Mailing List Archive

> for xl). Are people still dedicating CPU cores to dom0 in modern
> versions of Xen with credit2 scheduler?
>
> If I put this on the hypervisor command line:
>
> dom0_max_vcpus=2 dom0_vcpus_pin
>
> how do I know which PCPUs dom0 will be pinned to?


I do this, but then I take it a little bit further, especially with NUMA hosts.


Let’s suppose I have a dual-socket 8-core CPU machine with HT enabled. 32 logical CPUs, right?

I use those hypervisor command line configs and some commands in rc.local to end up with the following:


# xl cpupool-list
Name CPUs Sched Active Domain count
Pool-dom0 4 credit y 1
Pool-CPU1 12 credit y 5
Pool-CPU2 16 credit y 3


dom0 ends up being assigned the first 2 physical cores on the first CPU (4 logicals)

I then have cpu pools for the remainder of the first cpu and then the whole entire second CPU.

This gives me plenty of dedicated horsepower for handling IO and s/w RAID ops.

xl cpupool-numa-split
xl cpupool-rename Pool-0 Pool-dom0;

xl cpupool-cpu-remove Pool-dom0 4
xl cpupool-cpu-remove Pool-dom0 5
.
.
.

xl cpupool-create name=\"Pool-CPU1\" cpus=["4","5","6","7","8","9","10","11","12","13","14","15"]
xl cpupool-create name=\"Pool-CPU2\" cpus=["16","17","18","19","20","21","22","23","24","25","26","27","28","29","30","31"]


—
Brian

> cpus = "all,^0,^1"

I rather share all CPUs and let the scheduler prioritize dom0 as such
(default weight is 256):

xl sched-credit2 --domain=0 --weight=512

The only drawback is that I need to be careful on what I do on the host,
to let some juice available for the guests. The good thing about it is
that the dom0 also shares all cores, I guess)

Sorry it's not a definitive answer, as I did not compare and
stress-bench this against an old-school pinning setup.

--
Pierre-Philipp Braun
SMTP Health Campaign: enforce STARTTLS and verify MX certificates
<https://nethence.com/smtp/>

Some things I feel comfortable entrusting to the hypervisor scheduler; some things I don’t.

For the latter, it gives me more piece of mind to dedicated full blocks of resources.

IO and RAID ops (being core to almost any other operation performed on the machine) fall into that second category.


My $0.02.



> On Jan 31, 2021, at 3:43 PM, Pierre-Philipp Braun <pbraun@nethence.com> wrote:
>
>> cpus = "all,^0,^1"
>
> I rather share all CPUs and let the scheduler prioritize dom0 as such (default weight is 256):
>
> xl sched-credit2 --domain=0 --weight=512
>
> The only drawback is that I need to be careful on what I do on the host, to let some juice available for the guests. The good thing about it is that the dom0 also shares all cores, I guess)
>
> Sorry it's not a definitive answer, as I did not compare and stress-bench this against an old-school pinning setup.