The O2-All Sky Gravitational Wave Search on GPUs - discussion thread.

Betreger
Betreger
Joined: 25 Feb 05
Posts: 987
Credit: 1421641903
RAC: 798008

Well I spoke too soon, I got

Well I spoke too soon, I got 1 yesterday and 2 more today. 

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4266
Credit: 244924956
RAC: 16565

The problem depends on the

The problem depends on the data you are processing, not on the time when you are doing it.

Unfortunately the last two weeks saw a lot of empty desks here, due to illness and vacation, so there wasn't much progress from our side. However it looks like we found the problem (or at least one more), I'll publish a new app version in a few hours.

Then again the current charge of O2AS 'work' is going to end today. A new run is being set up already ("O2MD1") which will use the same application, in basically the same setting (CPU main application, GPU Beta Test, GPU results validated w. CPU results).

BM

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109393700068
RAC: 35831442

Bernd Machenschalk wrote:...

Bernd Machenschalk wrote:
... I'll publish a new app version in a few hours.

Thank you for the information about the new search and the new version.  Lots of people will be happy to know that things are progressing and that you've identified "at least one more" problem :-).

For the information of all interested volunteers, the new version of the GW GPU app (V1.09) is now listed on the applications page.  It's still listed under the heading "Continuous Gravitational Wave search O2 All-Sky" but it seems highly likely that it is the app for what Bernd has described as the "O2MD1" search.  It will take a little time for all the correct details to be put in place.

For anyone needing to comment/report/discuss anything related to the old search, please continue to do so here.  Shortly, I'll start a new discussion thread for the upcoming O2MD1 search so when tasks for the new version (V1.09 or any subsequent version) start to flow, please direct your comments about the new search to the new thread.  Once the transition has happened and the new search is well underway, I'll look at closing this thread at an appropriate time.

Cheers,
Gary.

Zalster
Zalster
Joined: 26 Nov 13
Posts: 3117
Credit: 4050672230
RAC: 0

Bernd Machenschalk wrote:The

Bernd Machenschalk wrote:

The problem depends on the data you are processing, not on the time when you are doing it.

Unfortunately the last two weeks saw a lot of empty desks here, due to illness and vacation, so there wasn't much progress from our side. However it looks like we found the problem (or at least one more), I'll publish a new app version in a few hours.

Then again the current charge of O2AS 'work' is going to end today. A new run is being set up already ("O2MD1") which will use the same application, in basically the same setting (CPU main application, GPU Beta Test, GPU results validated w. CPU results).

 

Bernd, can you speak to if the work units are actually using more than 1 CPU for each work unit?

Looks like I will need to modify my app_config for the new work units.  Sunday at the earliest.

Thanks Bernd.

 

Z

Jim1348
Jim1348
Joined: 19 Jan 06
Posts: 463
Credit: 257957147
RAC: 0

Zalster wrote:Bernd, can you

Zalster wrote:
Bernd, can you speak to if the work units are actually using more than 1 CPU for each work unit?

I can.  My RX 570 uses 103% of a core on an i7-4771 (Win7 64-bit).  There are two cores free from other duties.  I expect if you reserved one core, that would be enough.  It will probably take what it wants beyond that.

However, that depends of course on your GPU, CPU and OS.  The faster the GPU, the more support it will take, and the faster the CPU, the less of a core it will require.  Also of course there is Nvidia v. AMD.  I think Nvidia will use more, since it translates OpenCl to CUDA, but I have not checked it out recently.

If you use BoincTasks, you can readily check out the CPU usage and make the appropriate reservations in your case.

Richie
Richie
Joined: 7 Mar 14
Posts: 656
Credit: 1702989778
RAC: 0

I haven't seen O2MD1 arriving

I haven't seen O2MD1 arriving yet but in this transition phase the v1.09 O2AS20 tasks seem to be running without problems. All of my results are still 'pending' though. Speed for 1x is roughly 1 percent per 1 min with all my cards (GTX960, RX570/580, R9 390) and Xeon / i7 @ ~3.8GHz

edit: First one of those validated... Nvidia (windows) against cpu (windows)...
https://einsteinathome.org/workunit/419780840

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109393700068
RAC: 35831442

Richie wrote:I haven't

Richie wrote:
I haven't seen O2MD1 arriving yet but in this transition phase the v1.09 O2AS20 tasks seem to be running without problems.

That's interesting, thanks for reporting.  Looks like they've decided to perhaps 'pre-test' the new app by using it to finish off the remaining O2AS tasks which were supposedly finishing yesterday.

I was just setting up a couple of hosts for when the O2MD1 run started.   These hosts still have plenty of FGRPB1G to churn through so I'll just wait a while until the search name changes in case there are some data changes (as well as the app change) which need to happen.  It's probably a bit of a nightmare for poor Bernd trying to get everything into place for the transition.  I'm sure there a lots of little things that, if overlooked, could easily become showstoppers :-).

It was interesting to note that your immediate validation was facilitated by the fact that you had a resend for an existing 2 CPU task quorum where there was a completed task as well as the failed task.

Cheers,
Gary.

Zalster
Zalster
Joined: 26 Nov 13
Posts: 3117
Credit: 4050672230
RAC: 0

I see some v1.09 but they

I see some v1.09 but they still say this

h1_0583.75_O2C02Cl1In0__O2AS20-500_583.90Hz_254

1 Oct 2019 22:15:41 UTC

Maybe tomorrow we will see the O2MD1?

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109393700068
RAC: 35831442

Zalster wrote:Maybe tomorrow

Zalster wrote:
Maybe tomorrow we will see the O2MD1?

Seeing as it's already 'tomorrow' in Hannover, it might actually happen in just a few more hours :-).

However, there's probably lots of stuff to tweak so it could easily slip a day or two.  It's that standard old saying - it'll be ready when it's ready :-).

Cheers,
Gary.

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109393700068
RAC: 35831442

Seeing as there will likely

Seeing as there will likely be a bit of a delay until O2MD1 gets going and seeing as V1.09 has been described as "fixing a problem" that was in earlier versions, I've decided to try running some O2AS tasks with this version on some of the older hosts of mine with decent GPUs to see if there is any change for the better in the validation department.  A quick recap is to point out that when I previously tried crunching with older CPUs like my Q6600s from 2008, using relatively modern Polaris GPUs and running the V1.07 app, tasks would crunch without compute errors but most or all of the results would eventually end up being declared invalid, even with tasks crunched singly.

I've downloaded the V1.09 app and a bunch of O2AS tasks on one such host and one task has already completed without issue and has been returned.  As you would expect, it shows as a 'pending'.

The full batch of new tasks includes around 6 'resends' and I've promoted all of these to now crunch first in the hope of speeding up the validation process.  Hopefully, there will either be valid results or inconclusives quite quickly.  Either way, this should help me decide whether or not quite old CPU architectures can support the production of valid results - or not.  This is something I would like to know before O2MD1 actually starts sending work for GPUs.

Cheers,
Gary.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.