David Hoover

Scientific Computing Branch, Division of Computer System Services

CIT, NIH

Swarms and Bundles: Bioinformatics and Biostatistics

on Biowulf

Embarrassingly Parallel Problems

• GWAS, with huge numbers of SNPs

• Sequence analysis, assembly, and mapping

• Testing and validating statistical models

• Protein folding and threading

• Molecular docking and compound screening

• Tomographic reconstruction

Tsai et al., Mol. Biochem. Parasitology, online preprint 2008

Protein folding calculations with Rosetta++100,000 cpu hours

Characterization of Surface Protein 3 from Malaria Parasite P. Falciparum

How to run multiple independent processes in parallel

16 independent processes

input

command

output input output

command

Biowulf Cluster Batch System

batch

job1

job1.out

script

batch

job16

job16.out

script

Node 1 Node 2 Node 3 Node 4

job1 job2 job3 job4

job1.out job2.out job3.out job4.out

biowulf% swarm -f file

Swarm

Node 1

job1

job1.out

biowulf% swarm -f file -b 4

Bundled Swarm

Swarm Facts

• Written and maintained by Helix Systems Staff• swarm introduced in late 2000

• 82% of all batch jobs run on the cluster since 2002 are swarm jobs

• ~60% of all wall time spent on swarm jobs

• swarm has been shared with clusters around the world

Swarm World Records

• Largest swarm: 683,445 commands

• Largest bundle: 24,000 commands per CPU

Future Challenges

• How to deal with larger multicore nodes?

Node 1 Node 2 Node 3