[Gluster-devel] High CPU consumption on Windows guest OS when libgfapi is used

Vijay Bellur vbellur at redhat.com
Tue Apr 19 23:18:24 UTC 2016


On Tue, Apr 19, 2016 at 1:24 AM, qingwei wei <tchengwee at gmail.com> wrote:
> Hi Vijay,
>
> I rerun the test with gluster 3.7.11and found that the utilization still
> high when i use libgfapi. The write performance is also no good.
>
> Below are the info and results:
>
> Hypervisor host:
> libvirt 1.2.17-13.el7_2.4
> qemu-kvm 2.1.2-23.el7.1
>
>
> WIndows VM:
> Windows 2008R2
> IOmeter
>
> Ubuntu VM:
> Ubuntu 14.04
> fio 2.1.3
>
>
> Windows VM (libgfapi)
>
> 4k random read
>
> IOPS 22920.76
> Avg. response time (ms) 0.65
> CPU utilization total (%) 82.02
> CPU Privilegde time (%) 77.76
>
> 4k random write
>
> IOPS 4526.39
> Avg. response time (ms) 3.26
> CPU utilization total (%) 93.61
> CPU Privilegde time (%) 90.24
>
> Windows VM (fuse)
>
> 4k random read
>
> IOPS 14662.86
> Avg. response time (ms) 1.08
> CPU utilization total (%) 27.66
> CPU Privilegde time (%) 24.45
>
> 4k random write
>
> IOPS 16911.66
> Avg. response time (ms) 0.94
> CPU utilization total (%) 26.74
> CPU Privilegde time (%) 22.64
>
> Ubuntu VM (libgfapi)
>
> 4k random read
>
> IOPS 34364
> Avg. response time (ms) 0.46
> CPU utilization total (%) 6.09
>
> 4k random write
>
> IOPS 4531
> Avg. response time (ms) 3.53
> CPU utilization total (%) 1.2
>
> Ubuntu VM (fuse)
>
> 4k random read
>
> IOPS 17341
> Avg. response time (ms) 0.92
> CPU utilization total (%) 4.22
>
> 4k random write
>
> IOPS 17611
> Avg. response time (ms) 0.91
> CPU utilization total (%) 4.65
>
> Any comments on this or things i should try?
>

Can you please share your gluster volume configuration? It might be
worth checking if the tunables in profile virt
(extras/group-virt.example) are applied on this volume.

Additionally I would also try to:
1 disable write-behind in gluster to see if there is any peformance
difference for writes
2 use perf record -p <qemu> followed by perf annotate to observe the
hot threads.

HTH,
Vijay


More information about the Gluster-devel mailing list