OK thanks arkayn

Please, how much is the advantage of crunching 2 tasks in parallel?
This question is, as the GPU load for these files is very high - 90% or even more - and the CPU load very little - much less than 1% - for single task crunching. So one would expext only some advantage during the change of tasks at the end of a task.
Kind reagrds
Martin

Please, how much is the advantage of crunching 2 tasks in parallel?
This question is, as the GPU load for these files is very high - 90% or even more - and the CPU load very little - much less than 1% - for single task crunching. So one would expext only some advantage during the change of tasks at the end of a task.
Kind reagrds
Martin

It is not less than 1%. Every GPU work unit uses several seconds of CPU time at the end. For example, my card crunches a WU in about 1:06, the last 0:06 is CPU only. Crunching two WUs at once allows my card to work on the other GPU WU while the CPU is finishing the first.

Hallo Swiftmallard!
I´m using a GT440 and a GTX550Ti and here are some data I took in the last few days:

GT440:
Run Time 1498.22 +/- 122.06 [sec]
CPU Time 8.63 +/- 3.55 [sec]
That does give a total relative CPU Load of 0.576 +/- 0.245 [%]
That was derived from 61 observed tasks.

GTX550Ti:
Run Time 660.83 +/- 82.60 [sec]
CPU Time 5.94 +/- 1.61 [sec]
That does give a total relative CPU Load of 0.899 +/- 0.268 [%]
That was derived from 277 observed tasks.

The Run Times and CPU Times are taken from my Results Page at MW.
All GPU taks where running at priority level of "Above Normal".

I´d observe for a short time only the CPU load by the MS Process Explorer and found only about 0.2% load for these tasks, if I remember correctly. (I didn´t put this data to paper.)

As you mentioned

Every GPU work unit uses several seconds of CPU time at the end.

All that explaines easily, why the the realtive CPU Load for the slower card is less. From this I would conclude, that only very fast GPU cards will realy profit from crunching multitasks on a single GPU.

I would be pleased to get your and other comments.

Kind regards
Matin

---

Hallo Swiftmallard!
I´m using a GT440 and a GTX550Ti and here are some data I took in the last few days:

GT440:
Run Time 1498.22 +/- 122.06 [sec]
CPU Time 8.63 +/- 3.55 [sec]
That does give a total relative CPU Load of 0.576 +/- 0.245 [%]
That was derived from 61 observed tasks.

GTX550Ti:
Run Time 660.83 +/- 82.60 [sec]
CPU Time 5.94 +/- 1.61 [sec]
That does give a total relative CPU Load of 0.899 +/- 0.268 [%]
That was derived from 277 observed tasks.

The Run Times and CPU Times are taken from my Results Page at MW.
All GPU taks where running at priority level of "Above Normal".

I´d observe for a short time only the CPU load by the MS Process Explorer and found only about 0.2% load for these tasks, if I remember correctly. (I didn´t put this data to paper.)

As you mentioned

Every GPU work unit uses several seconds of CPU time at the end.

All that explaines easily, why the the realtive CPU Load for the slower card is less. From this I would conclude, that only very fast GPU cards will realy profit from crunching multitasks on a single GPU.

I would be pleased to get your and other comments.

Kind regards
Matin

It also depends on the memory your gpu has onboard, a card with 512 meg of memory would not do nearly as well as a card with 3 gig of memory when doing multiple units at once. You need to know how much of the available gpu memory each workunit is using. Then double that and see how much leftover there is before you get to the total memory the gpu has on it. If there is not enough headroom, which varies by project, then it is not effective to crunch more than one unit at a time. To crunch the fastest the WHOLE workunit needs to fit in the gpu cards memory, with enough to spare so it can actually crunch.

In principle correct, but MW doesn't need much memory. I think any card with DP should be fine running 2 WUs at once. And running more than that is pointless at MW anyway.

MrS

---

Scanning for our furry friends since Jan 2002

Anyone could post a new working app_info.xml for 2 WUs for use with the 670/690?

Anyone could post a new working app_info.xml for 2 WUs for use with the 670/690?

If MW would update the Server Software side we could do it via the website and NOT need an app_info file, like at Einstein now. Updating the Server Software OFTEN causes problems and can be very time consuming in the long run.

Anyone could post a new working app_info.xml for 2 WUs for use with the 670/690?

If MW would update the Server Software side we could do it via the website and NOT need an app_info file, like at Einstein now. Updating the Server Software OFTEN causes problems and can be very time consuming in the long run.

Einstein uses very old server code, with a few extra bells and whistles they've written themselves (this is one of them). So even with a server update, this feature wouldn't automatically appear.

However, all is not lost. With a new client feature, you can run multiple tasks without either a full app_info.xml file or a server upgrade.

BOINC 7.0.40-42 and new app_config.xml

Anyone could post a new working app_info.xml for 2 WUs for use with the 670/690?

If MW would update the Server Software side we could do it via the website and NOT need an app_info file, like at Einstein now. Updating the Server Software OFTEN causes problems and can be very time consuming in the long run.

Einstein uses very old server code, with a few extra bells and whistles they've written themselves (this is one of them). So even with a server update, this feature wouldn't automatically appear.

However, all is not lost. With a new client feature, you can run multiple tasks without either a full app_info.xml file or a server upgrade.

BOINC 7.0.40-42 and new app_config.xml

Now THAT is cool!!!! THANKS!!!