Skip to content

Latest commit

 

History

History
104 lines (71 loc) · 4.13 KB

README.md

File metadata and controls

104 lines (71 loc) · 4.13 KB
Raw

ISx is Scalable Integer Sort Application

  • ISx is a new scalable integer sort application designed for co-design in the exascale era, scalable to large numbers of nodes.

  • ISx belongs to the class of bucket sort algorithms which perform an all-to-all communication pattern.

  • ISx is inspired by the NAS Parallel Benchmark Integer Sort and its OpenSHMEM implementation of University of Houston. ISx addresses identified shortcomings of the NPB IS.

  • ISx is a highly modular application implemented in the OpenSHMEM parallel programming model and supports both strong and weak scaling studies.

  • ISx uses an uniform random key distribution and guarantees load balance.

  • ISx includes a verification stage.

  • ISx is not a benchmark. It does not define fixed problems that can be used to rank systems. Furthermore ISx has not been optimzed for the features of any particular system.

  • ISx has been presented at the PGAS 2015 conference

References: ISx, a Scalable Integer Sort for Co-design in the Exascale Era. Ulf Hanebutte and Jacob Hemstad. Proc. Ninth Conf. on Partitioned Global Address Space Programming Models (PGAS). Washington, DC. Sep 2015. http://hpcl.seas.gwu.edu/pgas15/ http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=7306013

Information about the NAS Parallel Benchmarks may be found here: https://www.nas.nasa.gov/publications/npb.html

The OpenSHMEM NAS Parallel Benchmarks 1.0a by the HPCTools Group University of Houston can be downloaded at http://www.openshmem.org/site/Downloads/Examples

STRONG SCALING (isx.strong): Total number of keys are fixed and the number of keys per PE are reduced with increasing number of PEs Invariants: Total number of keys, max key value Variable: Number of keys per PE, Bucket width

WEAK SCALING (isx.weak): The number of keys per PE is fixed and the total number of keys grow with increasing number of PEs Invariants: Number of keys per PE, max key value Variable: Total Number of Keys, Bucket width

WEAK_ISOBUCKET (isx.weak_iso): Same as WEAK except the maximum key value grows with the number of PEs to keep bucket width constant This option is provided in effort to keep the amount of time spent in the local sort per PE constant. Without this option, the local sort time reduces with growing numbers of PEs due to a shrinking histogram improving cache performance. Invariants: Number of keys per PE, bucket width Variable: Total number of keys, max key value

Compiling and Executing the ISx Application

Compilation options:

make

  • Compiles with basic flags

make debug

  • Compiles with all debug flags, including -DDEBUG which enables verbose debugging print statements.

make optimized

  • Compiles with optimization flags, including -DNDEBUG, which disables assert statements.

The params.h file has various definitions that may be modified to change application options.

Usage: ./bin/isx.strong <total_num_keys> <log_file> ./bin/isx.weak <keys_per_pe> <log_file> ./bin/isx.weak_iso <keys_per_pe> <log_file>

The log file stores the verbose timing information for the run. Each row of the file corresponds to a single PE's timing results for each component of the application, for every iteration of computation. Reuse of the same output file will concatenate results.

example command lines (assuming aprun) for 2^27 keys (=134217728)

Strong: aprun -n 24 -N 4 ./bin/isx.strong 134217728 output_strong

Weak: aprun -n 24 -N 4 ./bin/isx.weak 134217728 output_weak

Weak_iso: Note that the iso-bucket width is specified in params.h aprun -n 24 -N 4 ./bin/isx.weak_iso 134217728 output_weak_iso

Note: timing measurements (see timer.c) are obtained by calls to clock_gettime with the clk_id argument set to CLOCK_MONOTONIC. However, not all systems support this clk_id. For such situations, clk_id should be changed to CLOCK_REALTIME, which is supported by all systems.

For backwards compatibility the params.h contains #define OPENSHMEM_COMPLIANT which can be undefined to fall back to earlier APIs