|
Hi Boris,
[Appreciate for your support, and I have one last question]
Since we’re using the Xen hypervisor (not the KVM that is type-II), I think we cannot use the
perf kvm command. Here are the outputs:
--------
# perf kvm --host --guest record -C 1 sleep 1
[ perf record: Woken up 1 times to write data ]
Warning:
1 unprocessable samples recorded.
Do you have a KVM guest running and not using 'perf kvm'?
[ perf record: Captured and wrote 0.242 MB perf.data.kvm ]
# perf kvm --guest --host report --stdio
Warning:
1 unprocessable samples recorded.
Do you have a KVM guest running and not using 'perf kvm'?
Error:
The perf.data.kvm file has no samples!
# To display the perf.data header info, please use --header/--header-only options.
#
--------
In addition, we used perf to collect hardware performance counters, not the symbols, kernel functions, or other stuffs. Thus, what we really need is that we need to separate the hardware counters by each domain if it’s possible (i.e. we want to
verify some counters are from Dom0, other counters are from Dom1, and the others come from Dom2, etc.)
I’m pretty much sure that the code you used to distinguish each domain from samples will work. I wonder if you can agree with that.
Best,
Sanghyun.
On
08/29/2016 05:48 PM, Sanghyun Hong wrote:
I really appreciate for your answers. Last but not the least, I want to make it more clear:
The hypervisor will provide dom0 with a raw sample (guest's, dom0's or
hypervisor's) and then it's the job of dom0 kernel to properly tag and
format it and make it available to the userland (i.e. perf itself).
Does this mean if I run the perf command on Dom0 while other domains are running, we can collect the performance counter values without distinguishing each domain, right?
With
'self' mode each domain collects samples only for itself.
And
for 'all' mode apparently this kind of works. With perf build from
Linux
4.7 tree:
root@haswell>
xl vcpu-list
Name
ID VCPU CPU State Time(s)
Affinity
(Hard / Soft)
Domain-0
0 0 0 -b- 55.0 0 / all
Domain-0
0 1 1 r-- 35.9 1 / all
Domain-0
0 2 2 -b- 36.9 2 / all
Domain-0
0 3 3 -b- 36.2 3 / all
fedora
1 0 1 -b- 183.7 1 / all
root@haswell>
./perf kvm --guest --host --guestkallsyms=/tmp/kallsyms
record
-C 1 sleep 1
[
perf record: Woken up 1 times to write data ]
[
perf record: Captured and wrote 0.139 MB perf.data.kvm (60 samples) ]
root@haswell>
./perf kvm --guest --host --guestkallsyms=/tmp/kallsyms
report
--stdio
...
#
Overhead Command Shared Object
Symbol
#
........ ............... .......................
.............................................
#
14.48%
swapper [unknown] [k] 0xffff82d0801c7a37
6.06%
swapper [kernel.kallsyms] [k]
update_blocked_averages
4.38%
swapper [unknown] [k] 0xffff82d08013690f
2.75%
swapper [kernel.kallsyms] [k] update_rq_clock
2.34%
swapper [kernel.kallsyms] [k] irq_enter
2.33%
[guest/0] [guest.kernel.kallsyms] [g]
fb_pad_aligned_buffer
2.24%
1.hda-0 [kernel.kallsyms] [k]
_raw_spin_unlock_irqrestore
2.11%
[guest/0] [guest.kernel.kallsyms] [g] native_safe_halt
2.05%
[guest/0] [guest.kernel.kallsyms] [g]
update_blocked_averages
...
"unknown"
samples are from hypervisor.
The tool will then look at the tag and display the event as belonging to
host (dom0, really) or guest. This is supported for KVM, with 'perf kvm'
commands.
So if you can run perf kvm commands then you may be able to
differentiate dom0's and guest's samples (you will also see hypervisor
samples but they won't get resolved to symbols). I am just not sure per
kvm will run on a non-KVM host. I had a private copy where it did but
this was based on a fairly old version of Linux.
(Also, stack profiling is not supported at all).
I wonder if you can share the private copy of your code. I’m looking into the source code of the patches, and it seems there’s a way to do it. Even if your codes are outdated, I think I can manage to port it. I will promise once I finish the porting, I will
share my adjustment of your code as well.
Looks
like you may not need it.
-boris
Best,
Sanghyun.
On Aug 29, 2016, at 5:24 PM, Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx> wrote:
On 08/29/2016 05:08 PM, Sanghyun Hong wrote:
Yes, this will allow the hypervisor to collect samples from multiple
guests. However, the tool (perf) probably won't be able to properly
process these samples. But you can try.
I understand, thus, I applied the patches and set
the /pmu_mode/ to *all*. However, I’m really curious what you mean by
the tool (*perf*) probably won’t be able to properly process these
samples. Is there any things I have to have in mind while collecting
the counters, or will it have incorrect values of the counters?
The hypervisor will provide dom0 with a raw sample (guest's, dom0's or
hypervisor's) and then it's the job of dom0 kernel to properly tag and
format it and make it available to the userland (i.e. perf itself). The
tool will then look at the tag and display the event as belonging to
host (dom0, really) or guest. This is supported for KVM, with 'perf kvm'
commands.
So if you can run perf kvm commands then you may be able to
differentiate dom0's and guest's samples (you will also see hypervisor
samples but they won't get resolved to symbols). I am just not sure per
kvm will run on a non-KVM host. I had a private copy where it did but
this was based on a fairly old version of Linux.
(Also, stack profiling is not supported at all).
You will want to run dom0 on all physical processors (i.e. no
dom0_max_vcpus boot option) and pin all VCPUs for both dom0 and the
guest.
Sure, thanks! My machine has four physical cores, and I’ve set
the /dom0_max_vcpus=4/. I think it will have the same effect in
removing the /dom0_max_vcpus/ option.
Yes.
-boris
Best,
Sanghyun.
On Aug 29, 2016, at 4:59 PM, Boris Ostrovsky
<boris.ostrovsky@xxxxxxxxxx <mailto:boris.ostrovsky@xxxxxxxxxx>> wrote:
On 08/29/2016 02:42 PM, Sanghyun Hong wrote:
Hi Boris,
I’ve found the
documentations(https://github.com/torvalds/linux/blob/master/Documentation/ABI/testing/sysfs-hypervisor-pmu)
in the kernel source code, and it seems like if we change the mode
from *self* to *all* then we can collect counters for all the domains,
right?
Yes, this will allow the hypervisor to collect samples from multiple
guests. However, the tool (perf) probably won't be able to properly
process these samples. But you can try.
You will want to run dom0 on all physical processors (i.e. no
dom0_max_vcpus boot option) and pin all VCPUs for both dom0 and the
guest.
-boris
All the best,
Sanghyun.
On Aug 29, 2016, at 12:36 PM, Boris Ostrovsky
<boris.ostrovsky@xxxxxxxxxx <mailto:boris.ostrovsky@xxxxxxxxxx>
<mailto:boris.ostrovsky@xxxxxxxxxx>> wrote:
On 08/29/2016 09:18 AM, Sanghyun Hong wrote:
Dear Xen-Devel Community:
I’m a grad student working on measuring performance counters at the
Xen domains. I read this
thread(https://wiki.xenproject.org/wiki/Xen_Profiling:_oprofile_and_perf)
in
web, and it says using Linux perf command will let us collecting
performance counters in both dom0 and domU. Does it mean that we can
collect both of them at once if we run perf command on the dom0? (If
not, does it mean we can collect counters for each domain separately
once we run the perf command in each domain?
Profiling both guest and dom0 (and the hypervisor) requires changes to
perf and those are not there yet.
But you can run perf in each guest (including dom0) separately. Make
sure you have vpmu=true boot option.
-boris
|
