Re: RAxML GPU Cluster Implementation

[Aram Valafar]

Hi,

I am a PhD student researching Mycobacterium tuberculosis. RAxML is extremely useful for my research, but I suffer from performance issues. My lab has access to a university wide GPU cluster for high performance computing. I would like to port RAxML onto this cluster, but unfortunately it does not have a native GPU implementation. I came across your paper from 2005 regarding your GPU application (Initial Experiences Porting a Bioinformatics Application to a Graphics Processor), as well as your other parallelization papers, and was very intrigued. I am reaching out to see if you would be willing to share your source code, or provide any tips for tackling this project. Any advice is greatly appreciated.

Sincerely,

Aram Valafar

[Alexandros Stamatakis]

Dear Aram,

> I am a PhD student researching Mycobacterium tuberculosis. RAxML is
> extremely useful for my research,

:-)

> but I suffer from performance issues.

How large is your dataset, have you switched to RAxML-NG ?
https://github.com/amkozlov/raxml-ng

> My lab has access to a university wide GPU cluster for high performance
> computing. I would like to port RAxML onto this cluster, but
> unfortunately it does not have a native GPU implementation. I came
> across your paper from 2005 regarding your GPU application (Initial
> Experiences Porting a Bioinformatics Application to a Graphics
> Processor), as well as your other parallelization papers, and was very
> intrigued. I am reaching out to see if you would be willing to share
> your source code, or provide any tips for tackling this project. Any
> advice is greatly appreciated.

These old source codes, both from that paper and the one in 2013

https://ieeexplore.ieee.org/document/6650928

where proof of concept implementations that never made it into the
production level software versions of RAxML.

The reason is that the performance benefits were small as
likelihood-based codes are genereally memory bandwidth bound and under
models typically used in practice the partitions are not sufficiently
large to allow for efficient parallelization.

We have a student working on re-assessing GPU deployment for
accelerating phylogenetic likelihood calculations, but the performance
limitations are generally the same as in 2013 it seems.

In any case, there is no production level GPU implementation available
at present as our x86 implementation is highly tuned and efficient.

The key issue is rather the general cluster purchase strategy with too
many GPUs at almost all institution and computing centers I know.

Hope this helps,

Alexis

>
> Sincerely,
> Aram Valafar
>
> --
> You received this message because you are subscribed to the Google
> Groups "raxml" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to raxml+un...@googlegroups.com
> <mailto:raxml+un...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/raxml/502b5313-69cb-460a-8957-f9f64b1fa0abn%40googlegroups.com <https://groups.google.com/d/msgid/raxml/502b5313-69cb-460a-8957-f9f64b1fa0abn%40googlegroups.com?utm_medium=email&utm_source=footer>.

--
Alexandros (Alexis) Stamatakis

ERA Chair, Institute of Computer Science, Foundation for Research and
Technology - Hellas
Research Group Leader, Heidelberg Institute for Theoretical Studies
Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology

www.biocomp.gr (Crete lab)
www.exelixis-lab.org (Heidelberg lab)