Number of Threads for BWA MEM

Question

[haiying.kong]

Is it recommended to use number of threads for BWA MEM, 2**N (4, 8, 16, 32...)? Or any number is as good as well as long as the memory can support?

[GenoMax]

You can't use more than what is available on your computer for a start. Experiment with a small set of reads (100K) to see what core # works best with your system in terms of time to complete. You will saturate something (PCI-E/memory bus) and will see a plateau in performance beyond a certain number of cores/threads.

[st.ph.n]

If your machine/cluster can support it, using many threads with -t should be fine.

[haiying.kong]

It does not have advantage if the number of thread is 2**N, is this correct?

[CMosychuk]

I'd stick to the greatest number of threads your processor can support, unless you're on a workstation and you actually need to do other stuff while bwa runs.

Answer 1

[Raony Guimarães]

Are you looking for something like this ?

Source: http://en.community.dell.com/techcenter/high-performance-computing/b/genomics

Answer 2

[fwuffy]

Bump. You can not run bwa with 1000 threads on one host and expect it to be 1000x faster than single thread. There will be a theoretical optimum which is a function of memory available per thread and reference genome size. Has anyone actually done tests?

[WouterDeCoster]

This will very likely depend on your computer architecture, for example I/O speed.

[Chris Miller]

+1 - for most setups, alignment will be I/O bound before any other bottleneck gets hit.

[kloetzl]

There will be a theoretical optimum

Did you mean Amdahl's law?