Overview of HPC Interconnects
1 like385 views
In this deck from ATPESC 2019, Ken Raffenetti from Argonne presents an overview of HPC interconnects. "The Argonne Training Program on Extreme-Scale Computing (ATPESC) provides intensive, two-week training on the key skills, approaches, and tools to design, implement, and execute computational science and engineering applications on current high-end computing systems and the leadership-class computing systems of the future." Watch the video: https://wp.me/p3RLHQ-luc Learn more: https://extremecomputingtraining.anl.gov/ Sign up for our insideHPC Newsletter: http://insidehpc.com/newsletter
![U.S. DOE Potential System Architecture Targets
ATPESC (07/29/2019)
System attributes 2010 2018-2019 2021-2022
System peak 2 Peta 150-200 Petaflop/sec 1 Exaflop/sec
Power 6 MW 15 MW 20 MW
System memory 0.3 PB 5 PB 32-64 PB
Node performance 125 GF 3 TF 30 TF 10 TF 100 TF
Node memory BW 25 GB/s 0.1TB/sec 1 TB/sec 0.4TB/sec 4 TB/sec
Node concurrency 12 O(100) O(1,000) O(1,000) O(10,000)
System size (nodes) 18,700 50,000 5,000 100,000 10,000
Total Node
Interconnect BW
1.5 GB/s 20 GB/sec 200GB/sec
MTTI days O(1day) O(1 day)
Past
production
Current generation
(e.g., CORAL)
Exascale
Goals
[Includes modifications to the DOE Exascale report]](https://image.slidesharecdn.com/atpesc2019track-157-291130amraffenetti-interconnects-200209161315/85/Overview-of-HPC-Interconnects-2-320.jpg)
Ad
More Related Content
What's hot (20)
Similar to Overview of HPC Interconnects (20)
Ad
More from inside-BigData.com (20)
Ad
Recently uploaded (20)
Overview of HPC Interconnects
- 1. Interconnects Ken Raffenetti Software Development Specialist Programming Models and Runtime Systems Group Mathematics and Computer Science Division Argonne National Laboratory
- 2. U.S. DOE Potential System Architecture Targets ATPESC (07/29/2019) System attributes 2010 2018-2019 2021-2022 System peak 2 Peta 150-200 Petaflop/sec 1 Exaflop/sec Power 6 MW 15 MW 20 MW System memory 0.3 PB 5 PB 32-64 PB Node performance 125 GF 3 TF 30 TF 10 TF 100 TF Node memory BW 25 GB/s 0.1TB/sec 1 TB/sec 0.4TB/sec 4 TB/sec Node concurrency 12 O(100) O(1,000) O(1,000) O(10,000) System size (nodes) 18,700 50,000 5,000 100,000 10,000 Total Node Interconnect BW 1.5 GB/s 20 GB/sec 200GB/sec MTTI days O(1day) O(1 day) Past production Current generation (e.g., CORAL) Exascale Goals [Includes modifications to the DOE Exascale report]
- 3. General Trends in System Architecture § Number of nodes is increasing, but at a moderate pace § Number of cores/threads on a node is increasing rapidly § Each core is not increasing in speed (clock frequency) § What does this mean for networks? – More cores will drive the network – More sharing of the network infrastructure – The aggregate amount of communication from each node will increase moderately, but will be divided into many smaller messages – A single core may not be able fully saturate the NIC ATPESC (07/29/2019)
- 4. A Simplified Network Architecture § Hardware components – Processing cores and memory subsystem – I/O bus or links – Network adapters/switches § Software components – Communication stack § Balanced approach required to maximize user-perceived network performance ATPESC (07/29/2019) P0 Core0 Core1 Core2 Core3 P1 Core0 Core1 Core2 Core3 Memory MemoryI / O B u s Network Adapter Network Switch Processing Bottlenecks I/O Interface Bottlenecks Network End-host Bottlenecks Network Fabric Bottlenecks
- 5. Agenda ATPESC (07/29/2019) Network Adapters Network Topologies Network/Processor/Memory Interactions
- 6. Bottlenecks on Traditional Network Adapters § Network speeds plateaued at around 1Gbps – Features provided were limited – Commodity networks were not considered scalable enough for very large-scale systems ATPESC (07/29/2019) P0 Core0 Core1 Core2 Core3 P1 Core0 Core1 Core2 Core3 Memory MemoryI / O B u s Network Adapter Network Switch Network Bottlenecks Ethernet (1979 - ) 10 Mbit/sec Fast Ethernet (1993 -) 100 Mbit/sec Gigabit Ethernet (1995 -) 1000 Mbit /sec ATM (1995 -) 155/622/1024 Mbit/sec Myrinet (1993 -) 1 Gbit/sec Fibre Channel (1994 -) 1 Gbit/sec
- 7. End-host Network Interface Speeds § HPC network technologies provide high bandwidth links – InfiniBand EDR gives 100 Gbps per network link • Will continue to increase (HDR 200 Gbps, etc.) – Multiple network links becoming a common place • ORNL Summit and LLNL Sierra machines, Japanese Post T2K machine • Torus style or other multi-dimensional networks § End-host peak network bandwidth is “mostly” no longer considered a major limitation § Network latency still an issue – That’s a harder problem to solve – limited by physics, not technology • There is some room to improve it in current technology (trimming the fat) • Significant effort in making systems denser so as to reduce network latency § Other important metrics: message rate, congestion, … ATPESC (07/29/2019)
- 8. Simple Network Architecture (past systems) § Processor, memory, network are all decoupled ATPESC (07/29/2019) P0 Core0 Core1 Core2 Core3 P1 Core0 Core1 Core2 Core3 Memory MemoryI / O B u s Network Adapter Network Switch
- 9. Integrated Memory Controllers (current systems) § In the past 10 years or so, memory controllers have been integrated on to the processor § Primary purpose was scalable memory bandwidth (NUMA) § Also helps network communication – Data transfer to/from network requires coordination with caches § Several network I/O technologies exist – PCIe, HTX, NVLink – Expected to provide higher bandwidth than what network links will have ATPESC (07/29/2019) P1 Core0 Core1 Core2 Core3 Memory Memory I / O B u s Network Adapter Network Switch Memory Controller P0 Core0 Core1 Core2 Core3 Memory Controller
- 10. Integrated Networks (future systems?) ATPESC (07/29/2019) Network Switch P0 Core0 Core1 Core2 Core3 Memory Controller Network Interface Off-chip Memory § Several vendors are considering processor-integrated network adapters § May improve network bandwidth – Unclear if the I/O bus would be a bottleneck § Improves network latencies – Control messages between the processor, network, and memory are now on-chip § Improves network functionality – Communication is a first-class citizen and better integrated with processor features – E.g., network atomic operations can be atomic with respect to processor atomics In-package Memory P1 Core0 Core1 Core2 Core3 Memory Controller Network Interface Off-chip Memory In-package Memory
- 11. Processing Bottlenecks in Traditional Protocols § Ex: TCP/IP, UDP/IP § Generic architecture for all networks § Host processor handles almost all aspects of communication – Data buffering (copies on sender and receiver) – Data integrity (checksum) – Routing aspects (IP routing) § Signaling between different layers – Hardware interrupt on packet arrival or transmission – Software signals between different layers to handle protocol processing in different priority levels ATPESC (07/29/2019) P0 Core0 Core1 Core2 Core3 P1 Core0 Core1 Core2 Core3 Memory MemoryI / O B u s Network Adapter Network Switch Processing Bottlenecks
- 12. Network Protocol Stacks: The Offload Era § Modern networks are spending more and more network real-estate on offloading various communication features on hardware § Network and transport layers are hardware offloaded for most modern HPC networks – Reliability (retransmissions, CRC checks), packetization – OS-based memory registration, and user-level data transmission ATPESC (07/29/2019)
- 13. Comparing Offloaded Network Stacks with Traditional Network Stacks ATPESC (07/29/2019) Application Layer MPI, PGAS, File Systems Transport Layer Low-level interface Reliable/unreliable protocols Link Layer Flow-control, Error Detection Physical Layer Hardware Offload Copper or Optical HTTP, FTP, MPI, File Systems Routing Physical Layer Link Layer Network Layer Transport Layer Application Layer Traditional Ethernet Sockets Interface TCP, UDP Flow-control and Error Detection Copper, Optical or Wireless Network Layer Routing Management tools DNS management tools
- 14. Current State for Network APIs § A large number of network vendor specific APIs – InfiniBand verbs, Intel PSM2, IBM PAMI, Cray Gemini/DMAPP, … § Recent efforts to standardize these low-level communication APIs – Open Fabrics Interface (OFI) • Effort from Intel, CISCO, etc., to provide a unified low-level communication layer that exposes features provided by each network – Unified Communication X (UCX) • Effort from Mellanox, IBM, ORNL, etc., to provide a unified low-level communication layer that allows for efficient MPI and PGAS communication – Portals 4 • Effort from Sandia National Laboratory to provide a network hardware capability centric API ATPESC (07/29/2019)
- 15. User-level Communication: Memory Registration ATPESC (07/29/2019) 1. Registration Request • Send virtual address and length 2. Kernel handles virtual->physical mapping and pins region into physical memory • Process cannot map memory that it does not own (security !) 3. Network adapter caches the virtual to physical mapping and issues a handle 4. Handle is returned to application Before we do any communication: All memory used for communication must be registered 1 3 4 Process Kernel Network 2
- 16. User-level Communication: OS Bypass ATPESC (07/29/2019) Process Kernel Network User-level APIs allow direct interaction with network adapters § Contrast with traditional network APIs that trap down to the kernel § Eliminates heavyweight context switch § Memory registration caches allow for fast buffer re-use, further reducing dependence on the kernel
- 17. Send/Receive Communication ATPESC (07/29/2019) Network Interface Memory Memory Network Interface Send Recv Memory Segment Send entry contains information about the send buffer (multiple non-contiguous segments) Processor Processor Send Recv Memory Segment Receive entry contains information on the receive buffer (multiple non-contiguous segments); Incoming messages have to be matched to a receive entry to know where to place Hardware ACK Memory Segment Memory Segment Memory Segment
- 18. PUT/GET Communication ATPESC (07/29/2019) Network Interface Memory Memory Network Interface Send Recv Memory Segment Send entry contains information about the send buffer (multiple segments) and the receive buffer (single segment) Processor Processor Send Recv Memory Segment Hardware ACK Memory Segment Memory Segment
- 19. Atomic Operations ATPESC (07/29/2019) Network Interface Memory Memory Network Interface Send Recv Memory Segment Send entry contains information about the send buffer and the receive buffer Processor Processor Send Recv Source Memory Segment OP Destination Memory Segment
- 20. Network Protocol Stacks: Specialization § Increasing network specialization is the focus today – The next generation of networks plan to have further support for noncontiguous data movement, and multiple contexts for multithreaded architectures § Some networks, such as the Tofu network, Cray Aries and InfiniBand, are also offloading some MPI and PGAS features on to hardware – E.g., PUT/GET communication has hardware support – Increasing number of atomic operations being offloaded to hardware • Compare-and-swap, fetch-and-add, swap – Collective operations (NIC and switch support) • SHARP collectives – Hardware tag matching for MPI send/recv • Cray Seastar, Bull BXI, Mellanox Infiniband (ConnectX-5 and later) ATPESC (07/29/2019)
- 21. Agenda ATPESC (07/29/2019) Network Adapters Network Topologies Network/Processor/Memory Interactions
- 22. Traditional Network Topologies: Crossbar § A network topology describes how different network adapters and switches are interconnected with each other § The ideal network topology (for performance) is a crossbar – Alltoall connection between network adapters – Typically done on a single network ASIC – Current network crossbar ASICs go up to 64 ports; too expensive to scale to higher port counts – All communication is nonblocking ATPESC (07/29/2019)
- 23. Traditional Network Topologies: Fat-tree § The most common topology for small and medium scale systems is a fat-tree – Nonblocking fat-tree switches available in abundance • Allows for pseudo nonblocking communication • Between all pairs of processes, there exists a completely nonblocking path, but not all paths are nonblocking – More scalable than crossbars, but the number of network links still increases super-linearly with node count • Can get very expensive with scale ATPESC (07/29/2019)
- 24. Network Topology Trends § Modern topologies are moving towards more “scalability” (with respect to cost, not performance) § Blue Gene, Cray XE/XK, and K supercomputers use a torus-network; Cray XC uses dragonfly – Linear increase in the number of links/routers with system size – Any communication that is more than one hop away has a possibility of interference – congestion is not just possible, but common – Even when there is no congestion, such topologies increase the network diameter causing performance loss § Take-away: topological locality is important and its not going to get better ATPESC (07/29/2019)
- 25. Network Congestion Behavior: IBM BG/P 0 500 1000 1500 2000 2500 3000 3500 1 2 4 8 16 32 64 128 256 512 1K 2K 4K 8K 16K 32K 64K 128K 256K 512K 1M Bandwidth(Mbps) Message Size (bytes) P2-P5 P3-P4 No overlap ATPESC (07/29/2019) P0 P1 P2 P3 P4 P5 P6 P7
- 26. 2D Nearest Neighbor: Process Mapping (XYZ) ATPESC (07/29/2019) X-Axis Z-Axis Y-Axis
- 27. Nearest Neighbor Performance: IBM BG/P ATPESC (07/29/2019) 0 100 200 300 400 500 600 700 800 900 2 4 8 16 32 64 128 256 512 1K ExecutionTime(us) Grid Partition (bytes) System Size : 16K Cores XYZT TXYZ ZYXT TZYX 0 500 1000 1500 2000 2500 2 4 8 16 32 64 128 256 512 1K ExecutionTime(us) Grid Partition (bytes) System Size : 128K Cores XYZT TXYZ ZYXT TZYX 2D Halo Exchange
- 28. Agenda ATPESC (07/29/2019) Network Adapters Network Topologies Network/Processor/Memory Interactions
- 29. Network Interactions with Memory/Cache § Most network interfaces understand and work with the cache coherence protocols available on modern systems – Users do not have to ensure that data is flushed from cache before communication – Network and memory controller hardware understand what state the data is in and communicate appropriately ATPESC (07/29/2019)
- 30. Send-side Network Communication ATPESC (07/29/2019) North BridgeCPU NIC FSB Memory Bus I/OBus Appln. Buffer L3 $ Memory Appln. Buffer
- 31. Receive-side Network Communication ATPESC (07/29/2019) North BridgeCPU NIC FSB Memory Bus I/OBus Appln. Buffer L3 $ Memory Appln. Buffer
- 32. Network/Processor Interoperation Trends § Direct cache injection – Most current networks inject data into memory • If data is in cache, they flush cache and then inject to memory – Some networks are investigating direct cache injection • Data can be injected directly into the last-level cache • Can be tricky since it can cause cache pollution if the incoming data is not used immediately § Atomic operations – Current network atomic operations are only atomic with respect to other network operations and not with respect to processor atomics • E.g., network fetch-and-add and processor fetch-and-add might corrupt each other’s data – With network/processor integration, this is expected to be fixed ATPESC (07/29/2019)
- 33. Network Interactions with Accelerators § PCI Express peer-to-peer capabilities enables network adapters to directly access third-party devices – Coordination between network adapter and accelerator (GPUs, FPGAs, …) – Data does not need to be copied into to/from buffers when going over the network – GPUDirect RDMA for NVIDIA GPUs is one example ATPESC (07/29/2019)
- 34. Summary § These are interesting times for all components in the overall system architecture: processor, memory, interconnect – And interesting times for computational science on these systems § Interconnect technology is rapidly advancing – More hardware integration is the key to removing bottlenecks and improve functionality • Processor/memory/network integration is already in progress and will continue for the foreseeable future – Offload technologies continue to evolve as we move more functionality to the network hardware – Network topologies are becoming more “shared” (cost saving) ATPESC (07/29/2019)