)
D
(0) denote the ith derivative of the deadlock size function n
D
(t) at point zero and
c
i
= n
(i)
D
(0)/i!.
2
Recall that n
D
(t) is a continuous approximation function whose curves between jumping points can
be chosen.
18
By the properties of the deadlock size function n
D
(t), we have n
D
(0) = 0 and n

D
(0) > 0.
It can be easily veried that c
0
= 0 and c
1
= n

D
(0) > 0. The resolution cost C
R
(t) can
be written as cnn
2
D
(t) for some constant c. By Theorem 1, the long-run mean average cost
becomes
C(T) =
2n
2
T
+ cn
_
T
0
n
2
D
(t)dt
T
. (21)
Inserting Eq(20) into Eq(21), we have
C(T) =
2n
2
T
+ cn
3
T
1
_
T
0
(

i=1
c
i
t
i
)
2
dt =
2n
2
T
+
cn
3
_
T
0
(c
1
t +

i=2
c
i
t
i
)
2
dt
T
. (22)
Through a lengthy calculation, Eq(22) can be simplied as
C(T) =
2n
2
T
+ cn
3
(
c
2
1
T
2
3
+
2c
1
c
2
T
3
4
) + cn
3
(

i=2

j=2
c
i
c
j
T
i+j
i + j + 1
). (23)
Taking derivative of Eq(23) with respect to T, we have
C

(T) =
2n
2
T
2
+ cn
3
(c
2
1
2T
3
+
3c
1
c
2
T
2
2
) + cn
3
(

i=2

j=2
c
i
c
j
(i + j)T
i+j1
i + j + 1
). (24)
By lemma 1, there exists a unique optimal detection frequency 1/T

when n is suciently
large, such that C(T

) C(T), T (0, ). We know that C

(T

) = 0. Based on (24), we
transform C

(T

) = 0 to the following equation.

1
n
=
c
2
(
2c
2
1
(T

)
3
3
+
3c
1
c
2
(T

)
4
2
+

i=2

j=2
c
i
c
j
(i + j)(T

)
i+j+1
i + j + 1
). (25)
Only n, T

, and are free variables and the rest are constants. By performing the Big-O
reduction we obtain
1
n
= (((T

)
3
+ (T

)
4
+ (T

)
5
+ ...)) (26)
19
When n is suciently large and T

is suciently small, we have

1
n
= (
(T

)
3
1 T

) = O((T

)
3
)
T

= (
1
(n)
1/3
) (27)
Therefore, the asymptotic optimal deadlock detection frequency 1/T

is O((n)
1/3
).
1 0
-2
1 0
-1
1 0
0
1 0
1
1 0
2
1 0
3
1 0
4
1 0
5
1 0
6
1 0
7
1 0
8
D e a d lo c k D e te c tio n In te rv a l (lo g )
L
o
n
g
-
r
u
n
m
e
a
n
a
v
e
r
a
g
e
c
o
s
t
(
lo
g
)
n=1000
n=500
n=200
n=100
n=50
=1/30s
Figure 2: Cost of Deadlock Handling vs. Detection Interval (n: number of processes)
10
-2
10
-1
10
0
10
1
10
2
10
6
10
7
10
8
10
9
Deadlock Detection Interval (log)
L
o
n
g
-
r
u
n
m
e
a
n
a
v
e
r
a
g
e
c
o
s
t
(
l
o
g
)
=1s
=1/30s
=1/60s
=1/90s
=1/120s
n=1000
Figure 3: Cost of Deadlock Handling vs. Deadlock Formation Rate
As an illustration, we consider an example as follows. Let C
R
(t) = n
3
(1 exp(t)),
C
D
= n
2
. In accordance with Theorem 1, the long-run mean average cost of deadlock
20
handling thus is written as
C(T) =
n
2
+ n
3
(T + exp(T) 1)
T
. (28)
Figs(2)-(3) show log-log plots of a family of curves illustrating the dependence of long-
run mean average cost of deadlock handling upon detection interval. The x-axis denotes
the deadlock detection interval and the y-axis denotes the long-run mean average cost of
deadlock handling.
# of Processes Optimal Detection Interval ( = 1)
50 0.214699(s)
100 0.148555(s)
200 0.103495(s)
500 0.064189(s)
1000 0.045402(s)
# of Processes Optimal Detection Interval ( = 1/30)
50 2.0223(s)
100 1.0973(s)
200 0.6832(s)
500 0.3942(s)
1000 0.2675(s)
Table 1: Optimal Detection Interval vs. # of Processes
In Fig(2), we present plots of the deadlock detection interval and cost of deadlock handling
under dierent the total number of processes, 50, 100, 200, 500, and 1000, respectively. Fig(3)
shows the relationship between the overall cost of deadlock handling and deadlock detection
interval under the dierent deadlock formation rates, 1s, 1/30s, 1/60s, 1/90s, and 1/120s,
respectively. Figs(2)-(3) visualizes convexity that suggests the existence of an optimal de-
tection frequency, illustrating that the overall cost of deadlock handling increases with the
total number of processes and deadlock formation rate.
A detailed calculation given in Table 1 shows that as the number of processes in a dis-
tributed system increases, the optimal detection interval decreases, which is clearly in line
21
with our theoretical analysis. In the sequel, we study the impact of coordinated v.s. random
deadlock detection scheduling on the performance of deadlock handling. We consider two
strategies of deadlock detection scheduling: (1) centralized, coordinated deadlock detection
scheduling, and (2) fully distributed, uncoordinated deadlock detection scheduling.
The centralized scheduling excels in its simplicity in implementation and system main-
tenance, but undermines the reliability and resilience against failures because one and only
one process is elected as the initiator of deadlock detections in a distributed system. In
contrast, the fully distributed scheduling excels in the reliability and resilience against fail-
ures because every process in the distributed system can independently initiate detections
[15], without a single point of failure. However, due to the lack of coordination in deadlock
detection initiation among processes, it presents a dierent mathematical problem from the
centralized deadlock detection scheduling.
In the previous discussions we have focused on the derivation of optimal frequency of
deadlock detection in connection with the rate of deadlock formation and the message com-
plexities of deadlock detection and resolution algorithms, assuming deadlock detections are
centrally scheduled at a xed rate of 1/T. To capture the lack of coordination in fully dis-
tributed scheduling, we will study the case where processes randomly, independently initiate
the detection of deadlocks.
Let n be the number of processes in a distributed system and T be the optimal time inter-
val between any two consecutive deadlock detections in the centralized scheduling. Consider
a fully distributed deadlock detection scheduling, where each process initiates deadlock de-
tection at a rate of 1/(nT) independently. Although the average interval between deadlock
detections in the fully distributed scheduling remains T (the same as its centralized counter-
22
part), the actual occurring times of those detections are likely to be non-uniformly spaced
because the initiation of deadlock detection is performed by the processes in a completely
uncoordinated fashion.
In the following we will study the fully distributed (random) scheduling and compare
it with the centralized scheduling. Consider a sequence of independently and identically
distributed iid random variables Y
i
, i 1 dened on (0, ) following certain distribution
H. The sequence Y
i
, i 1 represents the inter-arrival times of deadlock detections initiated
by the fully distributed scheduling, and it is assumed to be independent of the arrival of
deadlock formations. It is obvious that the centralized scheduling is a special case of the
fully distributed scheduling.
Let H be the family of all distribution functions on (0, ) with nite rst moment.
Namely,
H =
_
H: H is a CDF on (0, ),
_

0

H(t)dt <
_
(29)
where

H(t) 1 H(t), t 0.
The following theorem states that the lack of coordination in deadlock detection initia-
tion by fully distributed scheduling will introduce additional overhead in deadlock handling.
Therefore the fully distributed scheduling in general cannot perform as eciently as its
centralized counterpart.
Theorem 3 Let C
H
denote the long-run mean average cost under fully distributed scheduling
with a random detection interval Y characterized by certain distribution H H with the mean
of , and C(T) denote the long-run mean average cost under centralized scheduling with a
23
xed detection interval T. Then
C
H
C(T), (30)
when E(Y ) = = T.
Proof: Since the sequence Y
i
, i 1 of interarrival times of deadlock detection is assumed
to be independent of the Poisson deadlock formations, it is easy to see that the random costs
over the intervals (0, Y
1
], (Y
1
, Y
1
+ Y
2
], . . . are iid. Using the same line of reasoning in the
proof of Theorem 1, the long-run mean average cost is expressed as
C
H
=
E(random cost over Y )
E(Y )
, (31)
where Y His a random variable representing the interval between two consecutive deadlock
detections. Let (Y ) be the random cost in the interval Y . The expected cost over the interval
Y is given by
E((Y )) = EE[(Y )[Y ] =
_

0
E(C
D
+
N(y)

n=1
C
R
(y S
n
)I
{N(y)>0}
)dH(y), (32)
where S
n
=
n

i=1
X
i
denotes the time of the nth deadlock formation and N(y) represents the
number of independent deadlocks occurred in the time interval (0, y). It follows from the
independence of X
i
, i 1 and Y
i
, i 1, and from Eq(32), the long-run mean average
cost is
C
H
=
E((Y ))
E(Y )
=
_

0
(C
D
+
_
y
0
C
R
(t)dt)dH(y)
E(Y )
=
C
D
E(Y )
+
_

0
__

t
C
R
(t)dH(y)
_
dt
E(Y )
=
C
D
E(Y )
+

0
C
R
(t)

H(t)dt
E(Y )
. (33)
When E(Y ) = = T, meaning that the xed deadlock detection interval T equals to the
mean value of the random detection interval Y , we compare the centralized (xed) detection
24
scheduling with the rate of 1/T with the fully distributed (random) one with the mean rate of
1/E(Y ) = 1/. According to Theorem 1, the long run mean average cost of xed detection
is given as
C(T) =
C
D

0
C
R
(t)dt

. (34)
Subtracting Eq(34) from Eq(33) yields
C
H
C(T) =

__

0
C
R
(t)

H(t)dt
_

0
C
R
(t)dt
_
=

__

C
R
(t)

H(t)dt
_

0
C
R
(t)H(t)dt
_

_
C
R
()
_

H(t)dt C
R
()
_

0
H(t)dt
_
=
C
R
()

__

H(t)dt
_

0
(1

H(t))dt
_
=
C
R
()

__

0

H(t)dt
_
= 0. (35)
Hence we have
C
H
C(T). (36)
Theorem 3 is thus established.
It can be seen from Eq(36) that C
H
C(T) and the equality holds if and only if Y is a
degenerate random variable when Prob(Y = T) = 1. Theorem 3 asserts that the fully dis-
tributed (random) deadlock detection scheduling in general results in an increased overhead
in overall deadlock handling.
5 Conclusion
Deadlock detection scheduling is an important, yet often overlooked aspect of distributed
deadlock detection and resolution. The performance of deadlock handling not only depends
upon per-execution complexity of deadlock detection/resolution algorithms, but also depends
fundamentally upon deadlock detection scheduling and the rate of deadlock formation. Ex-
cessive initiation of deadlock detection results in an increased number of message exchange
25
in the absence of deadlocks, while insucient initiation of deadlock detection incurs an in-
creased cost of deadlock resolution in the presence of deadlocks. As a result, reducing the
per-execution cost of distributed deadlock detection/resolution algorithms alone does not
warrant the overall performance improvement on deadlock handling.
The main thrust of this paper is to bring an awareness to the problem of deadlock
detection scheduling and its impact on the overall performance of deadlock handling. The key
element in our approach is to develop a time-dependent model that associates the deadlock
resolution cost with the deadlock persistence time. It assists the study of time-dependent
deadlock resolution cost in connection with the rate of deadlock formation and the frequency
of deadlock detection initiation, diering signicantly from the past research that focuses on
minimizing per-detection and per-resolution costs.
Our stochastic analysis, which solidies the ideas presented in [10, 26, 23, 11], shows that
there exists a unique deadlock detection frequency that guarantees a minimum long-run
mean average cost for deadlock handling when the total number of processes in a distributed
system is suciently large, and that the cost of overall deadlock handling grows linearly with
the rate of deadlock formation.
In addition, we study the fully distributed (random) deadlock detection scheduling and
its impact on the performance of deadlock handling. We prove that in general the lack of
coordination in deadlock detection initiation among processes will increase the overall cost
of deadlock handling.
Theoretical results obtained in this paper could help system designers/practitioners to
better understand the fundamental performance tradeo between deadlock detection and
deadlock resolution costs, as well as the innate dependency of optimal detection frequency
26
upon deadlock formation rate. However, there are still a lot of questions regarding how to
use theoretical results to ne-tune the performance of a distributed system. Determination
of the actual rate of deadlock formation and verication of the Poisson process are problems
of great complexity that can be inuenced by many known/unknown factors such as the
granularity of locking, actual distribution of resource, process mix, and resource request and
release patterns [26]. Tapping into system logging les and inferring the actual deadlock
formation rate via data mining could provide an eective and feasible way to translate
theoretical insights into actual system performance gain.
6 Acknowledgements
We would like to thank Drs. Marek Rusinkiewicz and Ritu Chadh at Applied Research,
Telcordia Technologies for their constructive comments on the manuscript of this paper. We
would also like to thank three anonymous reviewers for critically reviewing the manuscript
and for their insightful comments. We would like to especially thank Dr. Shu-Chan Hsu in
Department of Cell Biology and Neuroscience at Rutgers University for her encouragement
and support.
References
[1] Roberto Baldoni and Silvio Salz. Deadlock Detection in Multidatabase Systems: a
Performance Analysis. DIstributed Systems Engineering, 4:244252, December 1997.
[2] Azzedine Boukerche and Carl Tropper. A Distributed Graph Algorithm for the Detec-
tion of Local Cycles and Knots. IEEE Transactions on Parallel and Distributed Systems,
9(8):748757, August 1998.
27
[3] G. Bracha and S. Toueg. Distributed Deadlock Detection. Distributed Computing,
2:127138, 1987.
[4] K.M. Chandy, J. Misra, and L. Hass. Distributed Deadlock. ACM Transaction on
Computer Systems, 1(2):144156, May 1983.
[5] Shigang Chen, Yi Deng, and Wei Sun. Optimal Dealock Detection in Distributed Sys-
tems Based on Locally Constructed Wait-for Graph. In Proceedings of the 16th Inter-
national Conference on Distributed Computing Systems, pages 613619, 1996.
[6] Shigang Chen and Yibei Ling. Stochastic Analysis of Distributed Deadlock Scheduling.
In Proceedings of the 24th ACM Symposium on Principles of Distributed Computing,
pages 265273, July 17-20 2005.
[7] Jose Ramon Gonzales de Mendivil, Jose Ramon Garitagoitia, Carlos F. Alastruey, and
J.M. Bernabeu-Auban. A Distributed Deadlock Resolution Algorithm for the AND
Model. IEEE Transactions on Parallel and Distributed Systems, 10(5):433447, May
1999.
[8] Jim Gray, P. Homan, Ron Obermarck, and Henry Korth. A Straw-man Analysis of the
Probability of Waiting and Deadlock in a Database System. IBM Research, RJ3066,
February 1981.
[9] Young M. Kim, Tan H. Lai, and Neelam Soundarajan. Ecient Distributed Deadlock
Detection and Resolution Using Probes, Tokens, and Barriers. In Proceedings of the
International Conference on Parallel and Distributed Systems, pages 584591, 1997.
28
[10] Edgar Knapp. Deadlock Detection in Distributed Databases. ACM Computing Surveys,
19(4):303328, 1987.
[11] Natalija Krivokapic, Alfons Kemper, and Ehud Gudes. Deadlock Detection in Dis-
tributed Database Systems: A New Algorithm and a Comparative Performance Analy-
sis. VLDB Journal: Very Large Data Bases, 8(2):79100, 1999.
[12] Ajay D. Kshemkalyani and Mukesh Singhal. Ecient Detection and Resolution of Gen-
eralized Distributed Deadlocks. IEEE Transactions on Software Engineering, 20(1):43
54, January 1994.
[13] Ajay D. Kshemkalyani and Mukesh Singhal. Distributed Detection of Generalized Dead-
locks. In Proceedings of the 1997 International Conference on Distributed Computing
Systems, pages 553560, 1997.
[14] Ajay D. Kshemkalyani and Mukesh Singhal. A One-Phase Algorithm to Detect Dis-
tributed Deadlocks in Replicated Databases. IEEE Transactions on Knowledge and
Data Engineering, 11(6):880895, 1999.
[15] Soojung Lee. Fast, Centralized Detection and Resolution of Distributed Deadlocks in
the Generalized model. IEEE Transactions on Software Engineering, 30(8):561573,
September 2004.
[16] Soojung Lee and Junguk L. Kim. Performance Analysis of Distributed Deadlock Dectec-
tion Algorithms. IEEE Transactions on Knowledge and Data Engineering, 13(3):623
636, 2001.
29
[17] Xuemin Lin and Jian Chen. An Optimal Deadlock Resolution Algorithm in Multi-
database Systems. In Proceedings of the 1996 International Conference on Parallel and
Distributed Systems, pages 516521, 1996.
[18] Yibei Ling, Jie Mi, and Xiaola Lin. A Variational Calculus Approach to Optimal
Checkpoint Placement. IEEE Transactions on Computers, 50(7):699708, July 2001.
[19] Philip P. Macri. Deadlock Detection and Resolution in a CODASYL based Data Man-
agement System. In Proceedings of the 1976 ACM SIGMOD International Conference
on Management of Data, pages 4549, 1976.
[20] William A. Massey. A Probabilistic Analysis of a Database System. ACM SIGMETRICS
Performance Evaluation Review, 14(1):141146, 1986.
[21] Jayadev Misra. Distributed Discrete-Event Simulation. ACM Computing Surveys,
18(1):3965, March 1986.
[22] Ron Obermarck. Distributed Deadlock Detection Algorithm. ACM Transactions on
Database Systems, 7(2):187208, June 1982.
[23] Young Chul Park, Peter Scheuermann, and Snag Ho Lee. A Periodic Deadlock Detec-
tion and Resolution Algorithm with a New Graph Model for Sequential Transaction
Processing. In Proceedings of the Eighth International Conference of Data Engineering,
pages 202209, February 1992.
[24] M. Roesler and W. A. Burkhard. Semantic Lock Models in Object-oriented Distributed
Systems and Deadlock Resolution. In Proceedings of the 1988 ACM SIGMOD Interna-
tional Conference on Management of Data, pages 361370, 1988.
30
[25] Sheldon M. Ross. Stochastic Processes. John Wiley & Sons, Inc., New York, 1996.
[26] Mukesh Singhal. Deadlock detection in distributed systems. IEEE Computer Magazine,
40(8):3748, November 1989.
[27] Igor Terekhov and Tracy Camp. Time Ecient Deadlock Resolution Algorithms. In-
formation Processing Letters, 69:149154, 1999.
[28] Carl Tropper and Azzedine Boukerche. Parallel simulations of communicating nite
state machines. In Proceedings of the SCS Multiconf on Parallel and Distributed Simu-
lation, pages 143150, May 1993.
[29] Jesus Villadangos, Federico Farina, Jose Ramon Gonzales de Mendivil, Jose Ramon
Garitagoitia, and Alberto Cordoba. A Safe Algorithm for Resolving OR Deadlocks.
IEEE Transactions on Software Engineering, 29(7):608622, July 2003.
[30] J.W. Wang, Shing-Tsaan Huang, and Nian-Shing Chen. A Distributed Algorithm for
Detecting Generalized Deadlocks. Technical Report (SF-C-010-1), Computer Science,
National Tsing-Hua University, 1990.
[31] Yi-Min Wang, Michael Merritt, and Alexander B. Romanovsky. Guaranteed Deadlock
Recovery: Deadlock Resolution with Rollback Propagation. In Technical Report Number
648, 1998.
[32] Sugath Warnakulasuriya and Timothy Mark Pinkston. A Formal Model of Message
Blocking and Deadlock Resolution in Interconnection Networks. IEEE Transactions on
Parallel and Distributed Systems, 11(3):212229, March 2000.
31