How to run different number of work units on different cards?
23 Mar 2014 19:56:09 UTC
Topic 197478
(moderation:
I have a GTX 660 and a GTX 650 Ti running under WinXP (BOINC 7.3.11) on the same machine. I would like to run 2 BRP4G/BRP5 work units on the 650 Ti, and 3 work units on the 660. Is this possible?
(I have used cc_config.xml and app_config.xml files as necessary to do it on different machines, but don't see how to do it on the same machine.)
How to run different number of work units on different cards?
)
Not without going through a few hoops. You would have to run 2 instances of Boinc and exclude one GPU from each one. I don't know how to do that but I know it's been done by participants on other projects. This would also mean 2 host IDs, the server would think it's 2 separate machines.
Another way to do something similar (I think, I haven't tried it myself) would be to use the tag in cc_config.xml to only run BRP4G on one GPU and BRP5 on the other, you could then use the app_config.xml to set how many to run at once.
Something like this should exclude BRP4G from device 0:
[pre]
einstein.phys.uwm.edu
0
NVIDIA
einsteinbinary_BRP4G
[/pre]
Holmis, Thanks. I think
)
Holmis,
Thanks. I think your second option is better for me. I have looked into running multiple instances of BOINC myself for other purposes, and it is not a pleasant thought.
RE: Not without going
)
Exactly I saw this at PrimeGrid and the guy says it works for him:
http://www.primegrid.com/
1
genefer_wr
http://www.primegrid.com/
0
genefer
He said the genefer_wr and the genefer units are NOT the same units. So he is doing exactly what you said.
And then I think an app_info.xml file with similar exclusions like this:
milkyway
milkyway_separation_1.02_windows_x86_64__opencl_amd_ati.exe
milkyway
102
ATI
0.5
milkyway_separation_1.02_windows_x86_64__opencl_amd_ati.exe
mikey, It looks like that
)
mikey,
It looks like that will do it too. I have never seen those units used that way before in cc_config.xml and would never have found them myself. Thanks also.
EDIT: It looks like this is the discussion you are referring to, and helps illuminate the subject.
http://www.primegrid.com/forum_thread.php?id=5488&sort=6
RE: mikey, It looks like
)
Yes that's the one. That is one of the advantages in running multiple projects sometimes, you see something one place that can help you or someone else on another project.
So I know this is an older
)
So I know this is an older thread, but is there a way to run different number of the same work units on different GPUs?
Depending on the type of motherboard, the PCIe can be either a x16, x8, x4
So on the faster slot I would want to run 2 or more and on the slower I might just want to run 1
Is there a way to do this?
RE: So I know this is an
)
Do you mean in the same pc or different pc's? Different pc's sure thru some config files, the same pc the only way I know would be by running separate projects on each gpu.
Same computer. Thanks
)
Same computer.
Thanks
RE: Depending on the type
)
I presume you are talking about the second or subsequent PCIe full length slot. If you are referring to LGA 1150 type boards and CPUs, there are a total of 20 lanes available so the first slot will use 16 of those leaving only 4 for the second slot. If you populate both slots, I think you will get x8 x8 but you would need to check the specs of an individual board to be sure about that. EDIT: Correction - LGA 1155 CPUs (Sandy Bridge / Ivy Bridge) has 20 lanes. LGA 1150 (Haswell) has 16 lanes.
When HB (Bikeman) introduced the optimisations for the BRP6 app earlier this year, it made a huge difference to the PCIe bandwidth requirements. Before the change, using a version 1.x PCIe x16 slot suffered a large penalty compared to a V2 slot. I know because I was running HD7850 cards in both types of slots. Now there is virtually no difference. This leads me to believe that a board with 2 full length slots running x8 x8 would be able to run these GPUs to pretty much their full potential. I'm intending to test this out shortly. My current hosts only have single full length slots but yesterday I found a nice cheap B85 based board with 2 full length slots. I'll probably get around to trying this out shortly.
EDIT: I've done a bit more research and it seems like I should be looking at Z87 chipset rather than B85 if I wanted to get x8 x8 for the two slots. So exactly what motherboard do you have and what two GPUs are you wanting to run in it?
If the GPUs are the same, why don't you just insert them both and see if you get x8 x8? If so, try them both running the same number of tasks and see how that compares with running just one GPU (at x16). You might find there isn't a problem to solve after all.
Apart from running 2 separate BOINCs in separate trees with separate host IDs, I don't think you could run the same task type at different concurrencies.
Cheers,
Gary.