Request to increase task limit per core
16 Jun 2012 6:08:11 UTC
Topic 196400
(moderation:
Hello,
I have request. I am running a 6-core processor on one of my hosts and have the max CPU usage set to 59% currently. Normally I run 50% as to avoid slow down from Hyperthreading and also due to summer heat which drives up CPU temps. HT drags down performance of the GPU tasks. I am hitting the quota limit of 280 tasks (7x40) in a day and 240 tasks yesterday when I was set to 50% usage (6x40). My question is if it would be possible to have the per core limit of 40 tasks increased to 50 tasks? This way my host can run up to 300 tasks (6x50) in a day with 50% max CPU usage for avoiding HT performance loss and extra heat. Thanks.
Request to increase task limit per core
)
Yes, I also see this happening on a less extreme system of mine, a Core2 Dual with 2 GPUs and one core set aside to drive the ATI/AMD card. The resulting 40 tasks per day is limiting.
However, the real fix would probably to include the GPUs in the calculation, not just bumping up the limit of jobs per core per day which does have some justification in the absence of GPUs.
The workaround I have until this is fixed is to have a little cc_config.xml file which fools BOINC into assuming the host has more CPUs (say twice as many) than it really has. I tell boinc to re-read this file on-the-fly (with boinccmd command line interface, see boinccmd --help) , ask boinc to update the project a few times to fetch work (using boinccmd as well), and then revert to a cc_config file that has the true number of CPUs (and re-load it again with cc_config.xml).
But yes, this is only a workaround and we need to address the root cause.
Cheers
HBE
I also find the daily
)
I also find the daily download limit based only on allocated CPU cores to be an awkward limitation on my GPU-employing systems. My plea posted here a couple of months ago may perhaps be added to yours now.
RE: Yes, I also see this
)
Thank you for the suggestion with the cc_config.xml. I will give this configuration change a try and also see if I can script something for automation.
Jeroen
Hi! Bernd has just changed
)
Hi!
Bernd has just changed the server config to allow more work being distributed to our GPU crunching volunteers. The new settings are
Max tasks per day = 32 * (number of usable CPU cores) + 160 * (number of reported GPUs) , where the number of cores is limited to max 64 and the max # of GPUs is 8.
The previous formula was:
Max tasks per day = 40 * (number of usable CPU cores) , where the number of cores was limited to max 16.
We'll see how this works out, let us know if there is still a problem with GPUs running dry because of a daily quota issue.
All this is a compromise between allowing clients to download as many WUs as they request (which could be, for a (say) 7 day cache and 3 top notch graphics cards ca 1500 jobs on the initial day!!!) and protecting the server infrastructure against clients that accidentally try to fetch work in an endless loop.
Cheers
HB
Hello, That is great news.
)
Hello,
That is great news. Thank you and to Bernd for updating the server to allow for a greater number of tasks per GPU. I had temporarily disabled CPU tasks to help prevent GPUs from running out of tasks but have been wanting to get the CPU projects up and running again. With this change, I have re-enabled running Gravitational-wave and Gamma-ray searches. I am glad to have those projects running.
Thanks again.
Jeroen
Has anyone observed any
)
Has anyone observed any problems regarding the new quota setting?
BM
BM
I have not seen any problems
)
I have not seen any problems and my host is no longer running into a daily task limit.
No new problems here, and it
)
No new problems here, and it solved a problem, in that I was able to reduce the CPU jobs running to one while still being able to get an adequate supply of GPU work. This suits my summertime power reduction scheme, as the power productivity of incremental CPU work is far lower than that of incremental GPU work on my hosts. I also throttle the hosts using Fred's tool, cutting back power consumption more when the room is hotter than when cooler. I should be back up to full production by early October.
Thanks for the feedback guys.
)
Thanks for the feedback guys. The suspected download throughput problem (moved to a separate thread) was also resolved (was actually a user defined throttling), so everything looks fine now in the high performance GPU camp :-)
Happy crunching
HB
RE: Has anyone observed any
)
I have 2 systems that I think are running into this limit:
http://einsteinathome.org/host/5771544
and
http://einsteinathome.org/host/5860433
At least they seem to be productive on GPU tasks but neither have any running at certain times.
How can I tell if the reason is this limit?
Joe