On Thu, Dec 15, 2016 at 3:04 AM, Stefan Hajnoczi <stefa...@gmail.com> wrote:
> On Wed, Dec 14, 2016 at 08:06:10PM -0500, Weiwei Jia wrote:
>> Hi Stefan,
>>
>> Thanks for your reply. Please see the inline replies.
>>
>> On Wed, Dec 14, 2016 at 2:31 PM, Stefan Hajnoczi <stefa...@gmail.com> wrote:
>> > On Wed, Dec 14, 2016 at 12:58:11AM -0500, Weiwei Jia wrote:
>> >> I find the timeslice of vCPU thread in QEMU/KVM is unstable when there
>> >> are lots of read requests (for example, read 4KB each time (8GB in
>> >> total) from one file) from Guest OS. I also find that this phenomenon
>> >> may be caused by lock contention in QEMU layer. I find this problem
>> >> under following workload.
>> >>
>> >> Workload settings:
>> >> In VMM, there are 6 pCPUs which are pCPU0, pCPU1, pCPU2, pCPU3, pCPU4,
>> >> pCPU5. There are two Kernel Virtual Machines (VM1 and VM2) upon VMM.
>> >> In each VM, there are 5 vritual CPUs (vCPU0, vCPU1, vCPU2, vCPU3,
>> >> vCPU4). vCPU0 in VM1 and vCPU0 in VM2 are pinned to pCPU0 and pCPU5
>> >> separately to handle interrupts dedicatedly. vCPU1 in VM1 and vCPU1 in
>> >> VM2 are pinned to pCPU1; vCPU2 in VM1 and vCPU2 in VM2 are pinned to
>> >> pCPU2; vCPU3 in VM1 and vCPU3 in VM2 are pinned to pCPU3; vCPU4 in VM1
>> >> and vCPU4 in VM2 are pinned to pCPU4. Besides vCPU0 in VM2 (pinned to
>> >> pCPU5), other vCPUs all have one CPU intensive thread (while(1){i++})
>> >> upon each of them in VM1 and VM2 to avoid the vCPU to be idle. In VM1,
>> >> I start one I/O thread on vCPU2, which the I/O thread reads 4KB from
>> >> one file each time (reads 8GB in total). The I/O scheduler in VM1 and
>> >> VM2 is NOOP. The I/O scheduler in VMM is CFQ. I also pinned the I/O
>> >> worker threads launched by QEMU to pCPU5 (note: there is no CPU
>> >> intensive thread on pCPU5 so the I/O requests will be handled by QEMU
>> >> I/O thread workers ASAP). The process scheduling class in VM and VMM
>> >> is CFS.
>> >
>> > Did you pin the QEMU main loop to pCPU5? This is the QEMU process' main
>> > thread and it handles ioeventfd (virtqueue kick) and thread pool
>> > completions.
>>
>> No, I did not pin main loop to pCPU5. Do you mean If I pin QEMU main
>> loop to pCPU5 under above workload, the timeslice of vCPU2 thread will
>> be stable even though there are lots of I/O requests? I didn't use
>> virtio for VM and I use SCSI. My whole VM xml configuration file is as
>> follows.
>
> Pinning the main loop will probably not solve the problem but it might
> help a bit. I just noticed it while reading your email because you
> pinned everything carefully except the main loop, which is an important
> thread.
Yes, even pin main loop to dedicated pCPU, the timeslice still jitters
very much.
>> >>
>> >> Linux Kernel version for VMM is: 3.16.39
>> >> Linux Kernel version for VM1 and VM2 is: 4.7.4
>> >> QEMU emulator version is: 2.0.0
>> >>
>> >> When I test above workload, I find the timeslice of vCPU2 thread
>> >> jitters very much. I suspect this is triggered by lock contention in
>> >> QEMU layer since my debug log in front of VMM Linux Kernel's
>> >> schedule->__schedule->context_switch is like following. Once the
>> >> timeslice jitters very much, following debug information will appear.
>> >>
>> >> 7097537 Dec 13 11:22:33 mobius04 kernel: [39163.015789] Call Trace:
>> >> 7097538 Dec 13 11:22:33 mobius04 kernel: [39163.015791]
>> >> [<ffffffff8176b2f0>] dump_stack+0x64/0x84
>> >> 7097539 Dec 13 11:22:33 mobius04 kernel: [39163.015793]
>> >> [<ffffffff8176bf85>] __schedule+0x5b5/0x960
>> >> 7097540 Dec 13 11:22:33 mobius04 kernel: [39163.015794]
>> >> [<ffffffff8176c409>] schedule+0x29/0x70
>> >> 7097541 Dec 13 11:22:33 mobius04 kernel: [39163.015796]
>> >> [<ffffffff810ef4d8>] futex_wait_queue_me+0xd8/0x150
>> >> 7097542 Dec 13 11:22:33 mobius04 kernel: [39163.015798]
>> >> [<ffffffff810ef6fb>] futex_wait+0x1ab/0x2b0
>> >> 7097543 Dec 13 11:22:33 mobius04 kernel: [39163.015800]
>> >> [<ffffffff810eef00>] ? get_futex_key+0x2d0/0x2e0
>> >> 7097544 Dec 13 11:22:33 mobius04 kernel: [39163.015804]
>> >> [<ffffffffc0290105>] ? __vmx_load_host_state+0x125/0x170 [kv
>> >> m_intel]
>> >> 7097545 Dec 13 11:22:33 mobius04 kernel: [39163.015805]
>> >> [<ffffffff810f1275>] do_futex+0xf5/0xd20
>> >> 7097546 Dec 13 11:22:33 mobius04 kernel: [39163.015813]
>> >> [<ffffffffc0222690>] ? kvm_vcpu_ioctl+0x100/0x560 [kvm]
>> >> 7097547 Dec 13 11:22:33 mobius04 kernel: [39163.015816]
>> >> [<ffffffff810b06f0>] ? __dequeue_entity+0x30/0x50
>> >> 7097548 Dec 13 11:22:33 mobius04 kernel: [39163.015818]
>> >> [<ffffffff81013d06>] ? __switch_to+0x596/0x690
>> >> 7097549 Dec 13 11:22:33 mobius04 kernel: [39163.015820]
>> >> [<ffffffff811f9f23>] ? do_vfs_ioctl+0x93/0x520
>> >> 7097550 Dec 13 11:22:33 mobius04 kernel: [39163.015822]
>> >> [<ffffffff810f1f1d>] SyS_futex+0x7d/0x170
>> >> 7097551 Dec 13 11:22:33 mobius04 kernel: [39163.015824]
>> >> [<ffffffff8116d1b2>] ? fire_user_return_notifiers+0x42/0x50
>> >> 7097552 Dec 13 11:22:33 mobius04 kernel: [39163.015826]
>> >> [<ffffffff810154b5>] ? do_notify_resume+0xc5/0x100
>> >> 7097553 Dec 13 11:22:33 mobius04 kernel: [39163.015828]
>> >> [<ffffffff81770a8d>] system_call_fastpath+0x1a/0x1f
>> >>
>> >>
>> >> If true, I think this may be a scalability problem caused by QEMU I/O
>> >> part. Do we have a feature in QEMU to avoid this? Would you please
>> >> give me some suggestions about how to make the timeslice of vCPU2
>> >> thread stable even though there are lots of I/O Read requests on it.
>> >
>> > Yes, there is a way to reduce jitter caused by the QEMU global mutex:
>> >
>> > qemu -object iothread,id=iothread0 \
>> > -drive if=none,id=drive0,file=test.img,format=raw,cache=none \
>> > -device virtio-blk-pci,iothread=iothread0,drive=drive0
>> >
>> > Now the ioeventfd and thread pool completions will be processed in
>> > iothread0 instead of the QEMU main loop thread. This thread does not
>> > take the QEMU global mutex so vcpu execution is not hindered.
>> >
>> > This feature is called virtio-blk dataplane.
>> >
>> > You can query IOThread thread IDs using the query-iothreads QMP command.
>> > This will allow you to pin iothread0 to pCPU5.
>> >
>> > Please let us know if this helps.
>>
>> Does this feature only work for VirtIO? Does it work for SCSI or IDE?
>
> This only works for virtio-blk and virtio-scsi. The virtio-scsi
> dataplane support is more recent and I don't remember if it is complete.
> I've CCed Fam and Paolo who worked on virtio-scsi dataplane.
>
> Now that you have mentioned that you aren't using virtio devices, there
> is another source of lock contention that you will encounter. I/O
> request submission takes place in the vcpu thread when ioeventfd is not
> used. Only virtio uses ioeventfd so your current QEMU configuration is
> unable to let the vcpu continue execution during I/O request submission.
>
> If you care about performance then using virtio devices is probably the
> best choice. Try comparing against virtio-scsi dataplane - you should
> see a lot less jitter.
I will try virtio-scsi or virtio-blk dataplane solution. Thank you.
Cheers,
Weiwei Jia
