[mira_talk] MIRA GPU
- From: Frank Huiskamp <frank.huiskamp@xxxxxx>
- To: mira_talk@xxxxxxxxxxxxx
- Date: Thu, 23 Apr 2009 11:42:45 +0200
Dear Bastien, dear all,
Since I know there are people interested in this I'd thought I'd share
this with the mailing list. This topic already came up about 2 months
ago by Gregory and Jan.
For my MSc thesis I'm working on GPU programming and in agreement with
my supervisor we have decided to see how MIRA can be optimized for GPUs.
Reading through the PhD thesis I was wondering how the described
algorithms are still implemented in the current MIRA versions. The
DNASAND and ZEBRA algorithm for example. I assume this is the skim part?
I found no reference to either algorithm names in the code and
unfortunately the code is not well documented.
I was also wondering if you had the time already to extract the numbers
I asked for last month? I've managed to compile Mira with the -pg
options and got some gprofiler results on the demo data sets. These
don't seem very representative as they all finish within 30 seconds.
Nevertheless it seems that computeBSimMatrix takes without a doubt most
of the time. Is this the banded SW algorithm as described in your PhD
thesis?
Having read part of your thesis I do think we can gain profit here from
using the GPU. I will need to have a closer look on the algorithm,
current SW GPU/SIMD implementations and the CUDA architecture to see how
your algorithm can be ported to the GPU.
The Skim part is also a very suitable candidate to transfer to the GPU,
but if this doesn't take much time already it wouldn't be worth the time
to transfer it. I have not yet looked into the other algorithms being used.
Of course, if anybody else has any benchmark results on big data sets I
would love to see them and get a better idea on where the bottlenecks in
MIRA are.
With kind regards,
Frank
--
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
Other related posts:
