[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [HTCondor-users] GPUs_MONITOR resource usage




On 2019/01/10 23:40, Todd L Miller wrote:
Â39637 ioctl(0, _IOC(_IOC_READ|_IOC_WRITE, 0x46, 0x2a, 0x20),

ÂÂÂÂI'm assuming this is the nvml library talking to the GPU driver, but I don't know why that would use so much CPU (or even if that's where it is using all that CPU). If you get a chance, I'm curious what the CPU usage of the monitor is if the machine's GPUs are idle.

Hi Todd,

*) This was on an totally idle system. On all 9 test machines I have 6 to 9% CPU consistently.
*) 'perf top' shows that it is indeed the nvml library; more in detail: (to 5% only)

[root@postel ~]# perf top -p 1254
  32.47%  [kernel]                [k] _nv031472rm
  19.52%  [kernel]                [k] pci_conf1_read
   5.05%  [kernel]                [k] _raw_spin_lock_irqsave
   4.28%  [kernel]                [k] _nv033593rm
   1.59%  [kernel]                [k] _nv007572rm
   1.30%  [kernel]                [k] _nv034648rm
[...]

This is with libnvidia-ml.so.415.25 on a GeForce GTX 960

B.
 ÂÂÂÂIn the meantime, it shouldn't be too hard to turn off the GPU monitor entirely, if you'd prefer.

- ToddM