Argonne National Laboratory Mathematics and Computer Science Division
Argonne Home > MCS Division > Seminar & Events

Seminars & Events

Bookmark and Share

Argonne Leadership Computing Facility
"Experiences on Performance Enhancement of Parallel I/O and FFT on Blue Gene/L Supercomputer"

DATE: March 26, 2009
TIME: 10:30 AM - 12:00 AM
SPEAKER: Dr. Ramendra Sahoo, IBM Research Division, Blue Gene Systems Software Group
LOCATION: Building 221, Room A216, Argonne National Laboratory

Description:
The talk will cover our experiences on performance analysis, tuning and implementation of (1) Parallel I/O* and (2) FFT algorithm optimizations** carried out for Blue Gene/L Supercomputer.

In order to provide sustainable Parallel I/O performance, we designed and implemented highly scalable parallel file I/O architecture for the Blue Gene/L system. Our architecture leveraged the benefit of the hierarchical and functional partitioning design of the system software with separate computational and I/O cores. Exploiting the scalability aspect of GPFS (General Parallel File System) at the backend and using MPI I/O as an application interface, the architecture was able to deliver at least one order of magnitude higher I/O bandwidth for a real application; i.e., for HOMME application we achieved an aggregate bandwidth of 1.8 GB/Sec and 2.3 GB/Sec for write and read accesses, respectively). The implementation also included the support of high-level parallel I/O data interfaces such as parallel HDF5 and parallel NetCDF scaling up to thousands of processors.

To enhance the 2D/3D FFT algorithm (as a part of HPC Challenge Benchmark Suite), we have exploited (1) single-node FFT performance, (2) all-to-all collective performance, and (3) overlap of computation and communication. Through effective exploitation of Blue Gene/L's double-FPU intrinsics, careful placement of all-to-all operations and synchronizations to maximize the interleave of communications and computations, substantial performance enhancement was achieved; i.e., a highly scalable FFT implementation with 20% performance improvement over the FFTW baseline on the LLNL Blue Gene/L system.

__________________
*Joint work with ANL (R. B. Ross, R. Thakur, R. Latham, W. D. Gropp) and H. Yu, C. Hawson, J. Moreira, T Engelsiepen from IBM Research.
**Joint work with J. Gunnels, Y. Shabharwal, R. Garg from IBM Research.


Save the event to your calendar [schedule.ics]


The Office of Advanced Scientific Computing Research | UChicago Argonne LLC | Privacy & Security Notice | ContactUs