Supporting PGAS Models (UPC and OpenSHMEM) on MIC Clusters

Transcription

1 Supporting PGAS Models (UPC and OpenSHMEM) on MIC Clusters Presentation at IXPUG Meeting, July 2014 by Dhabaleswar K. (DK) Panda The Ohio State University Khaled Hamidouche The Ohio State University

2 Parallel Programming Models Overview P1 P2 P3 P1 P2 P3 P1 P2 P3 Shared Memory Memory Memory Memory Logical shared memory Memory Memory Memory Shared Memory Model DSM Distributed Memory Model MPI (Message Passing Interface) Partitioned Global Address Space (PGAS) Global Arrays, UPC, Chapel, X10, CAF, Programming models provide abstract machine models Models can be mapped on different types of systems e.g. Distributed Shared Memory (DSM), MPI within a node, etc. Additionally, OpenMP can be used to parallelize computation within the node Each model has strengths and drawbacks - suite different problems or applications IXPUG14-OSU-PGAS 2

3 Partitioned Global Address Space (PGAS) Models Key features - Simple shared memory abstractions - Light weight one-sided communication - Easier to express irregular communication Different approaches to PGAS - Languages Unified Parallel C (UPC) Co-Array Fortran (CAF) X10 Chapel - Libraries OpenSHMEM Global Arrays IXPUG14-OSU-PGAS 3

4 SHMEM SHMEM: Symmetric Hierarchical MEMory library One-sided communications library had been around for a while Similar to MPI, processes are called PEs, data movement is explicit through library calls Provides globally addressable memory using symmetric memory objects (more in later slides) Library routines for Symmetric object creation and management One-sided data movement Atomics Collectives Synchronization IXPUG14-OSU-PGAS 4

5 OpenSHMEM SHMEM implementations Cray SHMEM, SGI SHMEM, Quadrics SHMEM, HP SHMEM, GSHMEM Subtle differences in API, across versions example: SGI SHMEM Quadrics SHMEM Cray SHMEM Initialization start_pes(0) shmem_init start_pes Process ID _my_pe my_pe shmem_my_pe Made application codes non-portable OpenSHMEM is an effort to address this: A new, open specification to consolidate the various extant SHMEM versions into a widely accepted standard. OpenSHMEM Specification v1.0 by University of Houston and Oak Ridge National Lab SGI SHMEM is the baseline IXPUG14-OSU-PGAS 5

6 Compiler-based: Unified Parallel C UPC: a parallel extension to the C standard UPC Specifications and Standards: Introduction to UPC and Language Specification, 1999 UPC Language Specifications, v1.0, Feb 2001 UPC Language Specifications, v1.1.1, Sep 2004 UPC Language Specifications, v1.2, June 2005 UPC Language Specifications, v1.3, Nov 2013 UPC Consortium Academic Institutions: GWU, MTU, UCB, U. Florida, U. Houston, U. Maryland Government Institutions: ARSC, IDA, LBNL, SNL, US DOE Commercial Institutions: HP, Cray, Intrepid Technology, IBM, Supported by several UPC compilers Vendor-based commercial UPC compilers: HP UPC, Cray UPC, SGI UPC Open-source UPC compilers: Berkeley UPC, GCC UPC, Michigan Tech MuPC Aims for: high performance, coding efficiency, irregular applications, IXPUG14-OSU-PGAS 6

7 MPI+PGAS for Exascale Architectures and Applications Hierarchical architectures with multiple address spaces (MPI + PGAS) Model MPI across address spaces PGAS within an address space MPI is good at moving data between address spaces Within an address space, MPI can interoperate with other shared memory programming models Applications can have kernels with different communication patterns Can benefit from different models Re-writing complete applications can be a huge effort Port critical kernels to the desired model instead IXPUG14-OSU-PGAS 7

8 Hybrid (MPI+PGAS) Programming Application sub-kernels can be re-written in MPI/PGAS based on communication characteristics Benefits: Best of Distributed Computing Model Best of Shared Memory Computing Model Exascale Roadmap*: Hybrid Programming is a practical way to program exascale systems HPC Application Kernel 1 MPI Kernel 2 PGAS MPI Kernel 3 MPI Kernel N PGAS MPI * The International Exascale Software Roadmap, Dongarra, J., Beckman, P. et al., Volume 25, Number 1, 2011, International Journal of High Performance Computer Applications, ISSN IXPUG14-OSU-PGAS 8

9 MVAPICH2/MVAPICH2-X Software High Performance open-source MPI Library for InfiniBand, 10Gig/iWARP, and RDMA over Converged Enhanced Ethernet (RoCE) MVAPICH (MPI-1), MVAPICH2 (MPI-2.2 and MPI-3.0), Available since 2002 MVAPICH2-X (MPI + PGAS), Available since 2012 Support for GPGPUs and MIC Used by more than 2,150 organizations (HPC Centers, Industry and Universities) in 72 countries More than 218,000 downloads from OSU site directly Empowering many TOP500 clusters 7 th ranked 519,640-core cluster (Stampede) at TACC 11th ranked 74,358-core cluster (Tsubame 2.5) at Tokyo Institute of Technology 16 th ranked 96,192-core cluster (Pleiades) at NASA and many others Available with software stacks of many IB, HSE, and server vendors including Linux Distros (RedHat and SuSE) Partner in the U.S. NSF-TACC Stampede System IXPUG14-OSU-PGAS 9

10 MVAPICH2-X for Hybrid MPI + PGAS Applications MPI Applications, OpenSHMEM Applications, UPC Applications, Hybrid (MPI + PGAS) Applications OpenSHMEM Calls UPC Calls MPI Calls Unified MVAPICH2-X Runtime InfiniBand, RoCE, iwarp Unified communication runtime for MPI, UPC, OpenSHMEM available with MVAPICH2-X 1.9 onwards! Feature Highlights Supports MPI(+OpenMP), OpenSHMEM, UPC, MPI(+OpenMP) + OpenSHMEM, MPI(+OpenMP) + UPC MPI-3 compliant, OpenSHMEM v1.0 standard compliant, UPC v1.2 standard compliant (with initial support for UPC 1.3) Scalable Inter-node and intra-node communication point-to-point and collectives IXPUG14-OSU-PGAS 10

11 Outline OpenSHMEM and UPC for Host Hybrid MPI + PGAS (OpenSHMEM and UPC) for Host OpenSHMEM for MIC UPC for MIC IXPUG14-OSU-PGAS 11

12 OpenSHMEM Design in MVAPICH2-X Memory Management OpenSHMEM API Data Atomics Movement Collectives Minimal Set of Internal API Communication API Symmetric Memory Management API Active Messages One-sided Operations MVAPICH2-X Runtime Remote Atomic Ops Enhanced Registration Cache InfiniBand, RoCE, iwarp OpenSHMEM Stack based on OpenSHMEM Reference Implementation OpenSHMEM Communication over MVAPICH2-X Runtime Uses active messages, atomic and one-sided operations and remote registration cache J. Jose, K. Kandalla, M. Luo and D. K. Panda, Supporting Hybrid MPI and OpenSHMEM over InfiniBand: Design and Performance Evaluation, Int'l Conference on Parallel Processing (ICPP '12), September IXPUG14-OSU-PGAS 12

13 OpenSHMEM Collective Communication Performance Reduce (1,024 processes) Broadcast (1,024 processes) Time (us) X MV2X-SHMEM Scalable-SHMEM OMPI-SHMEM Time (us) X Time (us) K 4K 16K 64K Message Size Collect (1,024 processes) 180X IXPUG14-OSU-PGAS Message Size Time (us) K 4K 16K 64K 256K Message Size Barrier 3X No. of Processes 13

14 Time (s) OpenSHMEM Application Evaluation UH-SHMEM OMPI-SHMEM (FCA) Scalable-SHMEM (FCA) MV2X-SHMEM Heat Image Number of Processes Time (s) UH-SHMEM Daxpy OMPI-SHMEM (FCA) Scalable-SHMEM (FCA) MV2X-SHMEM Number of Processes Improved performance for OMPI-SHMEM and Scalable-SHMEM with FCA Execution time for 2DHeat Image at 512 processes (sec): - UH-SHMEM 523, OMPI-SHMEM 214, Scalable-SHMEM 193, MV2X- SHMEM 169 Execution time for DAXPY at 512 processes (sec): - UH-SHMEM 57, OMPI-SHMEM 56, Scalable-SHMEM 9.2, MV2X- SHMEM 5.2 J. Jose, J. Zhang, A. Venkatesh, S. Potluri, and D. K. Panda, A Comprehensive Performance Evaluation of OpenSHMEM Libraries on InfiniBand Clusters, OpenSHMEM Workshop (OpenSHMEM 14), March 2014 IXPUG14-OSU-PGAS 14

15 MVAPICH2-X Support for Berkeley UPC Runtime UPC Applications Compiler-generated Code Heavily based on active messages; directly on top of each individual network architectures Compiler-specific Runtime Extended APIs Core APIs GASNet High-level operations such as remote memory access and various collective operations SMP Conduit IB Conduit MPI Conduit MXM Conduit MVAPICH2-X Conduit GASNet (Global-Address Space Networking) is a language-independent, lowlevel networking layer that provides support for PGAS language Support multiple networks through different conduit: MVAPICH2-X Conduit is available in MVAPICH2-X release, which support UPC/OpenMP/MPI on InfiniBand IXPUG14-OSU-PGAS 15

16 Time (us) Time (ms) UPC Collectives Performance Broadcast (2048 processes) UPC-GASNet UPC-OSU K 2K 4K 8K 16K 32K 64K 128K 256K IXPUG14-OSU-PGAS Message Size Gather (2048 processes) UPC-GASNet UPC-OSU 25X 2X K 2K 4K 8K 16K 32K 64K 128K 256K Time (ms) Time (ms) Scatter (2048 processes) UPC-GASNet UPC-OSU K 2K 4K 8K 16K 32K 64K 128K 256K Message Size Exchange (2048 processes) Message Size Message Size J. Jose, K. Hamidouche, J. Zhang, A. Venkatesh, and D. K. Panda, Optimizing Collective Communication in UPC (HiPS 14, in association with IPDPS 14) 64 UPC-GASNet UPC-OSU 2X 35% K 2K 4K 8K 16K 32K 64K 128K 16

17 Outline OpenSHMEM and UPC for Host Hybrid MPI + PGAS (OpenSHMEM and UPC) for Host OpenSHMEM for MIC UPC for MIC IXPUG14-OSU-PGAS 17

18 Unified Runtime for Hybrid MPI + OpenSHMEM Applications Hybrid (OpenSHMEM + MPI) Applications MPI Applications, OpenSHMEM Applications, Hybrid (MPI + OpenSHMEM) Applications OpenSHMEM Calls MPI Calls OpenSHMEM Calls MPI Calls OpenSHMEM Runtime MPI Runtime MVAPICH2-X Runtime InfiniBand, RoCE, iwarp InfiniBand, RoCE, iwarp Optimal network resource usage No deadlock because of single runtime Better performance J. Jose, K. Kandalla, M. Luo and D. K. Panda, Supporting Hybrid MPI and OpenSHMEM over InfiniBand: Design and Performance Evaluation, Int'l Conference on Parallel Processing (ICPP '12), September IXPUG14-OSU-PGAS 18

19 Hybrid MPI+OpenSHMEM Graph500 Design Time (s) MPI-Simple MPI-CSC MPI-CSR Execution Time Hybrid (MPI+OpenSHMEM) 7.6X 4K 8K 16K No. of Processes 13X Performance of Hybrid (MPI+OpenSHMEM) Graph500 Design 8,192 processes - 2.4X improvement over MPI-CSR - 7.6X improvement over MPI-Simple 16,384 processes - 1.5X improvement over MPI-CSR - 13X improvement over MPI-Simple J. Jose, S. J. Potluri, Jose, S. K. Potluri, Tomko K. and Tomko D. K. and Panda, D. K. Designing Panda, Designing Scalable Graph500 Scalable Graph500 Benchmark Benchmark with Hybrid with MPI+OpenSHMEM Hybrid MPI+OpenSHMEM Programming Models, International Programming Supercomputing Models, Conference International (ISC 13), Supercomputing June 2013 Conference (ISC 13), June 2013 J. Jose, K. Kandalla, M. Luo and D. K. Panda, Supporting Hybrid MPI and OpenSHMEM over InfiniBand: Design and Performance Evaluation, Int'l Conference on Parallel Processing (ICPP '12), September 2012 IXPUG14-OSU-PGAS 19 Billions of Traversed Edges Per Second (TEPS) Billions of Traversed Edges Per Second (TEPS) Strong Scalability MPI-Simple MPI-CSC MPI-CSR Hybrid(MPI+OpenSHMEM) # of Processes Weak Scalability MPI-Simple MPI-CSC MPI-CSR Hybrid(MPI+OpenSHMEM) Scale

20 Hybrid MPI+UPC NAS-FT Time (s) % B-64 C-64 B-128 C-128 NAS Problem Size System Size Modified NAS FT UPC all-to-all pattern using MPI_Alltoall Truly hybrid program For FT (Class C, 128 processes) 34% improvement over UPC (GASNet) 30% improvement over UPC (MV2-X) UPC (GASNet) UPC (MV2-X) MPI+UPC (MV2-X) J. J. Jose, M. Luo, S. S. Sur Sur and and D. D. K. K. Panda, Panda, Unifying Unifying UPC and UPC MPI and Runtimes: MPI Runtimes: Experience Experience with MVAPICH, with Int'l MVAPICH, Conference Fourth on Conference Partitioned on Global Partitioned Address Space Global Programming Address Space Models Programming (PGAS) 2010 Model (PGAS '10), October 2010 IXPUG14-OSU-PGAS 20

21 Outline OpenSHMEM and UPC for Host Hybrid MPI + PGAS (OpenSHMEM and UPC) for Host OpenSHMEM for MIC UPC for MIC IXPUG14-OSU-PGAS 21

22 HPC Applications on MIC Clusters Flexibility in launching HPC jobs on Xeon Phi Clusters Xeon Xeon Phi Multi-core Centric Host-only OpenSHMEM Program Offload (/reverse Offload) OpenSHMEM Program Offloaded Computation Symmetric OpenSHMEM Program OpenSHMEM Program Coprocessor-only OpenSHMEM Program Many-core Centric IXPUG14-OSU-PGAS 22

23 Data Movement on Intel Xeon Phi Clusters CPU PCIe MIC MIC Node 0 Node 1 QPI MIC CPU CPU IB MIC OpenSHMEM Process 1. Intra-Socket 2. Inter-Socket 3. Inter-Node 4. Intra-MIC 5. Intra-Socket MIC-MIC 6. Inter-Socket MIC-MIC 7. Inter-Node MIC-MIC 8. Intra-Socket MIC-Host 9. Inter-Socket MIC-Host 10. Inter-Node MIC-Host 11. Inter-Node MIC-MIC with IB on remote socket and more... Critical for runtimes to optimize data movement, hiding the complexity IXPUG14-OSU-PGAS 23

24 Need for Non-Uniform Memory Allocation in OpenSHMEM MIC cores have limited memory per core Memory per core OpenSHMEM relies on symmetric memory, allocated using shmalloc() shmalloc() allocates same amount of memory on all PEs For applications running in symmetric mode, this limits the total heap size Similar issues for applications (even host-only) with memory load imbalance (Graph500, Out-of-Core Sort, etc.) How to allocate different amounts of memory on host and MIC cores, and still be able to communicate? IXPUG14-OSU-PGAS 24

25 OpenSHMEM Design for MIC Clusters Non-Uniform Memory Allocation: Team-based Memory Allocation (Proposed Extensions) Address Structure for non-uniform memory allocations IXPUG14-OSU-PGAS 25

26 Proxy-based designs for OpenSHMEM (1) IB REQ HOST1 HOST2 HOST1 HOST2 (2) SCIF Read (3) IB FIN (2) IB Write (1) IB REQ (2) IB Write (2) SCIF Read MIC1 H C A H C A MIC2 MIC1 H C A H C A MIC2 OpenSHMEM Put using Active Proxy (3) IB FIN OpenSHMEM Get using Active Proxy Current generation architectures impose limitations on read bandwidth when HCA reads from MIC memory Impacts both put and get operation performance Solution: Pipelined data transfer by proxy running on host using IB and SCIF channels Improves latency and bandwidth! IXPUG14-OSU-PGAS 26

27 OpenSHMEM Put/Get Performance Latency (us) MV2X-Def MV2X-Opt K 4K 16K 64K 256K 1M 4M K 4K 16K 64K 256K 1M 4M Message Size OpenSHMEM Put Latency MV2X-Def 3500 MV2X-Opt X 4.6X 2500 Latency (us) Message Size OpenSHMEM Get Latency Proxy-based designs alleviate hardware limitations Put Latency of 4M message: Default: 3911us, Optimized: 838us Get Latency of 4M message: Default: 3889us, Optimized: 837us IXPUG14-OSU-PGAS 27

28 OpenSHMEM Collectives Performance Latency (us) MV2X-Def MV2X-Opt K 4K 16K 64K 256K Message Size Broadcast Latency (1,024 processes) 7000 MV2X-Def MV2X-Opt 2.6X 1.9X 4000 Latency (us) K 2K 4K 8K 16K 32K 64K 128K Message Size Reduce Latency (1,024 processes) Optimized designs for OpenSHMEM collective operations Broadcast Latency of 256KB message at 1,024 processes: Default: 5955us, Optimized: 2268us Reduce Latency of 256KB message at 1,024 processes: Default: 6581us, Optimized: 3294us IXPUG14-OSU-PGAS 28

29 Performance Evaluations using Graph500 Execution Time (s) MV2X-Def MV2X-Opt Execution Time (s) MV2X-Def MV2X-Opt 28% Number of Processes Number of Processes Native Mode (8 procs/mic) Symmetric Mode (16 Host+16MIC) Graph500 Execution Time (Native Mode): At 512 processes, Default: 5.17s, Optimized: 4.96s Performance Improvement from MIC-aware collectives design Graph500 Execution Time (Symmetric Mode): At 1,024 processes, Default: 15.91s, Optimized: 12.41s Performance Improvement from MIC-aware collectives and proxy-based designs IXPUG14-OSU-PGAS 29

30 Graph500 Evaluations with Extensions Execution Time (s) Number of Processes Redesigned Graph500 using MIC to overlap computation/communication Data Transfer to MIC memory; MIC cores pre-processes received data Host processes traverses vertices, and sends out new vertices Graph500 Execution time at 1,024 processes: Host-Only:.33s, Host+MIC with Extensions:.26s Magnitudes of improvement compared to default symmetric mode Host Host+MIC (extensions) 26% Default Symmetric Mode: 12.1s, Host+MIC Extensions: 0.16s J. Jose, K. Hamidouche, X. Lu, S. Potluri, J. Zhang, K. Tomko and D. K. Panda, High Performance OpenSHMEM for Intel MIC Clusters: Extensions, Runtime Designs and Application Co-Design, IEEE International Conference on Cluster Computing (CLUSTER '14) (Best Paper Nominee) IXPUG14-OSU-PGAS 30

31 Outline OpenSHMEM and UPC for Host Hybrid MPI + PGAS (OpenSHMEM and UPC) for Host OpenSHMEM for MIC UPC for MIC IXPUG14-OSU-PGAS 31

32 HPC Applications on MIC Clusters Flexibility in launching HPC jobs on Xeon Phi Clusters Xeon Xeon Phi Multi-core Centric Host-only UPC Program Offload (/reverse Offload) UPC Program Offloaded Computation Symmetric UPC Program UPC Program Coprocessor-only UPC Program Many-core Centric IXPUG14-OSU-PGAS 32

33 UPC on MIC clusters Process Model Communication via shared memory single copy/double copy One connection per process Drawbacks: Communication Overheads, Network Connections Thread Model Direct access to shared array: same memory space No extra copy No shared memory mapping entries Drawback: sharing of network connections Our work on multi-endpoint addresses this issue M. Luo, J. Jose, S. Sur and D. K. Panda, Multi-threaded UPC Runtime with Network Endpoints: Design Alternatives and Evaluation on Multi-core Architectures, Int'l Conference on High Performance Computing (HiPC '11), Dec IXPUG14-OSU-PGAS 33

34 UPC Runtime on MIC: Remote Memory Access between Host and MIC Original all-to-all connection mode Leader-to-all connection mode CPU Global Memory Region on MIC CPU Global Memory Region on MIC CPU CPU Pin-down the whole global memory region CPU CPU CPU Intel Xeon Intel Xeon Phi Co-Processors CPU Intel Xeon Intel Xeon Phi Co-Processors Original connection mode: Each UPC thread pins down the global memory region that has affinity with this UPC thread; Establish all-to-all connections between every pair of UPC threads between MIC and host Leader-to-all connection mode: Only one UPC thread pinned down the whole global memory space on MIC; all UPC threads on the host will get access with the leader UPC thread Connections: N MIC *N HOST -> N MIC +N HOST IXPUG14-OSU-PGAS 34

35 UPC Runtime on MIC: Remote Memory Access between Host and MIC CPU Global Memory Region on MIC CPU CPU CPU Intel Xeon Intel Xeon Phi Co-Processors Process-based Runtime: The whole global memory region need to be mapped as shared memory region, such as PSHM Thread-based Runtime: As the threads share the same memory space, the leader can access other global memory region on MIC without mapping to shared memory IXPUG14-OSU-PGAS 35

36 Application Benchmark Evaluation for Native Mode native-host native-mic CG.B EP.B FT.B IS.B MG.B RandomAccess Host and MIC are occupied with 16 and 60 UPC threads, respectively MIC performs 80%, 67%, and 54% as 16-CPU host for MG, EP, and FT, respectively For communication-intensive benchmarks CG and IS, MIC spends 7x and 3x execution times Default applications without modification no representation of optimal performance 2.6X M. Luo, M. Li, A. Venkatesh, X. Lu and D. K. Panda, UPC on MIC: Early Experiences with Native and Symmetric Modes, Int'l Conference on Partitioned Global Address Space Programming Models (PGAS '13), October UTS 4X IXPUG14-OSU-PGAS 36

37 Concluding Remarks HPC Systems are evolving to support growing needs of exascale applications Hybrid programming (MPI + PGAS) is a practical way to program exascale systems Presented designs to demonstrate performance potential of PGAS and hybrid MPI+PGAS models MIC support for OpenSHMEM and UPC will be available in future MVAPICH2-X releases IXPUG14-OSU-PGAS 37

38 Funding Acknowledgments Funding Support by Equipment Support by IXPUG14-OSU-PGAS 38

39 Personnel Acknowledgments Current Students S. Chakraborthy (Ph.D.) N. Islam (Ph.D.) J. Jose (Ph.D.) M. Li (Ph.D.) R. Rajachandrasekhar (Ph.D.) M. Rahman (Ph.D.) D. Shankar (Ph.D.) R. Shir (Ph.D.) A. Venkatesh (Ph.D.) J. Zhang (Ph.D.) Current Senior Research Associates H. Subramoni X. Lu Current Post-Docs Current Programmers K. Hamidouche M. Arnold J. Lin J. Perkins Past Students P. Balaji (Ph.D.) D. Buntinas (Ph.D.) S. Bhagvat (M.S.) L. Chai (Ph.D.) B. Chandrasekharan (M.S.) N. Dandapanthula (M.S.) V. Dhanraj (M.S.) T. Gangadharappa (M.S.) K. Gopalakrishnan (M.S.) W. Huang (Ph.D.) W. Jiang (M.S.) S. Kini (M.S.) M. Koop (Ph.D.) R. Kumar (M.S.) S. Krishnamoorthy (M.S.) K. Kandalla (Ph.D.) P. Lai (M.S.) J. Liu (Ph.D.) M. Luo (Ph.D.) A. Mamidala (Ph.D.) G. Marsh (M.S.) V. Meshram (M.S.) S. Naravula (Ph.D.) R. Noronha (Ph.D.) X. Ouyang (Ph.D.) S. Pai (M.S.) S. Potluri (Ph.D.) G. Santhanaraman (Ph.D.) A. Singh (Ph.D.) J. Sridhar (M.S.) S. Sur (Ph.D.) H. Subramoni (Ph.D.) K. Vaidyanathan (Ph.D.) A. Vishnu (Ph.D.) J. Wu (Ph.D.) W. Yu (Ph.D.) Past Post-Docs H. Wang X. Besseron H.-W. Jin E. Mancini S. Marcarelli J. Vienne Past Research Scientist S. Sur Past Programmers D. Bureddy M. Luo IXPUG14-OSU-PGAS 39

40 Web Pointers MVAPICH Web Page IXPUG14-OSU-PGAS 40