[mira_talk] Re: Parallel Implementation of MIRA
- From: Habeeb Syed <habeeb.syed@xxxxxxx>
- To: mira_talk@xxxxxxxxxxxxx
- Date: Thu, 3 Oct 2013 00:21:08 -0400
Thanks for the long mail Bastien. It was very helpful to know your
insight on parallelisation of MIRA.
Laurent, Thanks for sharing analysis.
Just wondering where can we find pseudocode for computations like Fast
Read comparision, Smith-Waterman Step, and Partial Path finder algorithm.
-H Syed.
From:
Bastien Chevreux <bach@xxxxxxxxxxxx>
To:
mira_talk@xxxxxxxxxxxxx,
Date:
10/01/2013 02:38 PM
Subject:
[mira_talk] Re: Parallel Implementation of MIRA
Sent by:
mira_talk-bounce@xxxxxxxxxxxxx
On Oct 1, 2013, at 11:05 , Habeeb Syed <habeeb.syed@xxxxxxx> wrote:
> I was wondering if anybody tried on large machines, something like few
hundred cores? How is the scale up?
Someone (maybe Jorge? I don't remember) made a couple of test runs on
larger machines. He found out that on his machine, performance
improvements plateaued out at around 40 cores iirc. I suppose that either
the bus or the disk were a bottleneck, but this was never investigated.
MIRA itself has only one part left which could "easily" be multithreaded
and is not yet: the Smith-Waterman matrix calculation and recursive
alignment generation. Incidentally, I suppose large speed-ups there could
be achieved by using SIMD instructions (SSE2/3/4 family) combined with
more intelligent memory usage and plain old parallelism on multiple cores.
Of that, I may attack the latter sometime this year.
Almost all other parts of MIRA which can be parallelised have been
parallelised. However, one of the key steps in MIRA - the contig building
- cannot be parallelised unless MIRA does like other assemblers and stops
at repeats. Which I don't plan to implement because, well, a repeat is
only a true repeat if it is 100% identical and not, like other assemblers
think, just almost identical.
> GPU acceleration will be nice; have you tried?
> What about GPU implementation of MIRA. Can anybody share some insight
on this?
As far as I am aware of, GPU implementations of bioinformatics algorithms
have not had a terrible success in the wild. Either because they didn't
deliver that much of a speed-up or, equally important, because it's too
much work to implement several different GPU algorithms for all the
different cards out there. Even more vexing, quite a number of larger
servers have no GPU at all. SSE instructions on the other hand are quite
common nowadays, one would be hard pressed to find machines which have no
SSE4 (let alone 3 or 2).
As last point, the paper you cited in your initial posting (if it's the
one I have in mind, I did not check yet) had come as a complete surprise
to me when I initially discovered it, totally by chance, a couple of
months ago. Someone had been working on MIRA code without contacting me to
find out whether I could make use of the changes or integrate it back into
MIRA. That had been a quite strange experience at the time.
B.
--
You have received this mail because you are subscribed to the mira_talk
mailing list. For information on how to subscribe or unsubscribe, please
visit http://www.chevreux.org/mira_mailinglists.html
=====-----=====-----=====
Notice: The information contained in this e-mail
message and/or attachments to it may contain
confidential or privileged information. If you are
not the intended recipient, any dissemination, use,
review, distribution, printing or copying of the
information contained in this e-mail message
and/or attachments to it are strictly prohibited. If
you have received this communication in error,
please notify us by reply e-mail or telephone and
immediately and permanently delete the message
and any attachments. Thank you
Other related posts:
