Work balancing problem with mixed AMD/NVIDIA setup
27 Jan 2020 9:17:41 UTC
Topic 220553
(moderation:
Hi,
I've got this computer https://einsteinathome.org/host/12460398 crunching with 2 Radeon VII and 3 NVIDIA GTX 1080. Each AMD card is a factor of 4 faster than the NVIDIA card, however it seems that somehow the NVIDIA cards get "priority" when requesting new tasks. Currently I'm out of tasks for the Radeons but the NVIDIA cards have "too many tasks" (https://ibb.co/3M8Pr7j).
Any ideas how to tip the scales in favor of the AMD cards?
Thanks!
Aron wrote:Hi,I've got this
)
One way would be to install Boinc again in a different directory and exclude the AMD cards in one instance and exclude the Nvdidia gpu's in the other one, that way they both have their own cache.
Another way would be to just get another pc for one or the other brand. I used to find pc's on the curb on trash day when I first started, it helped build my collection. One warning is to be sure to wipe the harddrive BEFORE you hook it up to your network as some were FULL of viruses. I found Derek's Boot and Nuke an easy way to do this, just boot up the pc with the cd in the drive and after a few clicks it wipes the drive with ones and then zero's and then ones again. THE FBI says that doing that seven times makes the drive good enough to be put in the trash, I just did it three times and have never had a problem. https://dban.org/
What are your cache
)
What are your cache settings?
"too many tasks" is a message from BOINC, because the maximum limit is 1000 WUs per project.
I would set the cache to no more than 0.5 days if I were you. Then BOINC can request the needed WUs for AMD.
mikey wrote: One way would
)
Thanks for the reply! Regarding two boinc instances, are there any examples how this can be done? Would the work be counted towards two or one machines?
Regarding the second paragraph: all the cards are plugged into the same water-cooling loop (silent operation is a strict requirement from my partner), thus it would be a headache to try to jam two separate computers in there. So my current only option is rack-cases with pice-riser cards & one motherboard. Did someone try with virtual machines?
DF1DX wrote:What are your
)
Yes! This must be it! I increased the cache and then it got tons of new tasks, starting with NVIDIA, then got over 1000 wu and then was like "ok I'm full no more tasks thanks". Yikes! Is there any way to increase this upper WU limit?
IMHO this is a fixed limit of
)
IMHO this is a fixed limit of BOINC. I don't know of any configuration setting that can change this value.
Aron wrote:DF1DX wrote:What
)
It's a Serrver side setting so no we users can't change it.
mikey wrote:Aron wrote:DF1DX
)
Ok, thanks! I guess I'm going back to smaller cache lol!
DF1DX wrote:IMHO this is a
)
sounds like you are talking about the limit in the boinc client. In which case you certainly can go past this limit.
the easiest way is to edit the boinc client source code and recompile it, the limit is in the boinc client code. several of us at SETI have done this. I'm currently holding ~10,000 SETI tasks on each of my systems this way (a buffer against system outages, so I don't run out of work in the downtime).
_________________________________________________________________________
Ian&Steve C. wrote:DF1DX
)
Thanks for the info! I think recompiling the client is a bit above my abilities... I've set the cache down to 0.5 days and that did the trick for now.