1. Scott Michael Indiana University July 6, 2017
Performance Benchmarking of the R Programming Environment on Knight's Landing
Scott Michael
Indiana University
July 6, 2017
Intro Slide

2. Who am I? Theoretical Astrophysicist NOT a statistician
HPC application optimization and performance tuning
Lead the Research Analytics team in Research Technologies at Indiana University

3. Contributors IU Eric Wernert Jefferson Davis James McCombs Esen Tuna
TACC
Bill Barth
Tommy Minyard
David Walling

4. Talk Overview
Targeting productivity languages for Xeon Phi based architecture: Motivation and History
Benchmark results and lessons learned
The RHPCBenchmark package
Future directions
Conclusions

5. IU, The Stampede Supercomputer, and Xeon Phi
IU Research Technologies has a partnership with TACC collaborating on systems and support
Stampede – largest XSEDE machine by core count
Wrangler – data intensive computing and 20 PB out of region replication
Jetstream – XSEDE production science cloud
IU supports data intensive and “high productivity” languages on Stampede
Including R, python, and Matlab
Large transition between Stampede 1 & 2

6. Evolution of Xeon Phi Knight’s Corner Knight’s Landing
Coprocessor only
Coprocessor or Self-hosted
1 TF peak (DP)
3 TF peak (DP)
8GB device + system memory
16GB MCDRAM + system memory

7. R Support on Stampede 1 & 2 Primary support on Stampede 1 for R
Support several methods for distributed R (pbdR, Rmpi, snow, etc.)
R built in offload mode
Configured R to use GPUs in portion of Stampede via HiPLAR
However, much of the R workload on Stampede didn’t rely on KNC
Stampede 1
Nodes
6,400
Interconnect
FDR IB
Filesystem
14 PB Lustre
Node Configuration
Processor
Dual E “SandyBridge”
Phi SE10P
Memory
32GB DDR3
8GB GDDR5
Stampede 2
Nodes
4,200
Interconnect
OmniPath v1
Node Configuration
Processor
Phi 7250
Memory
16GB GDDR4

8. R Performance on KNL KNL the sole processor on Stampede 2
Has shown good performance for large scale HPC codes (MD, climate, astro, etc.)
How does KNL perform with a language like R?

9. KNL Architecture
Intel(R) Xeon Phi(TM) CPU 1.60GHz (68 physical cores)
Features of note for KNL
Tiled architecture supporting 4 SMT threads per physical core

10. KNL Architecture (cont.)
Features of note for KNL
16GB on chip MCDRAM to act as fast memory can be configured into several modes

11. Benchmarking Strategy
Look at industry standard performance benchmarks for R on KNL and compare to SNB
Further explore some exemplar workflows in each language and compare to benchmark results
Compare both single node and multinode benchmarks

12. Benchmarking Strategy
R standard benchmark: R-25 benchmark
Very old, fixed (small) problem sizes, report output challenging to parse
Reasonable mix of mini-kernels focused on dense matrix operations and linear solvers
R benchmark for scalability focused on similar kernels to R-25
Built to distribute and for flexibility, currently available on CRAN at RHPCBenchmark

13. Talk Overview
Targeting productivity languages for Xeon Phi based architecture: Motivation and History
Benchmark results and lessons learned
The RHPCBenchmark package
Future directions
Conclusions

14. R Benchmark Results
Generally R lacks multithreading (some exceptions include mclapply) so we rely on the threading in MKL
Standard profiling/tracing tools are challenging to employ
Instrumenting entire R interpreter creates too much overhead

15. R Benchmark Results Benchmarks include
Cholesky decomp, eigendecomp, LS fit, linear solve, QR decomp, matrix cross, matrix det, matrix-matrix, matrix-vector
Multiple threads per core aren’t useful
Contrast to KNC

16. R Benchmark Results
For some benchmarks single core KNL outperforms SNB

17. R Benchmark Results
Need large matrices to make full use of all 68 cores

18. R Benchmark Results
For math intensive kernels R interpreter overhead isn’t bad

19. Talk Overview
Targeting productivity languages for Xeon Phi based architecture: Motivation and History
Benchmark results and lessons learned
The RHPCBenchmark package
Future directions
Conclusions

20. RHPCBenchmark Package
The RHPCBenchmark initial release is available on CRAN
Provides a variety of dense matrix, sparse matrix, and machine learning benchmarks
Users can configure the set of benchmarks to run and benchmark parameters
Results are provided in .csv files and a data frame for further analysis

21. Talk Overview
Targeting productivity languages for Xeon Phi based architecture: Motivation and History
Benchmark results and lessons learned
The RHPCBenchmark package
Future directions
Conclusions

22. Next Steps for R Performance
Internode performance
Higher level functions
Many R packages don’t rely on the building blocks tested (e.g. nnet, cluster)
Other classes of functions
Sparse matrix operations
Data wrangling operations

23. Talk Overview
Targeting productivity languages for Xeon Phi based architecture: Motivation and History
Benchmark results and lessons learned
The RHPCBenchmark package
Future directions
Conclusions

24. Conclusions
R performance on KNL better for dense matrix operations (3x SNB) and close to native C performance
Performance is best for large matrices
SNB does perform better for small matrices
New RHPCBenchmark offers flexibility in benchmarking your hardware and R build

25. Questions? Suggestions?
Scott Michael James McCombs

26. Backups: KNL Speedup in R

27. Backups: KNL vs. IvyBridge

28. Backups: KNL Flat vs. Cached