The Effect of HPC Cluster Architecture on the Scalability Performance of CAE Simulations Pak Lui HPC Advisory Council June 7, 2016 1
Agenda Introduction to HPC Advisory Council Benchmark Configuration Performance Benchmark Testing/Results Summary Q&A / For More Information
The HPC Advisory Council Mission Statement World-wide HPC non-profit organization (429+ members) Bridges the gap between HPC usage and its potential Provides best practices and a support/development center Explores future technologies and future developments Leading edge solutions and technology demonstrations
HPC Advisory Council Members
HPC Advisory Council Cluster Center Dell PowerEdge R730 32-node cluster Dell PowerVault MD3420 Dell PowerVault MD3460 HP Proliant XL230a Gen9 10-node cluster HP Cluster Platform 3000SL 16-node cluster InfiniBand Storage (Lustre) Dell PowerEdge C6145 6-node cluster Dell PowerEdge R815 11-node cluster Dell PowerEdge R720xd/R720 32-node cluster Dell PowerEdge M610 38-node cluster Dell PowerEdge C6100 4-node cluster White-box InfiniBand-based Storage (Lustre)
HPC Training HPC Training Center CPUs GPUs Interconnects Clustering Storage Cables Programming Applications Network of Experts Ask the experts
Special Interest Subgroups HPC Scale Subgroup Explore usage of commodity HPC as a replacement for multi-million dollar mainframes and proprietary based supercomputers HPC Storage Subgroup Demonstrate how to build highperformance storage solutions and their affect on application performance and productivity HPC Cloud Subgroup Explore usage of HPC components as part of the creation of external/public/internal/private cloud computing environments. HPC GPU Subgroup Explore usage models of GPU components as part of next generation compute environments and potential optimizations for GPU based computing HPC Works Subgroup Provide best practices for building balanced and scalable HPC systems, performance tuning and application guidelines. HPC Music To enable HPC in music production and to develop HPC cluster solutions that further enable the future of music production
University Award Program University award program Universities / individuals are encouraged to submit proposals for advanced research Selected proposal will be provided with: Exclusive computation time on the HPC Advisory Council s Compute Center Invitation to present in one of the HPC Advisory Council s worldwide workshops Publication of the research results on the HPC Advisory Council website 2010 award winner is Dr. Xiangqian Hu, Duke University Topic: Massively Parallel Quantum Mechanical Simulations for Liquid Water 2011 award winner is Dr. Marco Aldinucci, University of Torino Topic: Effective Streaming on Multi-core by Way of the FastFlow Framework 2012 award winner is Jacob Nelson, University of Washington Runtime Support for Sparse Graph Applications 2013 award winner is Antonis Karalis Topic: Music Production using HPC 2014 award winner is Antonis Karalis Topic: Music Production using HPC 2015 award winner is Christian Kniep Topic: Dockers To submit a proposal please check the HPC Advisory Council web site
Exploring All Platforms X86, Power, GPU, FPGA and ARM based Platforms x86 Power GPU FPGA ARM
158+ Applications Best Practices Published Abaqus CPMD LS-DYNA MILC AcuSolve Dacapo minife OpenMX Amber Desmond MILC PARATEC AMG DL-POLY MSC Nastran PFA AMR Eclipse MR Bayes PFLOTRAN ABySS FLOW-3D MM5 Quantum ANSYS CFX GADGET-2 MPQC ESPRESSO ANSYS FLUENT GROMACS NAMD RADIOSS ANSYS Mechanics Himeno Nekbone SPECFEM3D BQCD HOOMD-blue NEMO WRF CCSM HYCOM NWChem CESM ICON Octopus COSMO Lattice QCD OpenAtom CP2K LAMMPS OpenFOAM For more information, visit: http://www.hpcadvisorycouncil.com/best_practices.php
HPCAC - ISC 16 Student Cluster Competition University-based teams to compete and demonstrate the incredible capabilities of state-of- the-art HPC systems and applications on the International Super Computing HPC (ISC HPC) show-floor The Student Cluster Challenge is designed to introduce the next generation of students to the high performance computing world and community
ISC'15 Student Cluster Competition Award Ceremony
ISC'16 Student Cluster Competition Teams
Getting Ready to 2016 Student Cluster Competition
HPCAC Conferences 2015 Conferences
2016 HPC Advisory Council Conferences Introduction HPC Advisory Council (HPCAC) 429+ members, http://www.hpcadvisorycouncil.com/ Application best practices, case studies Benchmarking center with remote access for users World-wide workshops Value add for your customers to stay up to date and in tune to HPC market 2016 Conferences USA (Stanford University) February 24-26 Switzerland (CSCS) March 21-23 Mexico TBD Spain (BSC) September 21 China (HPC China) October 26 For more information www.hpcadvisorycouncil.com info@hpcadvisorycouncil.com
RADIOSS by Altair Altair RADIOSS Structural analysis solver for highly non-linear problems under dynamic loadings Consists of features for: multiphysics simulation and advanced materials such as composites Highly differentiated for Scalability, Quality and Robustness RADIOSS is used across all industry worldwide Improves crashworthiness, safety, and manufacturability of structural designs RADIOSS has established itself as an industry standard for automotive crash and impact analysis for over 20 years
Test Cluster Configuration Dell PowerEdge R730 32-node (896-core) Thor cluster Dual-Socket 14-core Intel E5-2697v3 @ 2.60 GHz CPUs (Turbo on, Max Perf set in BIOS) OS: RHEL 6.5, OFED MLNX_OFED_LINUX-2.4-1.0.5 InfiniBand SW stack Memory: 64GB memory, DDR3 2133 MHz Hard Drives: 1TB 7.2 RPM SATA 2.5 Mellanox ConnectX-4 EDR 100Gb/s InfiniBand VPI adapters Mellanox Switch-IB SB7700 100Gb/s InfiniBand VPI switch Mellanox ConnectX-3 40/56Gb/s QDR/FDR InfiniBand VPI adapters Mellanox SwitchX SX6036 56Gb/s FDR InfiniBand VPI switch MPI: Intel MPI 5.0.2, Mellanox HPC-X v1.2.0 Application: Altair RADIOSS 13.0 Benchmark datasets: Neon benchmarks: 1 million elements (8ms, Double Precision), unless otherwise stated
PowerEdge R730 Massive flexibility for data intensive operations Performance and efficiency Intelligent hardware-driven systems management with extensive power management features Innovative tools including automation for parts replacement and lifecycle manageability Broad choice of networking technologies from GbE to IB Built in redundancy with hot plug and swappable PSU, HDDs and fans Benefits Designed for performance workloads from big data analytics, distributed storage or distributed computing where local storage is key to classic HPC and large scale hosting environments High performance scale-out compute and low cost dense storage in one package Hardware Capabilities Flexible compute platform with dense storage capacity 2S/2U server, 7 PCIe slots Large memory footprint (Up to 1.5TB / 24 DIMMs) High I/O performance and optional storage configurations HDD: SAS, SATA, nearline SAS; SSD: SAS, SATA 16 x 2.5 up to 29TB via 1.8TB hot-plug SAS hard drives 8 x 3.5 up to 64TB via 8TB hot-plug nearline SAS hard drives
RADIOSS Performance Interconnect (MPP) EDR InfiniBand provides higher scalability than Ethernet 70 times better performance than 1GbE at 16 nodes / 448 cores 4.8x better performance than 10GbE at 16 nodes / cores Ethernet solutions does not scale beyond 4 nodes with pure MPI 70x 4.8x Intel MPI Higher is better 28 Processes/Node
RADIOSS Profiling % Time Spent on MPI RADIOSS utilizes point-to-point communications in most data transfers The most time MPI consuming calls is MPI_Waitany() and MPI_Wait() MPI_Recv(55%), MPI_Waitany(23%), MPI_Allreduce(13%) MPP Mode 28 Processes/Node
RADIOSS Performance Interconnect (MPP) EDR InfiniBand provides better scalability performance EDR IB improves over QDR IB by 28% at 16 nodes / 448 cores EDR InfiniBand outperforms FDR InfiniBand by 25% at 16 nodes 28% 25% Higher is better 28 Processes/Node
RADIOSS Performance CPU Cores Running more cores per node generally improves overall performance Seen improvement of 18% from 20 to 28 cores per node at 8 nodes Improvement seems not as consistent at higher node counts Guideline: Most optimal workload distribution is 4000 elements/process For test case of 1 million elements, most optimal core sizes is ~256 cores 4000 elements per process should provides sufficient workload for each process Hybrid MPP (HMPP) provides way to achieve additional scalability on more CPUs 18% 6% Higher is better Intel MPI
RADIOSS Performance IMPI Tuning (MPP) Tuning Intel MPI collective algorithm can improve performance MPI profile shows about 20% of runtime spent on MPI_Allreduce communications Default algorithm in Intel MPI is Recursive Doubling The default algorithm is the best among all tested for MPP Intel MPI Higher is better 28 Processes/Node
RADIOSS Performance Hybrid MPP version Enabling Hybrid MPP mode unlocks the RADIOSS scalability At larger scale, productivity improves as more threads involves As more threads involved, amount of communications by processes are reduced At 32 nodes/896 cores, best configuration is 1 process per socket to spawn 14 threads each 28 threads/1 PPN is not advised due to breach of data locality across different CPU socket The following environment setting and tuned flags are used for Intel MPI: I_MPI_PIN_DOMAIN auto I_MPI_ADJUST_ALLREDUCE 5 I_MPI_ADJUST_BCAST 1 KMP_AFFINITY compact KMP_STACKSIZE 400m ulimit -s unlimited 3.7x 32% 70% Intel MPI EDR InfiniBand
RADIOSS Profiling Number of MPI Calls For MPP utilizes most non-blocking calls for communications MPI_Recv, MPI_Waitany, MPI_Allreduce are used most of the time For HMPP, communication behavior has changed Higher time percentage in MPI_Waitany, MPI_Allreduce, and MPI_Recv MPI Communication behavior changed from previous RADIOSS version Most likely due to more CPU cores available on the current cluster MPP, 28PPN HMPP, 2PPN / 14 Threads
RADIOSS Profiling MPI Message Sizes The most time consuming MPI communications are: MPI_Recv: Messages concentrated at 640B, 1KB, 320B, 1280B MPI_Waitany: Messages are: 48B, 8B, 384B MPI_Allreduce: Most message sizes appears at 80B MPP, 28PPN HMPP, 2PPN / 14 Threads Pure MPP 28 Processes/Node
RADIOSS Performance Intel MPI Tuning (DP) For Hybrid MPP DP, tuning MPI_Allreduce shows more gain than MPP For DAPL provider, Binomial gather+scatter #5 improved perf by 27% over default For OFA provider, tuned MPI_Allreduce algorithm improves by 44% over default Both OFA and DAPL improved by tuning I_MPI_ADJUST_ALLREDUCE=5 Flags for OFA: I_MPI_OFA_USE_XRC 1. For DAPL: ofa-v2-mlx5_0-1u provider 27% 44% Intel MPI Higher is better 2 PPN / 14 OpenMP
RADIOSS Performance Interconnect (HMPP) EDR InfiniBand provides better scalability performance than Ethernet 214% better performance than 1GbE at 16 nodes 104% better performance than 10GbE at 16 nodes InfiniBand typically outperforms other interconnect in collective operations 214% 104% Intel MPI Higher is better 2 PPN / 14 OpenMP
RADIOSS Performance Interconnect (HMPP) EDR InfiniBand provides better scalability performance than FDR IB EDR IB outperforms FDR IB by 27% at 32 nodes Improvement for EDR InfiniBand occurs at high node count 27% Intel MPI Higher is better 2 PPN / 14 OpenMP
RADIOSS Performance System Generations Intel E5-2680v3 (Haswell) cluster outperforms prior generations Performs faster by 100% vs Jupiter, by 238% vs Janus at 16 nodes System components used: Thor: 2-socket Intel E5-2680v3@2.6GHz, 2133MHz DIMMs, EDR IB, v13.0 Jupiter: 2-socket Intel E5-2680@2.7GHz, 1600MHz DIMMs, FDR IB, v12.0 Janus: 2-socket Intel X5670@2.93GHz, 1333MHz DIMMs, QDR IB, v12.0 238% 100% Single Precision
RADIOSS Summary RADIOSS is designed to perform at scale in HPC environment Shows excellent scalability over 896 cores/32 nodes and beyond with Hybrid MPP Hybrid MPP version enhanced RADIOSS scalability 2 MPI processes per node (or 1 MPI process per socket), 14 threads each Additional CPU cores generally accelerating time to solution performance Network and MPI Tuning EDR IB outperforms other Ethernet in scalability EDR IB delivers higher scalability performance than FDR/QDR IB Tuning environment/parameters to maximize performance Tuning MPI collective ops helps RADIOSS to achieve even better scalability
STAR-CCM+ STAR-CCM+ An engineering process-oriented CFD tool Client-server architecture, object-oriented programming Delivers the entire CFD process in a single integrated software environment Developed by CD-adapco
Test Cluster Configuration Dell PowerEdge R730 32-node (896-core) Thor cluster Dual-Socket 14-Core Intel E5-2697v3 @ 2.60 GHz CPUs BIOS: Maximum Performance, Home Snoop Memory: 64GB memory, DDR4 2133 MHz (Snoop Mode: Home Snoop) OS: RHEL 6.5, MLNX_OFED_LINUX-3.0-1.0.1 InfiniBand SW stack Hard Drives: 2x 1TB 7.2 RPM SATA 2.5 on RAID 1 Mellanox ConnectX-4 EDR 100Gb/s InfiniBand Adapters Mellanox Switch-IB SB7700 36-port EDR 100Gb/s InfiniBand Switch Mellanox ConnectX-3 FDR VPI InfiniBand and 40Gb/s Ethernet Adapters Mellanox SwitchX-2 SX6036 36-port 56Gb/s FDR InfiniBand / VPI Ethernet Switch Dell InfiniBand-Based Lustre Storage based on Dell PowerVault MD3460 and Dell PowerVault MD3420 MPI: Platform MPI 9.1.2 Application: STAR-CCM+ 10.02.012 Benchmarks: lemans_poly_17m civil_trim_20m reactor_9m, LeMans_100M.amg
STAR-CCM+ Performance Network Interconnects EDR InfiniBand delivers superior scalability in application performance IB delivers 66% higher performance than 40GbE, 88% higher than 10GbE at 32 nodes Scalability stops beyond 4 nodes for 1GbE; scalability is limited for 10/40GbE Input data: Lemans_poly_17m: A race car model with 17 million cells 88%66% 748% Higher is better 28 MPI Processes / Node
STAR-CCM+ Performance Network Interconnects EDR InfiniBand delivers superior scalability in application performance EDR IB provides 177 higher performance than 40GbE, 194% than 40GbE at 32 nodes InfiniBand demonstrates continuous performance gain at scale Input data: reactor_9m: A reactor model with 9 million cells 194% 177% Higher is better 28 MPI Processes / Node
STAR-CCM+ Profiling % of MPI Calls For the most time consuming MPI calls: Lemans_17m: 55% MPI_Allreduce, 23% MPI_Waitany, 7% MPI_Bcast, 7% MPI_Recv Reactor_9m: 59% MPI_Allreduce, 21% MPI_Waitany, 7% MPI_Recv, 4% MPI_Bcast MPI as a percentage in wall clock times: Lemans_17m: 12% MPI_Allreduce, 5% MPI_Waitany, 2% MPI_Bcast, 2% MPI_Recv Reactor_9m: 15% MPI_Allreduce, 5% MPI_Waitany, 2% MPI_Recv, 1% MPI_Bcast lemans_17m 32 nodes / 896 Processes reactor_9m 32 Nodes / 896 Processes
STAR-CCM+ Profiling MPI Message Size Distribution For the most time consuming MPI calls Lemans_17m: MPI_Allreduce 4B (30%), 16B (19%), 8B (6%), MPI_Bcast 4B (4%) Reactor_9m: MPI_Allreduce 16B (35%), 4B (15%), 8B (8%), MPI_Bcast 1B (4%) lemans_17m 32 nodes / 896 Processes reactor_9m 32 Nodes / 896 Processes
STAR-CCM+ Profiling Time Spent in MPI Majority of the MPI time is spent on MPI collective Operations and nonblocking communications Heavy use of MPI collective operations (MPI_Allreduce, MPI_Bcast) and MPI_Waitany Some node imbalances characteristics shown on both input dataset Some processes appeared to take more time in communications, in MPI_Allreduce lemans_17m 32 nodes / 896 Processes reactor_9m 32 Nodes / 896 Processes
STAR-CCM+ Performance Scalability Speedup EDR InfiniBand demonstrates linear scaling for STAR-CCM+ STAR-CCM+ is able to achieve linear scaling with EDR InfiniBand Other interconnects only provided limited scalability As demonstrated in previous slides Higher is better 28 MPI Processes / Node
STAR-CCM+ Performance System Generations Current system generations of HW & SW configuration outperform prior generations Current Haswell systems outperformed Ivy Bridge by 38%, Sandy Bridge by 149%, Westmere by 409% Dramatic performance benefit due to better system architecture in compute and network scalability System components used: Haswell: 2-socket 14-core E5-2697v3@2.6GHz, DDR4 2133MHz DIMMs, ConnectX-4 EDR InfiniBand, v10.02.012 Ivy Bridge: 2-socket 10-core E5-2680v2@2.8GHz, DDR3 1600MHz DIMMs, Connect-IB FDR InfiniBand, v9.02.005 Sandy Bridge: 2-socket 8-core E5-2680@2.7GHz, DDR3 1600MHz DIMMs, ConnectX-3 FDR InfiniBand, v7.02.008 Westmere: 2-socket 6-core x5670@2.93ghz, DDR3 1333MHz DIMMs, ConnectX-2 QDR InfiniBand, v5.04.006 409% 149% 38% Higher is better
STAR-CCM+ Summary Compute: cluster of the current generation outperforms system architecture of previous generations Outperformed Ivy Bridge by 38%, Sandy Bridge by 149%, Westmere by 409% Dramatic performance benefit due to better system architecture in compute and network scalability Network: EDR InfiniBand demonstrates superior scalability in STAR-CCM+ performance EDR IB provides higher performance by over 4-5 times vs 1GbE, 10GbE and 40GbE, 15% vs FDR IB at 32 nodes Lemans_17m: Scalability stops beyond 4 nodes for 1GbE; scalability is limited for 10/40 GbE Reactor_9m: EDR IB provides 177 higher performance than 40GbE, 194% than 40GbE at 32 nodes EDR InfiniBand demonstrates linear scalability in STAR- CCM+ performance on the test cases
ANSYS Fluent Computational Fluid Dynamics (CFD) is a computational technology Enables the study of the dynamics of things that flow Enable better understanding of qualitative and quantitative physical phenomena in the flow which is used to improve engineering design CFD brings together a number of different disciplines Fluid dynamics, mathematical theory of partial differential systems, computational geometry, numerical analysis, Computer science ANSYS FLUENT is a leading CFD application from ANSYS Widely used in almost every industry sector and manufactured product
Test Cluster Configuration Dell PowerEdge R730 32-node (896-core) Thor cluster Dual-Socket 14-Core Intel E5-2697v3 @ 2.60 GHz CPUs Turbo enabled (Power Management: Maximum Performance) Memory: 64GB memory, DDR4 2133 MHz (Memory Snoop: Home Snoop) OS: RHEL 6.5, MLNX_OFED_LINUX-3.0-1.0.1 InfiniBand SW stack Hard Drives: 2x 1TB 7.2 RPM SATA 2.5 on RAID 1 Mellanox Switch-IB SB7700 36-port 100Gb/s EDR InfiniBand Switch Mellanox ConnectX-4 EDR 100Gbps EDR InfiniBand Adapters Mellanox SwitchX-2 SX6036 36-port 56Gb/s FDR InfiniBand / VPI Ethernet Switch Mellanox ConnectX-3 FDR InfiniBand, 10/40GbE Ethernet VPI Adapters MPI: Mellanox HPC-X v1.2.0-326, Platform MPI 9.1 Application: ANSYS Fluent 16.1 Benchmark datasets: eddy_417k truck_111m
Fluent Performance Network Interconnects InfiniBand delivers superior scalability performance EDR InfiniBand provides higher performance than Ethernet InfiniBand delivers ~20 to 44 times higher performance and continuous scalability Ethernet performance stays flat (or stops scaling) beyond 2 nodes 20x 44x 20x Higher is better 28 MPI Processes / Node
Fluent Performance EDR vs FDR InfiniBand EDR InfiniBand delivers superior scalability in application performance As the number of nodes scales, performance gap of EDR IB becomes widen Performance advantage of EDR InfiniBand increases for larger core counts EDR InfiniBand provides 111% versus FDR InfiniBand at 16 nodes (448 cores) 111% Higher is better 28 MPI Processes / Node
Fluent Performance MPI Libraries HPC-X delivers higher scalability performance than Platform MPI by 16% Support of HPC-X on Fluent Based on the support of Open MPI on Fluent The new yalla pml reduces the overhead. Flags used for HPC-X: Tuning Parameters used: -mca coll_fca_enable 1 -mca pml yalla -map-by node -x MXM_TLS=self,shm,ud --bind-to core 16% 12% Higher is better 28 MPI Processes / Node
Fluent Profiling Time Spent by MPI Calls Different communication patterns seen depending on data Eddy_417k: Most time spent in MPI_Recv, MPI_Allreduce, MPI_Waitall Truck_111m: Most time spent in MPI_Bcast, MPI_Recv, MPI_Allreduce eddy_417k truck_111m
Fluent Profiling Time Spent by MPI Calls Different communication patterns seen with different input data Eddy_417k: Most time spent in MPI_Recv, MPI_Allreduce, MPI_Waitall Truck_111m: Most time spent in MPI_Bcast, MPI_Recv, MPI_Allreduce eddy_417k truck_111m
Fluent Profiling MPI Message Sizes The most time consuming transfers are from small messages: Eddy_417k : MPI_Recv@16B (28%wall), MPI_Allreduce@4B (14% wall), MPI_Bcast @4B (6%wall) Truck_111m: MPI_Bcast @24B (22%wall), MPI_Recv @16B (19%wall), MPI_Bcast @4B (13%wall) eddy_417k truck_111m 32 Nodes
Fluent Summary Performance Compute: Intel Haswell cluster outperforms system architecture of previous generations Haswell cluster outperforms Ivy Bridge cluster by 26%-49% at 32 node (896 cores) depending on workload Network: EDR InfiniBand delivers superior scalability in application performance EDR InfiniBand provides 20 to 44 times higher performance and more scalable compared to 1GbE/10GbE/40GbE Performance for Ethernet (1GbE/10GbE/40GbE) stays flat (or stops scaling) beyond 2 nodes EDR InfiniBand provides 111% versus FDR InfiniBand at 16nodes / 448 cores MPI: HPC-X delivers higher scalability performance than Platform MPI by 16%
Thank You! All trademarks are property of their respective owners. All information is provided As-Is without any kind of warranty. The HPC Advisory Council makes no representation to the accuracy and completeness of the information contained herein. HPC Advisory Council undertakes no duty and 52assumes no obligation to update or correct any information presented herein