Optimize Single Particle Orbital (SPO) Evaluations Based on B-splines

Amrita Mathuriya
HPC Application Engineer
DCG/Intel Corporation
November 2016
Optimizations of B-spline based SPO evaluations in QMC for
Multi/many-core Shared Memory Processors
In collaboration with
Jeongnim Kim (Intel), Victor Lee(Intel)
Ye Luo, Anouar Benali (Argonne
National Laboratory)
Luke Shulenburger (Sandia National
Laboratories )

11/12/2016
Presenter: Amrita Mathuriya
3
§  HPC application Engineer at HPC Ecosystem Application Engineering Team; working for code
modernization and optimization on Xeon and Xeon Phi™
§  Working at Intel for past 8 Years.
–  Expert at algorithms and optimizations for IA architectures.
–  Worked on HPC applications in areas of Computational Geometry, Optical Proximity Correction
(OPC), Electromagnetics, Computational Biology, Quantum Monte Carlo.
–  Working on code modernization for Intel® Xeon and Xeon Phi™ architectures.
§  MS in Computer Science with the specialization in Computational Science and Engineering from
Georgia Tech, USA under the guidance of Professor David Bader.
§  Obtained B. Tech degree in Computer Science from Indian Institute of Technology (IIT) Roorkee, India.

11/12/2016
Systems
§  KNC: Intel® Xeon Phi™ coprocessor 7120P
•  61 cores @ 1.238 GHz, 4-way Intel® Hyper-Threading Technology, Memory: 15872 MB
•  Intel® Many-core Platform Software Stack Version 3.6.1
•  OS Version : 3.10.0-229.el7.x86_64
§  Intel® Xeon Phi™ 7250P (code-named Knights Landing, KNL), 68 cores, 1.4GHz with
16GB MCDRAM (used in flat mode), cluster boot mode=Quad, Turbo=enable. KNL used in
Quad/Flat mode.
§  Intel® Xeon® E5-2697v4(BDW) node single socket, 18 cores HT Enabled @2.3GHz 145W
(E5-2697v4 w/128GB RAM DDR4 2400 8*16GB DIMMS.
§  Bluegene/Q (BG/Q) processor from Mira Supercomputer, at Argonne National lab facility.
§  Compilers and MPI and math library.
•  icc version 16.0.2 (gcc version 4.8.3 compatibility)
•  Intel(R) MPI Library for Linux* OS, Version 5.1.3 Build 20160120 (build id: 14053)
4

11/12/2016
Agenda
§  KNL overview and motivation
§  Intro to quantum Monte Carlo and QMCPACK
§  Current status of QMCPACK
§  Analysis of CORAL graphite benchmark
§  Optimizations to B-spline based SPO evaluations for QMC
§  Summary
5

11/12/2016
Important Characteristics of KNL
7
§  Increasing core count per node on both Intel® Xeon® and Xeon Phi™
processors.
§  Large SIMD units – AVX512 supporting 16 single precision floating point
simultaneously.
§  Two level Cache system L1/L2 and high memory bandwidth.

11/12/2016
How to gain performance?
8
§  Scalability
–  Enable data sharing with hybrid parallelism using MPI + threading.
–  Design and implement scalable algorithms
§  SIMD Parallelism – adapt Data layouts to enable efficient vectorization.
§  Efficiently utilize caches and memory bandwidth with Tiling (cache-blocking).

11/12/2016 9
Performance tests and ratings are measured using specific computer systems and/or components and reflect the approximate performance of Intel products as measured by those tests. Any
difference in system hardware or software design or configuration may affect actual performance. Buyers should consult other sources of information to evaluate the performance of systems
or components they are considering purchasing. For more information on performance tests and on the performance of Intel products, visit Intel Performance Benchmark Limitations.
7% of peak GFLOPS achieved with the current AoS version
Roofline Performance Analysis on KNL
Peak GFLOPS at
(0.22 Flops/Byte
VGH roofline performance model for N=2048. Circles denote GFLOPS at the cache-aware AI.
0.22
GFLOPS NOW, 7%
Scalar add peak

11/12/2016
Performance Portable on Intel® Xeon®, Intel Xeon Phi™
and BG/Q Processors
10
§  Optimizations for efficiently utilizing SIMD units and caches.
–  SoA data layout transformation.
–  Tiling or AoSoA data layout transformation.
–  Nested thread parallelization to reduce time-to-solution and memory usage.
§  Optimization work done on KNC.
§  Later ported on KNL – works out of the box.
§  Optimizations result in significant performance improvement on BG/Q.

11/12/2016
Parallel efficiency of QMCPACK on US-DOE facilities. The legend
shows the MPI tasks and OpenMP threads of the reference computing
unit (CU) and the maximum number of nodes on each platform.
QMCPACK
An open-source US-DOE flagship many-body ab initio quantum Monte Carlo (QMC) code for
computing the electronic structure of atoms, molecules, and solids. http://qmcpack.org/
(a) DMC charge-density of AB stacked graphite and (b) the ball-and-
stick rendering and the 4-Carbon unit cell in blue.
11
J. Kim, K. P. Esler, J. McMinis, M. A. Morales, B. K. Clark, L. Shulenburger, and D. M. Ceperley,
“Hybrid algorithms in quantum monte carlo,” Journal of Physics: Conference Series, vol. 402, no.
1, p. 012008, 2012. [Online]. Available: http://stacks.iop.org/1742-6596/402/i=1/a=012008

11/12/2016
Diffusion Monte Carlo Schematics

Ensemble
evolves
according to
•  Diffusion
•  Drift
•  Branching
Possible new
configurations
Old configurations Random
Walking
New configurations
ensemble
w=0.8
w=1.6
w=2.4
w=0.3

11/12/2016
How is QMCPACK parallelized
QMCPACK utilizes OpenMP to optimize memory usage and to take
advantage of the growing number of cores per SMP node.
13
§  Walkers within a MPI task is distributed among
the cores in CPU.
§  Big common data is shared by all the walkers
like wave function coefficients.
§  Frequency stops increasing.
§  Node count stops growing.
§  Nodes are getting more powerful but require
applications to expose more concurrency.
Free lunch is over. On node performance is challenging.

11/12/2016
QMCPACK status
§  Excellent MPI & OpenMP parallel efficiency at the walker level
§  All in double precision except 3D cubic B-Spline.
–  Work done recently to implement mixed precision. Speeds up by 1.2-1.5x.
§  SIMD efficiency low
§  Basically scalar performance with few exceptions
–  B-Spline – SSE/SSE2/QPX
–  Distance tables with QPX
§  Array of Structure (AoS) for D-dim N-particle attributes, e.g., R (N,3),
Gradients (N,3), Hessian matrices (N,9)
Pretty good and we can even do better!
14

11/12/2016
CORAL Benchmark – KNL Profiling
29%
34%
18%
18%
Coral Benchmark Profile on KNL
Einspline
Distance Table
Jastraw
Others
4x4x1 AB-stacked graphite
64 carbon
256 electrons
15
The three compute kernels account for 80% run time in QMCPACK on KNL.

11/12/2016
QMC: Single particle orbital (SPO) representation with B-
spline basis set
16
One Dimensional cubic B-spline function
Precomputed coefficients
4D Read only array.
Stored in SOA format, P[nx][ny][nz][N]
Provided by DFT or HF computations using
Quantum Espresso
Tensor product in each Cartesian direction,
Representation for 3D orbital,

11/12/2016
Simplified miniQMC
17
§  Only contains B-spline
evaluation routines.
§  Mimics the computational and
data access patterns of B-
spline SPO evaluations in
QMC.
B-spline SPO evaluation kernels
Random position generation

11/12/2016
Array-of-Structs (AoS)
§  Pros:
Logical for expression of
physical abstractions in
3D or higher dimensions.
Struct-of-Arrays (SoA)
§  Pros:
Contiguous loads/stores
for efficient vectorization.
Hybrid (AoSoA)
§  Pros:
Potentially useful for
increasing cache locality.
Also supports efficient
vectorization.
x x xx x x
y y yy y y
z z zz z z
x x
x x
x x
y y
y y
y y
z z
z z
z z
Data Layout – Performance Considerations
18
x x
x x
x x
…
…
…
yy
yy
yy
…
…
…
z z
z z
z z
…
…
…

11/12/2016
Pseudocode - VGH
Computes value, gradient, Hessian at random (x,y,z)
19
Random
Data access pattern of read-only B-spline
coefficients P at a random position (x; y; z)
and j0=floor(y/dy) etc. The outermost x
dimension is not shown.
Strided access
for output arrays.

11/12/2016
SoA transformation for output arrays
20
Output arrays in SoA
(Structure of arrays)
format
x x xx x x
y y yy y y
z z zz z z
…
…
…

11/12/2016
How to evaluate performance of QMC
§  Rate of Monte Carlo sample generations (throughputs) per resource
§  For the miniapp,
Throughput = (number of evaluations)/(T)
Evaluations = (Number of walkers) X (Number of iterations) X (Number of splines)
T = Time per call of a function ( such as VGH )
§  Throughput represents work done on a node.
§  Ideally, it should stay constant across problem sizes.
21

11/12/2016
VGH throughput by AoS-to-SoA transformation
Higher the better
2x-4x Performance improvement for small to medium problem sizes.
22

11/12/2016
Pseudocode - VGH
Computes value, gradient, Hessian at random (x,y,z)
23
Random
Data access pattern of read-only B-spline
coefficients P at a random position (x; y; z)
and j0=floor(y/dy) etc. The outermost x
dimension is not shown.
Strided access
for output arrays.

11/12/2016
Why low performance for large N?
•  AoS-to-SoA improves SIMD efficiency
•  But, caches can be utilized better
•  Reduction on the arrays G& H of
size N
•  Streaming access at 4x4x4 block
•  Pressure on resources with large N,
e.g., TLB
•  How to keep the write data in L1/L2
•  How to maximize LLC sharing
Core Core
HUB
24
Reduction of
output arrays
over 64N values

11/12/2016
AoSoA Data Layout
Transformation
25
Tiled Input array
Tiled output arrays
Data access pattern of read-only
B-spline table
a) Current b) Tiled
x x
x x
x x
…
…
…
yy
yy
yy
…
…
…
z z
z z
z z
…
…
…
Efficient cache utilization, by tiling both input and output arrays along the innermost dimension.

11/12/2016
Performance gain with tiling/AoSoA - Higher the better
AoSoA helps achieve sustained throughput across problem sizes for all architectures.
VGH Performance with SoA to AoSoA transformation (tiling)
26

11/12/2016
VGH throughput with tiling, higher the better
Tiling improves performance for all three processors.
Performance of VGH at N = 2048 with respect to tile size.
27
§  BDW – peak at 64
§  The tiled input array fits in L3
cache.
§  KNC, KNL – peak at 512
§  For tile size > 512, output arrays
fall out of caches.

11/12/2016
Hybrid OpenMP/MPI Parallelism in QMCPACK
28
§  Current parallelism over walkers (Nw).
§  Working set size in QMCPACK grows
with number of walkers.
§  Parallelizing each walker update
§  Specifically, for Intel Xeon Phi, with large
number of cores/threads, next level of
parallelism becomes essential for strong
scaling. Parallel efficiency of QMCPACK on US-DOE facilities. The legend
shows the MPI tasks and OpenMP threads of the reference computing
unit (CU) and the maximum number of nodes on each platform.
J. Kim, K. P. Esler, J. McMinis, M. A. Morales, B. K. Clark, L. Shulenburger, and D. M. Ceperley,
“Hybrid algorithms in quantum monte carlo,” Journal of Physics: Conference Series, vol. 402, no.
1, p. 012008, 2012. [Online]. Available: http://stacks.iop.org/1742-6596/402/i=1/a=012008

11/12/2016
Parallelism within a walker – nested threading
29
#pragma omp parallel
Strong Scaling:- Independent execution of tiles in different threads.
•  Reduces memory requirement
and time to solution on a node,
by reducing the number of
walkers on a node.
•  miniQMC replaces OpenMP
nested threading with manual
assignment of work.

11/12/2016
Strong Scaling Results on
KNL
30
Reduces time to solution by ~14x with 16 threads per walker
Speedup on KNL w.r.t. number of walkers per thread.

11/12/2016
Strong Scaling Results on
KNL
31
Reduces time to solution by ~14x with 16 threads per walker
Speedup on KNL w.r.t. number of walkers per thread.
Performance of VGH at N = 2048 with
respect to tile size.

11/12/2016
32
§  SoA data layout conversion
–  Increases cache aware AI
from 0.22 to 0.32
–  ~7% of the achievable peak.
–  1.5x speedup wrt. AoS
version.
VGH roofline performance model for N=2048. Circles denote GFLOPS
at the cache-aware AI and X (b) the best performance (AoSoA) on DDR.
0.22 à 0.32
SoA,
~7% of peak
GFlops

11/12/2016
33
§  AoSoA version increases
cache reuse with the same
AI.
–  Better cache utilization.
–  ~2.25x gain in performance.
0.22 à 0.32
AoSoA,
11% of peak
GFlops

11/12/2016
34
§  AoSoA version with
MCDRAM ~3.3x faster than
DDR.
AoSoA,
3.3x speedup
With
MCDRAM

11/12/2016
Roofline performance analysis on BDW
35
Performance improved to ~50% of peak GFLOPS with the AoSoA version.
660 GFLOPS SP Vector FMA Peak
VGH roofline performance model for N=2048. Circles denote GFLOPS at the cache-aware AI
AoSoA,
~50% of
achievable
GFlops

11/12/2016
Performance Summary
36
§  The improvements are portable to 4 types of CPUs, even from different vendors.
§  Significant speedups even on BG/Q.
On VGH routine
BGQ
BDW
KNC
KNL
SOA and basic
1.9x
1.7x
2.6x
1.7x
AoSoA/Tiling
2.7x
3.7x
5.2x
2.3x
Strong scaling
5.2x
6.4x
35.2x
33.1x
Number of threads per
walker
(The optimal tile size)
2(32)
2(32)
8(256)
16(128)

11/12/2016
Symmetric Distance table computation
AoS to SoA transformation of particle positions.
37
0.8 0.6 0.6
7.5
13.1 13.0
18
30 30
0.0
5.0
10.0
15.0
20.0
25.0
30.0
35.0
256 512 800
Speedupwrt.BDWBaseline
Number of electrons
Speedup vs. Problem Size
Higher the better
KNL Baseline(256 TH) BDW Opt(2MPI/36TH) KNL Opt(256 TH)
KNL 50x
Faster with
SoA data
layout
•  KNL used in Quad/Cache
mode for these
experiments
•  Here, TH = threads.
•  BDW has 2 sockets for
these experiments.

11/12/2016
Results
§  Array of structures (AOS) to structure of arrays (SOA) transform helps
achieve efficient vectorization.
§  Tiling for better memory access helps achieve approximately constant
throughput across problem sizes.
§  Nested parallelism over the AoSoA objects on KNL helps reduce the time-to-
solution by ~14x speedup with 16 threads.
§  Optimizations result in significant performance gain on all three distinct
cache-coherent architectures.
38

11/12/2016
Ways we increased the performance!
39
§  SIMD Parallelism
–  SoA data layout adaption.
§  Efficient cache utilization
–  Tiling/Cache-Blocking.
§  Scalability
–  Next level of threading to reduce time to solution.
–  Takes advantage of reduced working set size.

11/12/2016
Reference
40
Amrita Mathuriya, Ye Luo, Anouar Benali, Luke Shulenburger, Jeongnim Kim
“Optimization and parallelization of B-spline based orbital evaluations in QMC on multi/
many-core shared memory processors”
arXiv:1611.02665

11/12/2016
Legal Disclaimers
INFORMATION IN THIS DOCUMENT IS PROVIDED IN CONNECTION WITH INTEL PRODUCTS. NO LICENSE, EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO ANY
INTELLECTUAL PROPERTY RIGHTS IS GRANTED BY THIS DOCUMENT. EXCEPT AS PROVIDED IN INTEL'S TERMS AND CONDITIONS OF SALE FOR SUCH PRODUCTS, INTEL
ASSUMES NO LIABILITY WHATSOEVER AND INTEL DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY, RELATING TO SALE AND/OR USE OF INTEL PRODUCTS INCLUDING
LIABILITY OR WARRANTIES RELATING TO FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY PATENT, COPYRIGHT OR OTHER
INTELLECTUAL PROPERTY RIGHT.
A "Mission Critical Application" is any application in which failure of the Intel Product could result, directly or indirectly, in personal injury or death. SHOULD YOU PURCHASE OR
USE INTEL'S PRODUCTS FOR ANY SUCH MISSION CRITICAL APPLICATION, YOU SHALL INDEMNIFY AND HOLD INTEL AND ITS SUBSIDIARIES, SUBCONTRACTORS AND
AFFILIATES, AND THE DIRECTORS, OFFICERS, AND EMPLOYEES OF EACH, HARMLESS AGAINST ALL CLAIMS COSTS, DAMAGES, AND EXPENSES AND REASONABLE ATTORNEYS'
FEES ARISING OUT OF, DIRECTLY OR INDIRECTLY, ANY CLAIM OF PRODUCT LIABILITY, PERSONAL INJURY, OR DEATH ARISING IN ANY WAY OUT OF SUCH MISSION CRITICAL
APPLICATION, WHETHER OR NOT INTEL OR ITS SUBCONTRACTOR WAS NEGLIGENT IN THE DESIGN, MANUFACTURE, OR WARNING OF THE INTEL PRODUCT OR ANY OF ITS
PARTS.
Intel may make changes to specifications and product descriptions at any time, without notice. Designers must not rely on the absence or characteristics of any features or
instructions marked "reserved" or "undefined". Intel reserves these for future definition and shall have no responsibility whatsoever for conflicts or incompatibilities arising from
future changes to them. The information here is subject to change without notice. Do not finalize a design with this information.
The products described in this document may contain design defects or errors known as errata which may cause the product to deviate from published specifications. Current
characterized errata are available on request.
Contact your local Intel sales office or your distributor to obtain the latest specifications and before placing your product order.
Copies of documents which have an order number and are referenced in this document, or other Intel literature, may be obtained by calling 1-800-548-4725, or go to:
http://www.intel.com/design/literature.htm
Knights Landing and other code names featured are used internally within Intel to identify products that are in development and not yet publicly announced for release.
Customers, licensees and other third parties are not authorized by Intel to use code names in advertising, promotion or marketing of any product or services and any such use of
Intel's internal code names is at the sole risk of the user
Intel, Look Inside, Xeon, Intel Xeon Phi, Pentium, Cilk, VTune and the Intel logo are trademarks of Intel Corporation in the U.S. and other countries.
*Other names and brands may be claimed as the property of others.
Copyright © 2016 Intel Corporation
41

11/12/2016
Intel's compilers may or may not optimize to the same degree for non-Intel
microprocessors for optimizations that are not unique to Intel microprocessors. These
optimizations include SSE2, SSE3, and SSE3 instruction sets and other optimizations. Intel
does not guarantee the availability, functionality, or effectiveness of any optimization on
microprocessors not manufactured by Intel.
Microprocessor-dependent optimizations in this product are intended for use with Intel
microprocessors. Certain optimizations not specific to Intel microarchitecture are reserved
for Intel microprocessors. Please refer to the applicable product User and Reference Guides
for more information regarding the specific instruction sets covered by this notice.
Notice revision #20110804
Legal Disclaimers
Optimization Notice
42

11/12/2016
§  Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark*
and MobileMark*, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the
results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of
that product when combined with other products. For more information go to http://www.intel.com/performance.
§  Estimated Results Benchmark Disclaimer:
Results have been estimated based on internal Intel analysis and are provided for informational purposes only. Any difference in system hardware or software design or
configuration may affect actual performance.
§  Software Source Code Disclaimer:
Any software source code reprinted in this document is furnished under a software license and may only be used or copied in accordance with the terms of that license.
§  Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the
Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to
permit persons to whom the Software is furnished to do so, subject to the following conditions:
§  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE
WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
Legal Disclaimers
43

Thank you for your time
Amrita Mathuriya
amrita.mathuriya@intel.com
www.intel.com/hpcdevcon

11/12/2016
SymmetricDTD::moveonsphere – Code Sinippet
§  For efficient auto-vectorization with the compiler
§  Three separate arrays for X, Y and Z instead of a single
array with (x, y, z) as a data member.
§  Similar SOA (structure of arrays) data layout for the output
array.
AoS
SoA
46

Optimize Single Particle Orbital (SPO) Evaluations Based on B-splines

Recommandé

Recommandé

Contenu connexe

Tendances

Tendances (20)

Similaire à Optimize Single Particle Orbital (SPO) Evaluations Based on B-splines

Similaire à Optimize Single Particle Orbital (SPO) Evaluations Based on B-splines (20)

Plus de Intel® Software

Plus de Intel® Software (20)

Dernier

Dernier (20)

Optimize Single Particle Orbital (SPO) Evaluations Based on B-splines