Thank you very much for the clarification.
Le lun. 29 oct. 2018 à 23:32, Wang, Hongbo <hongbo.wang(a)intel.com> a écrit :
Right. It’s by design.
At the earler phase, KVMGT was designed to use simple round robin
scheduler, which means 1 signel VM will solely occupy 100% GPU time, multi
VMs will share the GPU time equally.
With the latest KVMGT version, we have added QoS policy in the scheduler
which is designed for Cloud usage, “pay as you go”. Your VM can’t use other
VM’s timeslice anymore even the 2nd vGPU is idle.
You can config or modify the scheduler code based on your need.
Tel: +86-21-6116 7445 <+86%2021%206116%207445>
MP: +86-1364 1793 689
*From:* iGVT-g [mailto:email@example.com]
*Sent:* Tuesday, October 30, 2018 10:31 AM
*To:* Adel Belkhiri <adel.belkhiri(a)gmail.com>; igvt-g(a)lists.01.org
*Subject:* Re: [iGVT-g] bug in KVMGT scheduler ?
I think it’s a known behavior in current codes. Every vGPU has a time
slice according to the ratio of its vGPU type’s weight against the summary
of all vGPUs’ weight. It will not be scheduled if its time slice is used up
in one time period.
*From:* iGVT-g [mailto:firstname.lastname@example.org
<igvt-g-bounces(a)lists.01.org>] *On Behalf Of *Adel Belkhiri
*Sent:* Tuesday, October 30, 2018 12:52 AM
*Subject:* [iGVT-g] bug in KVMGT scheduler ?
Hi Gvt-g developers,
I'm doing a performance analysis of KVMGT and I need your help. I noticed
that when there are (for example) two VMs having each a similar virtual
GPU, one VM would achieve only half performance even though the second VM
is idle (not sending requests to its vGPU).
Is that a normal behavior or is it a bug?
I traced KVMGT and looked in its source code and saw that when a vGPU
consume entirely its timeslice (50ms in case there are two vGPUs) during
one period, then it is scheduled out by gvt_service_thr.
Can anyone confirm if this is a normal behavior or not?
Host OS : ubuntu 16.05 (Kernel 4.14.15)
Guest OS : ubuntu 16.05 (Kernel 4.15)