Hi Dan, On 3 May 2017 at 17:43, Dan van der Ster <dan@xxxxxxxxxxxxxx> wrote: > We use cpu_dma_latency=1, because it was in the latency-performance profile. > And indeed by setting cpu_dma_latency=0 on one of our OSD servers, > powertop now shows the package as 100% in turbo mode. I tried both 0 and 1 and didn't notice a difference in the effective frequency, though I think on most contemporary Intel systems this would/should allow the CPU to transition to C1/C1E. You can check the values that the pmqos interface uses to determine the transition latencies from various C-states via: `sudo find /sys/devices/system/cpu/cpu0/cpuidle -name latency -o -name name | xargs cat` `cpupower monitor` is another good option for inspecting this. > So I suppose we'll pay for this performance boost in energy. > But more importantly, can the CPU survive being in turbo 100% of the time? If it does not then Intel has messed up - the CPUs will still thermal throttle. You can see the natural production variation in any burn-in/acceptance testing of new homogenous cluster gear - anywhere from 5-8% CPU performance variation is expected. It's certainly not uncommon for HPC installations to run continuously with these options statically tuned (via kernel command line), in fact Mellanox recommends this in their tuning documentation, and we have certainly found it helps in a virtualised HPC environment - I've read anecdotal reports that the pstate driver is not as effective for such workloads but do not know if there is any hard evidence for that in the code. The real question is why the performance loss is so bad in a default configuration. -- Cheers, ~Blairo _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |