Interesting CPU behavior
29 Sep 2019 12:31:59 UTC
Topic 219674
(moderation:
My DL360 server is equipped with 2 Xeon 6 core processors. After a few startup bumps, it has been running and crunching without issue since mid July. About a week ago, it appears that one of the processors began to get lazy. Looking at the temperature graph it was apparent that the second processor was not doing it's share of work.
Archae86 had mentioned a program called Process Lasso which give the user more precise control over CPU assignment and process control. Once installed and configured, the split CPU behavior was eliminated.
This program looks like strong medicine with a considerable amount of options, but, I would highly recommend installing it on any Windows box with multiple cores, especially a server.
A big thank you to Archae86 for mentioning it!
Clear skies,
Matt
Matt White wrote:My DL360
)
I have a similar problem with a X8DTL-iF. I keep it in the garage and had to add liquid cooling for both Xeons. They are cooled in tandem so the one closes to the radiator is cooler. I had to solve the overheating by reducing the speed of all cores, not just the one that runs hotter.
Thinking about this I got to wishing that the Boinc client would tag each work unit with the core much like the GPU is tagged. Not sure if that is even possible as threads and cores can change as the OS sees fit unlike the GPUs. It might be useful to permanently assign a hyperthread or more likely just the core. One could tell if something nefarious was happening by comparing various completion times for cores.
JStateson wrote:Thinking
)
Process Lasso does just that. You can balance and assign tasks to a work unit forcing all the cores to load share. It will tell you what cores are assigned to each task, even if you use the auto assign procedure.
It's a very nice tool.
Just to clarify, the working
)
Just to clarify, the working processor wasn't overheating. The 2nd processor was not doing its share of the work. I generally keep the processor utilization at 60%, which allows one work unit for each physical core, plus two GPU tasks.