[CaCL] [Lab-related] Info on u070
Jin, Lifeng
jin.544 at buckeyemail.osu.edu
Tue Oct 15 14:13:37 EDT 2019
Hi Cocomo members,
There has been some issues with u070. We discovered last week that the
PBS scheduler somehow thought u070 had only 32 cores, which made some
jobs which requested 32 cores block all other queued jobs for this
machine. It has 40 true cores or 80 fake cores with hyper threading.
Now the issue has been addressed, and I would like to make the following
suggestions:
1. There is an upper limit of 49 cores per machine that can be
requested, so don't request an excessive amount of cores even if you
know there should be more cores on a machine, especially if your code is
not optimized for parallel processing. You can request a ton of memory
though.
2. u070 is our main (only) GPU powerhouse, so if you plan to use it but
not the GPUs, try to save a few cores for others who use GPUs. Normal
GPU jobs don't use a ton of CPU cores. For example, you can request 40
cores from u070 so that 9 (fake) cores will be left for GPU jobs.
Lifeng
More information about the CaCL
mailing list