平成25年度 後期 火曜 第2時限(10:40-12:10) 吉永 努(UEC) yosinaga@is.uec.ac.jp ネットワークコンピューティング論Ⅱ 平成25年度 後期 火曜 第2時限(10:40-12:10) 吉永 努(UEC) yosinaga@is.uec.ac.jp NC論2
内 容 分散・並列処理計算機における相互結合ネットワークとその上でのメッセージ・ルーティング技法などについて学ぶ 資料 http://comp.is.uec.ac.jp/yoshinagalab/yoshinaga/dp2.html http://ceng.usc.edu/smart/presentations/archives/AppendixE.ppt (253 slides, 13MB) http://booksite.mkp.com/9780123838728/references/appendix_f.pdf (P.118, 2MB) TA: 重信 裕政君 shige@comp.is.uec.ac.jp NC論2
References T. M. Pinkston and J. Duato: Interconnection Networks, Appendix E in Computer Architecture: A Quantitative Approach, 4th Edition, Morgan Kaufmann publishers (2006). 5th Edition, Morgan Kaufmann publishers (2011). J. Duato, S. Yalamanchili, L. Ni: Interconnection Networks - an Engineering Approach-, 第2版, Morgan Kaufmann publishers (2003) 富田眞治: 並列コンピュータ、昭晃堂(1996) W.D. Dally, B. Towles: Principles and Practices of Interconnection Networks, Morgan Kaufmann publishers (2003)
What is an interconnection Network? It is a programmable system that transports data between terminals, such as processors and memory. It is programmable in the sense that it makes different connections at different points. It is a system because it is composed of many components: buffers, channels, switches, and controls that works together to deliver data. NC論2
Interconnection Network (1/2) P P P M M M Multicomputer NC論2
Interconnection Network (2/2) P P P Interconnection Network M M M UMA type shared memory multiprocessor It is also called dance-hall architecture. NC論2
Trend Its performance is increasing with processor performance at a rate of 50% per year. Communication is a limiting factor in the performance of many modern systems. Buses have been unable to keep up with the bandwidth demand, and point-to-point interconnection networks are rapidly taking over. NC論2
Computer Classifications (%) 2013/06 2012/06 2011/06 MPP 16.6 18.6 17.4 Cluster 83.4 81.4 82.2 Others 0.0 0.4 http://www.top500.org/ share of the TOP500 June, 2013 – June, 2011 NC論2
Infiniband QDR (40Gbps) ×2 Examples of clusters Processors Accelerator Interconnect Tianhe-2 (天河2号)China 2013 Intel Xeon E5-2692 12C 2.2 GHz×2 ×16K Xeon Phi 31S1P (57 cores)×3 ×16K TH Express-2 (proprietary) Fat tree Tsubame 2.5 Tokyo Tech. Xeon X5670 2.93GHz×2 ×1,408 NVIDIA Kepler K20x ×3×1,048 Infiniband QDR (40Gbps) ×2 NC論2
Examples of MPPs #core Rmax Node Topology K computer @RIKEN Fujitsu 2011 SPARC64 VIIIfx 2 GHz (16 GFlops× 8 cores) 6D mesh/ 3D torus Tofu interconnect 80K-node x 8-core = 640K-core 10.51 PFlops 7,890 KW Titan@ORNL Cray XK7 2012 AMD Opteron 16C 2.2 GHz + NVIDIA K20x Gemini interconnect 18,688 nodes (200 Cabinets) 27.11 PFlops 8,209 KW NC論2
Other Networks of Supercomputers Sequoia (2011): 5D torus, proprietary IBM SeaStar Pleiades / NASA (2011): partial 11D hypercube topology with IB QDR/DDR Red Sky/ Sandia National Lab. (2010): 3D torus (12 bristled node) with IB QDR switches IBM Roadrunner (2009): fat-tree with IB DDR Earth Simulator2 / NEC SX-9E (2009): Fat-Tree (64GB/s/cpu, 8-CPU/node, 160 nodes) IBM Blue Gene/L (2004): 3D torus proprietary (64 x 32 x 32 = 64K nodes) NC論2
Architecture vs. software memory programming UMA (SMP) shared OpenMP NUMA (MPP) distributed (not shared) MPI (Message Passing Interface) NC論2
Network Design (1/3) Performance: latency and throughput (bandwidth) Scalability: #processors vs. network, memory, I/O bandwidth Incremental expandability: small to maximum size Partitionability: netwrok may be partitioned for several users NC論2
Network Design (2/3) Simplicity: simple design, higher clock frequency, easy to use Distance span: smaller system is preferred for noise and cable delay, etc. Physical constraints: packaging (pin count), wiring(wire length), and maintenance (power consumption) should meet physical limitation. NC論2
Network Design (3/3) Reliability: fault tolerant, reliable communication, hot swap Expected workload: robust performance over a wade range of traffic conditions. Cost: trade-offs between cost and performance. NC論2
Classifiction of Interconnection Networks Shared-Medium Networks Local area networks (ethernet, token ring) Backplane bus (e.g. SUN Gigaplane) Direct Networks (router-based) mesh, torus, hypercube, tree, … etc. Indirect Networks (switch-based) Hybrid Networks NC論2
Shared-Medium Networks (LAN) Arbitration that determines the mastership of the shared-medium network to resolve network access is needed. The most well-known protocol is carrier-sense multiple access with collision detection (CSMA/CD). Token bus and token ring pass a token from the owner which has the right to access the bus/ring and resolve nondeterministic waiting time. NC論2
Shared-Medium Networks (Backplane bus) It is commonly used to interconnect processor(s) and memory modules to provide SMP (Symmetrical Memory Processor) architecture. It is realized by printed lines on a circuit board by discrete wiring. Gigaplane in SUN Enterprise x000 server(1996): 2.6GB/s, 256 bits data, 42 bits address, 83.8MHz clock. NC論2
Direct (static) Networks Consists of a set of nodes. Each node is directly connected to a subset of other nodes in the network. Examples: 2D mesh (intel Paragon), 3D mesh (MIT J-Mahine) 2D torus (Fujitsu AP3000), 3D torus (Cray T3D, T3E) Hypercube (CM1, CM2, nCUBE) NC論2
Mesh topology node 2D 3D NC論2
Torus topology 2D (4-ary 2-cube) 3D (3-ary 3-cube) NC論2
Hypercube (binary n-cube) 4D (2-ary 4-cube) NC論2
tree x tree Binary tree fat tree NC論2
Hierarchical topology (1/2) Pyramid (Hierarchical 2D mesh) Hierarchical ring NC論2
Hierarchical topology (2/2) Cube-connected cycles RDT (Recursive Diagonal Torus) NC論2
Hypermesh (spaninng-bus hypercube) Single or multiple buses NC論2
Base-m n-cube (hyper-crossbar) 770 777 070 077 707 000 007 8x8 crossbar Base-8 3-cube (Toshiba Prodigy) NC論2
Diameter and degrees (1/2) 2D mesh 2D torus 3D torus binary n-cube #node N N N = 2n Diameter 2√N √N log N degree 4 6 3 NC論2
Diameter and degrees (2/2) Base-m n-cube CCC Binary tree ring #node N = mn N = n2n N Diameter logm N 3n/2 2log N N/2 degree 3 2 3 NC論2