[Qemu-devel] 4 CPUs with vexpress is slow?

[Qemu-devel] 4 CPUs with vexpress is slow?

[Frederic Konrad]

Hi,

Testing MTTCG patch-set performance I found strange slowness with _upstream_
qemu (46bca5404b08201bb9df1ac32bc88fc7e6db1f74).

Basically booting a vexpress-a15 with "-smp 4" takes approximately 
forever if I use a
vexpress-v2p-ca15-tc1.dtb (2x A15) dtb and 39 secs if I hack this dtb to 
have 4 CPUs.

This is definitely strange because if I use "-smp 2" with the same guest 
image it takes
only 4 secs. And MTTCG patch-set seems to fix the issue as it boot in 6 
secs with
"-smp 4".

Is that a known issue or maybe it's my guest (linux 3.13.5)?

Fred

Re: [Qemu-devel] 4 CPUs with vexpress is slow?

[Peter Maydell]

On 22 June 2015 at 19:18, Frederic Konrad <fred.konrad@greensocs.com> wrote:
> Testing MTTCG patch-set performance I found strange slowness with _upstream_
> qemu (46bca5404b08201bb9df1ac32bc88fc7e6db1f74).
>
> Basically booting a vexpress-a15 with "-smp 4" takes approximately forever
> if I use a
> vexpress-v2p-ca15-tc1.dtb (2x A15) dtb and 39 secs if I hack this dtb to
> have 4 CPUs.

Wait, so if you change the dtb it boots in a non-infinite time,
but MTTCG boots both dtbs?

> This is definitely strange because if I use "-smp 2" with the same guest
> image it takes
> only 4 secs. And MTTCG patch-set seems to fix the issue as it boot in 6 secs
> with
> "-smp 4".
>
> Is that a known issue or maybe it's my guest (linux 3.13.5)?

Dunno. I basically never run with SMP configs because they're
always slower than uniprocessor... Does your kernel boot on
KVM? It may be that it's configured to assume 2xSMP somehow,
given the h/w we're modelling here is 2xSMP.

Important question: when you boot this config, does /proc/cpuinfo
say you have two cores booted, or all four? If we've ended up with
the other two cores in the config spinning then that would explain
why MTTCG is doing better here, but it's not a very interesting
config to try to optimise :-)

There is an issue where IPIs may be inefficient:
http://lists.nongnu.org/archive/html/qemu-devel/2015-06/msg03824.html
but I'd expect that to merely slow things down a bit, not send
the boot time off to infinity.

-- PMM

Re: [Qemu-devel] 4 CPUs with vexpress is slow?

[Frederic Konrad]

On 22/06/2015 22:57, Peter Maydell wrote:
> On 22 June 2015 at 19:18, Frederic Konrad <fred.konrad@greensocs.com> wrote:
>> Testing MTTCG patch-set performance I found strange slowness with _upstream_
>> qemu (46bca5404b08201bb9df1ac32bc88fc7e6db1f74).
>>
>> Basically booting a vexpress-a15 with "-smp 4" takes approximately forever
>> if I use a
>> vexpress-v2p-ca15-tc1.dtb (2x A15) dtb and 39 secs if I hack this dtb to
>> have 4 CPUs.
> Wait, so if you change the dtb it boots in a non-infinite time,
> but MTTCG boots both dtbs?
Yes MTTCG boots both dtbs.

>> This is definitely strange because if I use "-smp 2" with the same guest
>> image it takes
>> only 4 secs. And MTTCG patch-set seems to fix the issue as it boot in 6 secs
>> with
>> "-smp 4".
>>
>> Is that a known issue or maybe it's my guest (linux 3.13.5)?
> Dunno. I basically never run with SMP configs because they're
> always slower than uniprocessor... Does your kernel boot on
> KVM? It may be that it's configured to assume 2xSMP somehow,
> given the h/w we're modelling here is 2xSMP.
>
> Important question: when you boot this config, does /proc/cpuinfo
> say you have two cores booted, or all four? If we've ended up with
> the other two cores in the config spinning then that would explain
> why MTTCG is doing better here, but it's not a very interesting
> config to try to optimise :-)
All four cores are in /proc/cpuinfo.

>
> There is an issue where IPIs may be inefficient:
> http://lists.nongnu.org/archive/html/qemu-devel/2015-06/msg03824.html
> but I'd expect that to merely slow things down a bit, not send
> the boot time off to infinity.

This is definitely a bug :
Seems that removing "--nographic" option removes the problem it just 
boots as per
normal..

Fred

>
> -- PMM