Scribd
Upload
38 views

Simplified Supercomputer Architecture Slide

On-node parallelism BWA, bowtie, Velvet, etc. Use pthreads or OpenMP to implement 'threads' MPI Parallel Tasks ABySS, MAKER, RaXML, etc. Make use of MPI (Message Passing Interface) any code where you can arbitrarily split the work across workers can be adapted to this model.

Uploaded by

Matthew Vaughn
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
38 views

Simplified Supercomputer Architecture Slide

On-node parallelism BWA, bowtie, Velvet, etc. Use pthreads or OpenMP to implement 'threads' MPI Parallel Tasks ABySS, MAKER, RaXML, etc. Make use of MPI (Message Passing Interface) any code where you can arbitrarily split the work across workers can be adapted to this model.

Uploaded by

Matthew Vaughn
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

Simplified Supercomputer Architecture Slide

Login
File system Network

Node 01

Node 02

Node 03

Node 04

4 nodes; 48 cores; Global file system; Connected by network; Controlled by Login node

Serial Tasks
SAMtools, Perl or Python scripts, R scripts, Other non-parallel code
Login
File system Network

Node 01

Node 02

Node 03

Node 04

Other 11 CPUs sit idle Task takes longer You are charged for all 12 CPUs

On-node Parallelism
BWA, Bowtie, Velvet, etc. use pthreads or OpenMP to implement threads
Login
File system Network

Node 01

Node 02

Node 03

Node 04

Makes use of all CPUs Your task can run up to 12x faster*

Caveats 1. Do you have enough RAM to run all 12 threads? 2. Is the code actually more efficient with additional threads? 3. What happens if you need more than 12 CPUs?

MPI Parallel Tasks


ABySS, MAKER, RaXML, etc. use MPI (Message Passing Interface)
Login
File system Network

Node 01

Node 02

Node 03

Node 04

You can make use of all available CPUs Tasks can run up to 48x faster!

Caveats 1. Really good networking is needed because tasks communicate with one another 2. Algorithm must be written to explicitly make use of MPI. Most bioinformatics code is NOT L

Simple Parallel Tasks


Any code where you can arbitrarily split the work across workers can be adapted to this model
Login
File system Network

Node 01

Node 02

Node 03

Node 04

You can make use of all available CPUs Tasks can run up to 48x faster!

Caveats 1. Shared file system is needed for this approach 2. You have to figure out how to shoe-horn your algorithm into this model

Why Does Parallelism Matter?


Login
File system Network

Node 01

Node 02

Node 03

Node 04

Upcoming Intel Xeon Phi systems (2013 / Not to scale) Notice anything different?

You might also like