Minor QTC Integration Items

vetter / shoc

The SHOC Benchmark Suite

Other

247 stars 102 forks source link

Minor QTC Integration Items #4

Open kspaff opened 12 years ago

kspaff commented 12 years ago

Testing issue tracker with a couple of minor action items for QTC integration.

@adanalis should check and make sure:

What I put on the wiki page (https://github.com/spaffy/shoc/wiki/Qtc) looks correct and maybe add the outputs. I haven't run QTC yet to see what metric the result is.
Move QTC to Level 2 directory
Add the -s parameter and sizes, then I will integrate into driver script and google results db

rothpc commented 12 years ago

Commit 18925dcf9fce8108558ffd0989946b219681ec4b moves QTC to level 2.

rothpc commented 12 years ago

Commit 1b83eb7cd7064a216f84ca5d3a5b95fa441e27e5 integrates QTC into driver script

kspaff commented 12 years ago

@adanalis QTC problem sizes need a little adjustment. -s 4 took 81 minutes on the Kepler in newark.

kspaff commented 12 years ago

Also, the only difference I notice in the output is that the Kepler (gtx680) is allocating much less texture memory (193MB vs. 1024MB on the Tesla M2090 in Keeneland)

adanalis commented 12 years ago

This is because "-s 4" is supposed to be badass :-)

On Jun 6, 2012, at 10:22 AM, Kyle Spafford wrote:

@adanalis QTC problem sizes need a little adjustment. -s 4 took 81 minutes on the Kepler in newark.

Reply to this email directly or view it on GitHub: https://github.com/spaffy/shoc/issues/4#issuecomment-6151762

rothpc commented 12 years ago

I don't think it takes that long on a M2090, so this seems to be a problem with the mapping of qtc to kepler.

On Jun 6, 2012, at 11:08 , adanalis wrote:

This is because "-s 4" is supposed to be badass :-)

On Jun 6, 2012, at 10:22 AM, Kyle Spafford wrote:

@adanalis QTC problem sizes need a little adjustment. -s 4 took 81 minutes on the Kepler in newark.

Reply to this email directly or view it on GitHub: https://github.com/spaffy/shoc/issues/4#issuecomment-6151762

Reply to this email directly or view it on GitHub: https://github.com/spaffy/shoc/issues/4#issuecomment-6153097

Philip C. Roth | +1 865 241-1543 | http://ft.ornl.gov/~rothpc

kspaff commented 12 years ago

Yeah, total runtime on 1 M2090 GPU on Keeneland is only ~68 seconds, so the slowdown is quite surprising.

adanalis commented 12 years ago

It's probably interesting to tune it. There are a couple of parameters that are easily tunable and should have a large effect on performance (i.e. number of threads/TB and numbers of registers per thread). I will look into it.

On Jun 6, 2012, at 11:27 AM, Kyle Spafford wrote:

Yeah, total runtime on 1 M2090 GPU on Keeneland is only ~68 seconds, so the slowdown is quite surprising.

Reply to this email directly or view it on GitHub: https://github.com/spaffy/shoc/issues/4#issuecomment-6153599