Best Practices and Performance Studies for High-Performance Computing Clusters

HPC Clusters:
Best Practices and
Performance Study

Agenda
 HPC at HPE
 System Configuration and Tuning
 Best Practices for Building Applications
 Intel Xeon Processors
 Efficient Methods in Executing Applications
 Tools and Techniques for Boosting Performance
 Application Performance Highlights
 Conclusions
2

HPE’s HPC Market and Share
4
HPE/HP, 33.7%
Dell, 17.4%Lenovo, 12.9%
IBM, 3.9%
Cray,
2.4%
SGI,
2.4%Sugon
(Dawning) ,
2.7%
Fujitsu, 1.3%
NEC, 1.2%
Bull Atos,
0.9%
Other, 16.0%
Wuxi,
5.4%
IDC HPC Market Share 2016
Top500 List (ISC2016, June 2016)

System Configuration and Tuning
5

Typical BIOS Settings: Processor Options
– Hyperthreading Options Disabled : Better scaling for HPC workloads
– Processor Core Disable 0 : Enables all available cores
– Intel Turbo Boost Technology Enabled : Increases clock frequency (increase affected by factors).
– ACPI SLIT Preferences Enabled : OS can improve performance by efficient allocation of
resources among processor, memory and I/O subsystems.
– QPI Snoop Configuration Home/Early/COD : Experiment and set the right configuration for your workload.
Home: High Memory Bandwidth for average NUMA workloads.
COD: Cluster On Die, Increased Memory Bandwidth for
optimized and aggressive NUMA workloads.
Early: Decreases latency but may also decrease memory
bandwidth compared to other two modes.
6

Typical BIOS Settings: Power Settings and Management
– HPE Power Profile should be set to Maximum Performance to get best performance (idle and average power
will increase significantly).
– Custom Power Profile will reduce idle and average power at the expense of 1-2% performance reduction.
– To get highest Turbo clock speeds (when partial cores are used), use Power Savings Settings.
– For Custom Power Profile, you will have to set the following additional settings:
7

Best Practices for Building Applications
8

9
-O2 enable optimizations ( = -O, Default)
-O1 optimize for maximum speed, but disable some optimizations which increases code
size for small speed benefit
-O3 enable –O2 plus more aggressive optimizations that may or may not improve
performance for all programs.
-fast enable –O3 –ipo –static
-xHOST optimize code based on the native node used for compilation
-xAVX enable advanced vector instructions set (for Ivy Bridge performance)
-xCORE-AVX2 enable advanced vector instructions set 2 (key Haswell/Broadwell performance)
-xMIC-AVX512 enable advanced vector instructions set 512 (for future KNL/SkyLake based systems)
-mp maintain floating-point precision (disables some optimizations)
-parallel enable the auto parallelizer to generate multi-threaded code
-openmp generate multi-threaded parallel code based on OpenMP directives
-ftz enable/disable flushing denormalized results to zero
-opt-streaming-stores [always auto never] generates streaming stores
-mcmodel=[small medium large] controls the code and data memory allocation
-fp-model=[fast precise source strict] controls floating point model variation
-mkl=[parallel sequential cluster] link to Intel MKL Lib. to build optimized code.
Building Applications: Intel Compiler Flags

10
Building Applications: Compiling Thread Parallel Codes
pgf90 –mp –O3 –Mextend –Mcache_align –k8-64 ftn.f
pathf90 –mp –O3 –extend_source –march=opteron ftn.f
ifort –openmp –O3 -132 –i_dynamic –ftz –IPF_fma ftn.f
pgcc –mp –O3 –Mcache_align –k8-64 code.c
opencc –mp –O3 –march=opteron code.c
icc –openmp –O3 –i_dynamic –ftz –IPF_fma code.c
Combination Flags
Intel: -fast => -O3 –ipo –static
PGI: -fast => -O2 –Munroll –Mnoframe
Open64: -Ofast => -O3 -ipa -OPT:Ofast -fno-math-errno
Notes:
• Must compile and link with –mp / –openmp
• Aggressive optimizations may compromise accuracy

11
mpicc C compiler wrapper to build parallel code
mpiCC C++ compiler wrapper
mpif77 Fortran77 compiler wrapper
mpif90 Fortran90 compiler wrapper
mpirun command to launch mpi parallel job
Environment Variables to specify the Compilers to use:
export I_MPI_CC=icc
export I_MPI_CXX=icpc
export I_MPI_F90=ifort
export I_MPI_F77=ifort
Building Applications: Compiling MPI based Codes

12
Building Applications: Compiling MPI based Codes (Contd…)
mpif90 –O3 –Mextend –Mcache_align –k8-64 ftn.f
mpif90 –O3 –extend_source –march=opteron ftn.f
mpif90 –O2 -xHOST –fp-model strict -openmp ftn.f
mpicc –O3 –Mcache_align –k8-64 code.c
mpicc –O3 –march=opteron code.c
mpicc –O3 -xAVX2 -openmp –ftz –IPF_fma code.c
Compilers and Interface chosen depend on:
what is defined in your PATH variable
what are defined by (for Intel MPI):
• I_MPI_CC, I_MPI_CXX
• I_MPI_F77, I_MPI_F90

Complete Specifications at:
http://www.intel.com/content/www/us/en/processors/xeon/xeon-e5-v3-spec-update.html
Intel Xeon Processors: Turbo, AVX and more

Intel Xeon Processors: Turbo, AVX and more (Contd …)

Intel Xeon Processors: Turbo, AVX and more (Contd…)

Intel Xeon Processors: Turbo, AVX and more (Contd …)
Intel publishes 4 different reference frequencies for every Xeon Processor:
1. Base Frequency 2. Non-AVX Turbo 3. AVX Base Frequency 4. AVX Turbo
• Turbo clock for a given model can vary as much as 5% from one processor to another
• Four possible scenarios exist:
• Turbo=OFF and AVX=NO => Clock is set to Base frequency
• Turbo=ON and AVX=NO => Clock range will be from Base to Non-AVX Turbo
• Turbo=OFF and AVX=YES => Clock range will be from AVX Base to Base Frequency
• Turbo=ON and AVX=YES => Clock range will be from AVX Base to AVX Turbo

Efficient Methods in Executing Applications
18

Running Parallel Programs in a Cluster: Intel MPI
– Environments in General
– export PATH
– export LD_LIBRARY_PATH
– export MPI_ROOT
– export I_MPI_FABRICS= shm:dapl
– export I_MPI_DAPL_PROVIDER=ofa-v2-mlx5_0-1u
– export NPROCS=256
– export PPN=16
– export I_MPI_PIN_PROCESSOR_LIST 0-15
– export OMP_NUM_THREADS=2
– export KMP_STACKSIZE=400M
– export KMP_SCHEDULE= static,balanced
– Example Command using Intel MPI
– time mpirun -np $NPROCS -hostfile ./hosts -genvall –ppn $PPN –genv I_MPI_PIN_DOMAIN=omp
./myprogram.exe
19

Profiling a Parallel Program: Intel MPI
– Using Intel MPS (MPI Performance Snapshot)
– Set all env variables to run Intel MPI based application
– Source the following additionally:
source /opt/intel/16.0/itac/9.1.2.024/intel64/bin/mpsvars.sh –papi | vtune
– Run your application as: mpirun –mps -np $NPROCS -hostfile ./hosts ….
– Two files app_stat_xxx.txt and stats_xxx.txt will be available at the end of the job.
– Analyze the these *.txt using mps tool
– Sample data you can gather from:
– Computation Time: 174.54 sec 51.93%
– MPI Time: 161.58 sec 48.07%
– MPI Imbalance: 147.27 sec 43.81%
– OpenMP Time: 155.79 sec 46.35%
– I/O wait time: 576.47 sec ( 0.08 %)
– Using Intel MPI built-in Profiling Capabilities
– Native mode: mpirun -env I_MPI_STATS 1-4 -env I_MPI_STATS_FILE native_1to4.txt …
– IPM mode: mpirun -env I_MPI_STATS ipm -env I_MPI_STATS_FILE ipm_full.txt
20

Tools and Techniques for Boosting Performance
21

Tools, Techniques and Commands
– Check Linux pseudo files and confirm the system details
– cat /proc/cpuinfo >> provides processor details (Intel’s tool cpuinfo.x)
– cat /proc/meminfo >> shows the memory details
– /usr/ sbin /ibstat >> shows the Interconnect IB fabric details
– /sbin/sysctl –a >> shows details of system (kernel, file system etc.)
– /usr/bin/lscpu >> shows cpu details including cache sizes
– /usr/bin/lstopo >> shows the hardware topology
– /bin/uname -a >> shows the system information
– /bin/rpm –qa >> shows the list of installed products including versions
– cat /etc/redhat-release >> shows the redhat version
– /usr/sbin/dmidecode >> shows system hardware and other details (need to be root)
– /bin/ dmesg >> shows system boot-up messages
– /usr/bin/numactl >> checks or sets NUMA policy for processes or shared memory
– /usr/bin/taskset >> shows cores and memory of numa nodes of a system

Top10 Practical Tips for Boosting Performance
– Check the system details thoroughly (Never assume !)
– Choose a compiler and MPI to build your application ( All are not same !)
– Start with some basic compiler flags and try additional flags one at a time (Optimization is incremental !)
– Use the built-in libraries and tools to save time and improve performance (Libs., Tools are your friends !)
– Change compiler and MPI if your code fails to compile or run correctly (Trying to fix things is futile !)
– Test your application at every level to arrive at an optimized code (Remember the 80-20 rule !)
– Customize your runtime environment to achieve desired goals (process parallel, hybrid run etc.)
– Always place and bind the processes and threads appropriately (Life saver !)
– Gather, check and correct your runtime environment (what you get may not be what you want !)
– Profile and adjust optimization and runtime environments accordingly (Exercise caution !)
23

Application Performance Highlights
24

Application: High Performance Linpack (HPL)
Description:
Benchmark to measure floating point rates (and times) by solving a random dense linear system of
equations in double-precision.
• Originally developed by Jack Dongarra at Univ. of Tennessee.
• Used Intel optimized HPL binary for this study.
• Ran the code in hybrid mode, one MPI process per processor and each process launched threads equal to
no. of cores on the processor.
• Used explicit placing and binding of threads.
• Attempted various choices of array sizes and other parameters to identify the best performance.
• The code provides a self-check to validate the results.
Additional details at: http://icl.eecs.utk.edu/hpl/

26
Proc.Type Processor Clock (GHz) # cores/proc #cores/node TDP (Watts) L3 Cache (MB) Rpeak (GFLOPS) Rmax (GFLOPS) % Peak
IvyBridge E5-2695 v2 2.4 12 24 115 30 461 458 99.35
IvyBridge E5-2670 v2 2.5 10 20 115 25 400 389 97.25
Haswell E5-2697 v3 2.6 14 28 145 35 1160 934 80.52
Haswell E5-2698 v3 2.3 16 32 135 40 1180 943 79.92
Broadwell E5-2690 v4 2.6 14 28 135 35 1164 1073 92.18
Broadwell E5-2697 v4 2.3 18 36 145 45 1324 1186 89.58
HPL Performance from a Single Node
461
400
1160 1180 1164
1324
458
389
934 943
1073
1186
99.35 97.25
80.52 79.92
92.18 89.58
0.00
20.00
40.00
60.00
80.00
100.00
120.00
0
200
400
600
800
1000
1200
1400
E5-2695 v2 E5-2670 v2 E5-2697 v3 E5-2698 v3 E5-2690 v4 E5-2697 v4
%PEAK
GFLOPS
PROCESSOR
HPL on One Node
Rpeak (GFLOPS) Rmax (GFLOPS) % Peak

27
HPL Performance from a Haswell Cluster
81.08
79.41
77.91 78.06
77.09
75.00
76.00
77.00
78.00
79.00
80.00
81.00
82.00
0.00
20.00
40.00
60.00
80.00
100.00
120.00
140.00
8 16 40 80 120
%PEAK
GFLOPS
NO. OF NODES
HPL on a Haswell Cluster
76.21
75.95 75.95
75.74
75.65
75.49
75.55
75.00
75.20
75.40
75.60
75.80
76.00
76.20
76.40
0.00
200.00
400.00
600.00
800.00
1000.00
1200.00
1400.00
200 280 288 300 600 1000 1200
%PEAK
GFLOPS
NO. OF NODES
HPL on a Haswell Cluster
System: BL460c Gen9, Intel E5-2680 v3, 2.5 GHz, 2P/24C, 128 GB (DDR4-2133 MHz) Memory, RHEL 6.5, IB/FDR 1:1,
Intel MPI 5.0.1, Intel Composer XE 15.0.0, Turbo ON, Hyperthreading OFF

Application: High Performance Conjugate Gradient (HPCG)
Description:
Benchmark designed to create a new metric for ranking HPC systems, complementing the current HPL
benchmark. HPCG is designed to exercise computational and data access patterns that more closely match a
different and broad set of important HPC applications.
• Supports various operations in a standalone and an unified code.
• Reference implementation is written in C++ with MPI and OpenMP support.
• Driven by multigrid preconditioned conjugate gradient algorithm that exercises the key kernels on a nested set
of coarse grids.
• Unlike the HPL, HPCG can be run for predetermined time (input).
• Local domain size (input) for a node is replicated to identify a global domain resulting in near-linear speed-up.
• Performance is measured by GFLOP/s rating reported by the code.
• An Intel optimized HPCG binary was used for this benchmark study.
• Ran the HPCG binary in hybrid mode, MPI processes + OpenMP threads.
Additional details at: http://hpcg-benchmark.org/

29
High Performance Conjugate Gradient (HPCG)
100.00 97.99 97.23 99.36 97.79 97.60
93.73 92.32
84.01 84.62
75.55
0.00
20.00
40.00
60.00
80.00
100.00
120.00
0
500
1000
1500
2000
2500
3000
3500
4000
1 10 20 40 50 80 100 200 300 400 500
%SPEEDUP
GFLOPS
NO. OF NODES
HPCG on a Haswell Cluster
Blade1
Standard Distribution
6.81
6.56
6.01
5.11
6.07
4.34
5.12
0.81
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
0.00
200.00
400.00
600.00
800.00
1000.00
1200.00
1 4 8 16 24 32 48 64
%IMPROVEMENT
GFLOPS
NO OF NODES
HPCG – Haswell Vs Broadwell
Haswell - FDR Broadwell - EDR % Improvement
HSW1 and BDW1
Intel Optimized Binary
Blade1: BL460c Gen9, Intel E5-2680 v3, 2.5 GHz, 2P/24C, 128 GB (DDR4-2133 MHz) Memory, RHEL 6.5, IB/FDR 1:1, Intel MPI 5.0.1, Intel Composer XE 15.0.0, Turbo ON, Hyperthreading OFF
HSW1: XL170r Gen9, Intel E5-2698 v3, 2.3 GHz, 2P/32C, 128 GB (DDR4-2133 MHz) Memory, RHEL 6.5, IB/FDR 1:1, Intel MPI 5.0.1, Intel Composer XE 15.0.0, Turbo ON, Hyperthreading OFF
BDW1: XL170r Gen9, Intel E5-2698 v4, 2.2 GHz, 2P/40C, 128 GB (DDR4-2133 MHz) Memory, RHEL 6.6, IB/EDR 1:1, Intel MPI / Intel Compiler 2016.2.181, Turbo ON, Hyperthreading OFF

Application: Graph500
Description:
Benchmark designed to address performance of data intensive HPC applications using Graph Algorithms
• The code generates problem size with a scale (input) creating vertices equal to 10scale
• The performance is measured in TEPS (Traversed Edges Per Second).
• The median_TEPS, in either GTEPS (Giga TEPS) or MTEPS (Mega TEPS) are reported.
• Used a source code optimized for a scale-out (DMP) system by Kyushu University (Japan).
• Application is written in C language.
• Compiled using GNU compiler, gcc.
• Code automatically detects the no. of processors and cores and runs optimally.
• No external placement and binding by the user are needed.
• Needs large memory foot-print to run very large scale problem.
Additional details at: http://www.graph500.org/

31
Graph500 Performance from a Haswell Cluster
39.87
74.27
133.39
179.14
212.42
264.15
281.57
0.00
50.00
100.00
150.00
200.00
250.00
300.00
32 64 128 192 256 384 512
8 16 32 48 64 96 128
MEDIAN_GTEPS
# NODES / # RANKS
Graph500 – DMP Version
System Type Architecture / Chip CPU Clock (GHz)
#
processors/node total # cores/node
Memory
(GB) OS Interconnect Memory Details MPI File System
Hestia
BL460c
Gen9 Intel64 / Haswell E5-2697 v3 2.60 2 28 128 RHEL 6.6
IB / FDR (Connect-
IB)
8x16 GB, 2R, DDR4
@ 2133 MHz Intel MPI 5.0.2.044 NFS

Application: Weather Research and Forecasting (WRF)
Description:
WRF is a Numerical Weather Prediction (NWP) model designed to serve both atmospheric research and
operational forecasting needs. NWP refers to the simulation and prediction of the atmosphere with a computer
model, and WRF is a set of software to accomplish this.
• The code was jointly developed by NCAR, NCEP, FSL , AFWA, NRL, Univ. of Oklahoma and FAA.
• WRF is freely distributed and supported by NCAR.
• Offers two dynamical solvers: WRF-ARW (Advanced Research WRF) and WRF-NMM (Nonhydrostatic
Mesoscale Model).
• Capabilities to mix and match modules to simulate various atmospheric conditions and couple with other
NWP models (Ocean Modeling codes)
• Can accommodate simulation with nested data domains, coarse to very fine grids in a single simulation.
• Popular data sets to port and optimize are: CONUS12 and CONUS2.5 (available at NCAR).
• Options to use dedicated processors for I/O (quilting) and various layout of processors (tiling) exist.
Additional details at: http://www.wrf-model.org/index.php

WRF (v 3.8.1) Results with CONUS 2.5km Data Set
0.0000
0.5000
1.0000
1.5000
2.0000
2.5000
0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
40.00
45.00
4 8 16 24 32 48 64 96 128
Ave.Time/step(s)
Speedup
No. of Nodes
WRF Scaling with CONUS 2.5km
Ave. Time/step (s) Speedup
202 448
1017
1585
2034
3234
4065
5245
7964
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
4 8 16 24 32 48 64 96 128
GFLOPS
No. of Nodes
WRF GFLOPS with CONUS 2.5km
System: XL170r Gen9, Intel E5-2699 v4, 2.2 GHz, 2P/44C, 256 GB (DDR4-2133 MHz) Memory, RHEL 6.7, IB/EDR 1:1,
Intel Composer XE and Intel MPI (2016.3.210), Turbo ON, Hyperthreading OFF

Application: Clover Leaf
Description:
Clover Leaf is an open-source Computational Fluid Dynamics (CFD) code developed and distributed by UK
Mini-Application Consortium (UK-MAC).
• Solves compressible Euler equations on a Cartesian grid using explicit second-order accurate method.
• Uses a ‘kernel’ (low level building block) approach with minimal control logic to increase compiler
optimization.
• Supports for accelerators (using both OpenACC and OpenCL) available.
• Scarifies memory (saving intermediate results and than re-computing) to improve performance.
• Available in two flavors, in 2-Dimenional (2D) and 3-Dimensional (3D) modeling.
• Rugged and easy to port and run and good candidate for evaluating and comparing systems.
• Available large no. of data sets for 2D and 3D models with control of run times, few seconds to hours.
Additional details at: http://uk-mac.github.io/CloverLeaf/

Clover Leaf (3D) Results with bm256 Data Set
0.00
20.00
40.00
60.00
80.00
100.00
120.00
140.00
0
2000
4000
6000
8000
10000
12000
1 4 8 16 24 32 48 64 96 128
Speedup
WallClock(s)
No. of Nodes
Clover Leaf Scaling with bm256 Data
Wall Clock (s) Speedup
System: XL230a Gen9, Intel E5-2697A v4, 2.6 GHz, 2P/32C, 128 GB (DDR4-2400 MHz) Memory, RHEL 7.2,
IB/EDR 1:1, Intel Composer XE and Intel MPI (2016.3.210), Turbo ON, Hyperthreading OFF

Conclusions
37
 HP is No. 1 vendor in HPC and Cluster Solutions
 Configure and tune the system first
 Check the system details (processor, clock, memory and BIOS settings)
 Investigate compiler and flags that best suit your application
 Profile the application and optimize further for boosting performance
 Explore and decide on the right interconnect and protocols
 Take advantage of tools and commands to improve performance
 Run the application the right way (environment, placement etc.)
 Choose the right file system (local disk, NFS, Lustre, IBRIX etc.)
 Settle on an environment that is best for your application, time and value
 Never assume and always check the cluster before benchmarking

Thank you
logan.sankaran@hpe.com

Best Practices and Performance Studies for High-Performance Computing Clusters

Recommended

More Related Content

What's hot (20)

Similar to Best Practices and Performance Studies for High-Performance Computing Clusters (20)

More from Intel® Software (20)

Recently uploaded (20)

Best Practices and Performance Studies for High-Performance Computing Clusters