Synergy-team

Download

synergy-team@cs.vt.edu

August 2014

1 participants
1 discussions

Re: [Synergy-team] Hackathon team submission from team Synergy@VT
by Foertter, Fernanda S. 01 Sep '14

01 Sep '14

Dear Synergy Team, Thank you for your submission. We will announce teams on Sep 5th. Please email me directly if you have any questions! Happy Computing! Fernanda -------------------------------------------------------- Fernanda Foertter National Center for Computational Sciences Oak Ridge National Laboratory O: 865-576-9391 F: 865-241-2850 foertterfs(a)ornl.gov<mailto:foertterfs@ornl.gov> On Aug 29, 2014, at 4:24 PM, "OpenACC Hackathon: Synergy@VT" <synergy-team(a)cs.vt.edu<mailto:synergy-team@cs.vt.edu>> wrote: Team Name Synergy@VT Email synergy-team(a)cs.vt.edu<mailto:synergy-team@cs.vt.edu> Tell us about your application Application Name: mpiBLAST (http://www.mpiblast.org/), 2004 R&D 100 Award. (Note: We are not confined to the above application. We have two other applications of interest as part of larger project in large-scale computing and big data, namely ABySS and Ray, which we briefly discuss at the end of this application.) Application Purpose: The BLAST family of sequence search algorithms serves as the foundation for much biological research. It searches for similarities between a query sequence and a large database of DNA or amino-acid sequences. Similarities between the new sequence and a gene of known function can help identify the function of the new sequence. Other uses include phylogenetic profiling, genome alignment, and understanding gene function and evolution. LOC: 1.2M Languages and Libraries: MPI and NCBI BLAST Current performance characteristics: Using a novel combination of database segmentation and query segmentation (the latter which is the "embarrassingly parallel" way of parallelizing BLAST), mpiBLAST delivers both better throughput and faster response time per query. It has been shown to scale up to 65,536 processors of an IBM BlueGene/L supercomputer at KAUST. Where does it run? CPUs only. Expected performance gains? The expected performance gain will depend on the granularity of parallelization that we will be able to tackle during the "Hackathon." If we parallelize at the coarse-grained level (as in our GPU-BLASTP algorithm from IPDPS'11), then the performance gain will be modest, e.g., 2x; if at the more sophisticated and fine-grained level (as in our cuBLASTP algorithm from IPDPS'14), then the performance gain will be a bit higher, e.g., 3x to 4x). The performance gain will also be dependent on the NCBI BLAST toolkit. Tell us about your team The team consists of members from the Synergy Laboratory, which is led by Professor Wu Feng, the PI for mpiBLAST. Of the team members, Jing Zhang, a PhD student of Professor Feng, is the most familiar with mpiBLAST due in part to his experience with related projects, e.g., cuBLASTP (IPDPS'14). He also is an experienced CUDA programmer but does not have any experience with OpenACC. The remaining members have a range of experience in parallel computing but are not familiar with mpiBLAST. 1. Jing Zhang, 3rd-year PhD student, CUDA and life sciences expertise. Recent publication on cuBLAST in IPDPS'14. 2. Kaixi Hou, 3rd-year PhD student, CUDA expertise and OpenCL and MPI experience. Recent publication on graph traversal on Xeon Phi. 3. Rubasri Kalidas, 2nd-year MS student, OpenCL experience. No publications yet. Female member of the team. 4. Yilong Jin, 2nd-year PhD student, OpenCL experience. Working on an OpenCL version of MapReduce for CPU+GPU. 5. Islam Harb, 1st-year PhD student but with 5+ years of industry experience in parallel hardware and architecture, CUDA experience. U.S. citizen. 6. Xiaodong Yu, 2nd-year PhD student, CUDA expertise but no experience with OpenACC or MPI. CUDA GPU publications as part of his MS. If selected, I understand at least 2 team members will travel to the hackathon on Oct 27th-31st * I understand What computing facilities does this application run on? Local clusters, desktops, and HPC centers. At one point, mpiBLAST was being distributed as part of Linux distributions. Please check skills your team members have experience with, if any * Developing in C/Fortran * Developing in C++ * Compiling applications * Profiling applications * Debugging applications using debuggers * GPU Architecture * GPU Programming (OpenACC, CUDA, OpenCL) * Linux environment (logging in remotely, using terminal, basic commands) * Experience with multiple compilers (GCC, PGI, Cray etc) Is there anything else you'd like to mention or ask us? If HPC = supercomputing, e.g., Titan, then the team's experience in HPC is limited at best, particularly with respect to MPI, where Kaixi Hou is the only one with adequate MPI experience. If HPC = massive parallelism within a node via CUDA, OpenCL, or OpenMP, then the team's experience is fairly extensive. In addition to the primary email contact, synergy-team(a)cs.vt.edu<mailto:synergy-team@cs.vt.edu>. Please feel free to contact Professor Wu Feng at feng(a)cs.vt.edu<mailto:feng@cs.vt.edu> or 540-231-1192 if you have any questions.

2 1