Re: [libvirt RFC] virFile: new VIR_FILE_WRAPPER_BIG_PIPE to improve performance

[Claudio Fontana]

On 4/5/22 10:35 AM, Dr. David Alan Gilbert wrote:
> * Claudio Fontana (cfontana@suse.de) wrote:
>> On 3/28/22 10:31 AM, Daniel P. Berrangé wrote:
>>> On Sat, Mar 26, 2022 at 04:49:46PM +0100, Claudio Fontana wrote:
>>>> On 3/25/22 12:29 PM, Daniel P. Berrangé wrote:
>>>>> On Fri, Mar 18, 2022 at 02:34:29PM +0100, Claudio Fontana wrote:
>>>>>> On 3/17/22 4:03 PM, Dr. David Alan Gilbert wrote:
>>>>>>> * Claudio Fontana (cfontana@suse.de) wrote:
>>>>>>>> On 3/17/22 2:41 PM, Claudio Fontana wrote:
>>>>>>>>> On 3/17/22 11:25 AM, Daniel P. Berrangé wrote:
>>>>>>>>>> On Thu, Mar 17, 2022 at 11:12:11AM +0100, Claudio Fontana wrote:
>>>>>>>>>>> On 3/16/22 1:17 PM, Claudio Fontana wrote:
>>>>>>>>>>>> On 3/14/22 6:48 PM, Daniel P. Berrangé wrote:
>>>>>>>>>>>>> On Mon, Mar 14, 2022 at 06:38:31PM +0100, Claudio Fontana wrote:
>>>>>>>>>>>>>> On 3/14/22 6:17 PM, Daniel P. Berrangé wrote:
>>>>>>>>>>>>>>> On Sat, Mar 12, 2022 at 05:30:01PM +0100, Claudio Fontana wrote:
>>>>>>>>>>>>>>>> the first user is the qemu driver,
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> virsh save/resume would slow to a crawl with a default pipe 
>>>>>>>>>>>>>>>> size (64k).
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> This improves the situation by 400%.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Going through io_helper still seems to incur in some penalty 
>>>>>>>>>>>>>>>> (~15%-ish)
>>>>>>>>>>>>>>>> compared with direct qemu migration to a nc socket to a file.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Signed-off-by: Claudio Fontana <cfontana@suse.de>
>>>>>>>>>>>>>>>> ---
>>>>>>>>>>>>>>>>  src/qemu/qemu_driver.c    |  6 +++---
>>>>>>>>>>>>>>>>  src/qemu/qemu_saveimage.c | 11 ++++++-----
>>>>>>>>>>>>>>>>  src/util/virfile.c        | 12 ++++++++++++
>>>>>>>>>>>>>>>>  src/util/virfile.h        |  1 +
>>>>>>>>>>>>>>>>  4 files changed, 22 insertions(+), 8 deletions(-)
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Hello, I initially thought this to be a qemu performance issue,
>>>>>>>>>>>>>>>> so you can find the discussion about this in qemu-devel:
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> "Re: bad virsh save /dev/null performance (600 MiB/s max)"
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> https://lists.gnu.org/archive/html/qemu-devel/2022-03/msg03142.html
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>> Current results show these experimental averages maximum throughput
>>>>>>>>>>> migrating to /dev/null per each FdWrapper Pipe Size (as per QEMU QMP
>>>>>>>>>>> "query-migrate", tests repeated 5 times for each).
>>>>>>>>>>> VM Size is 60G, most of the memory effectively touched before 
>>>>>>>>>>> migration,
>>>>>>>>>>> through user application allocating and touching all memory with
>>>>>>>>>>> pseudorandom data.
>>>>>>>>>>>
>>>>>>>>>>> 64K:     5200 Mbps (current situation)
>>>>>>>>>>> 128K:    5800 Mbps
>>>>>>>>>>> 256K:   20900 Mbps
>>>>>>>>>>> 512K:   21600 Mbps
>>>>>>>>>>> 1M:     22800 Mbps
>>>>>>>>>>> 2M:     22800 Mbps
>>>>>>>>>>> 4M:     22400 Mbps
>>>>>>>>>>> 8M:     22500 Mbps
>>>>>>>>>>> 16M:    22800 Mbps
>>>>>>>>>>> 32M:    22900 Mbps
>>>>>>>>>>> 64M:    22900 Mbps
>>>>>>>>>>> 128M:   22800 Mbps
>>>>>>>>>>>
>>>>>>>>>>> This above is the throughput out of patched libvirt with multiple 
>>>>>>>>>>> Pipe Sizes for the FDWrapper.
>>>>>>>>>>
>>>>>>>>>> Ok, its bouncing around with noise after 1 MB. So I'd suggest that
>>>>>>>>>> libvirt attempt to raise the pipe limit to 1 MB by default, but
>>>>>>>>>> not try to go higher.
>>>>>>>>>>
>>>>>>>>>>> As for the theoretical limit for the libvirt architecture,
>>>>>>>>>>> I ran a qemu migration directly issuing the appropriate QMP
>>>>>>>>>>> commands, setting the same migration parameters as per libvirt,
>>>>>>>>>>> and then migrating to a socket netcatted to /dev/null via
>>>>>>>>>>> {"execute": "migrate", "arguments": { "uri", 
>>>>>>>>>>> "unix:///tmp/netcat.sock" } } :
>>>>>>>>>>>
>>>>>>>>>>> QMP:    37000 Mbps
>>>>>>>>>>
>>>>>>>>>>> So although the Pipe size improves things (in particular the
>>>>>>>>>>> large jump is for the 256K size, although 1M seems a very good 
>>>>>>>>>>> value),
>>>>>>>>>>> there is still a second bottleneck in there somewhere that
>>>>>>>>>>> accounts for a loss of ~14200 Mbps in throughput.
>>>>>>>>
>>>>>>>>
>>>>>>>> Interesting addition: I tested quickly on a system with faster cpus 
>>>>>>>> and larger VM sizes, up to 200GB,
>>>>>>>> and the difference in throughput libvirt vs qemu is basically the same 
>>>>>>>> ~14500 Mbps.
>>>>>>>>
>>>>>>>> ~50000 mbps qemu to netcat socket to /dev/null
>>>>>>>> ~35500 mbps virsh save to /dev/null
>>>>>>>>
>>>>>>>> Seems it is not proportional to cpu speed by the looks of it (not a 
>>>>>>>> totally fair comparison because the VM sizes are different).
>>>>>>>
>>>>>>> It might be closer to RAM or cache bandwidth limited though; for an 
>>>>>>> extra copy.
>>>>>>
>>>>>> I was thinking about sendfile(2) in iohelper, but that probably
>>>>>> can't work as the input fd is a socket, I am getting EINVAL.
>>>>>
>>>>> Yep, sendfile() requires the input to be a mmapable FD,
>>>>> and the output to be a socket.
>>>>>
>>>>> Try splice() instead  which merely requires 1 end to be a
>>>>> pipe, and the other end can be any FD afaik.
>>>>>
>>>>
>>>> I did try splice(), but performance is worse by around 500%.
>>>
>>> Hmm, that's certainly unexpected !
>>>
>>>> Any ideas welcome,
>>>
>>> I learnt there is also a newer  copy_file_range call, not sure if that's
>>> any better.
>>>
>>> You passed len as 1 MB, I wonder if passing MAXINT is viable ? We just
>>> want to copy everything IIRC.
>>>
>>> With regards,
>>> Daniel
>>>
>>
>> Crazy idea, would trying to use the parallel migration concept for migrating 
>> to/from a file make any sense?
>>
>> Not sure if applying the qemu multifd implementation of this would apply, 
>> maybe it could be given another implementation for "toFile", trying to use 
>> more than one cpu to do the transfer?
> 
> I can't see a way that would help; well, I could if you could
> somehow have multiple io helper threads that dealt with it.
> 
> Dave
> 

I'll spend cycles on this;

another thing I noticed while doing the "splice" API experiments in iohelper,

we cannot use splice and --bypass-cache (O_DIRECT) together there, because as 
far as I could debug/understand, apparently the source data stream is not block 
aligned,
as the kernel iomap_dio_bio_iter return value during the splice call shows, so 
that's why I get EINVAL.

If we had one of the migration streams that consisted only of pages (instead of 
header+page, ...), that might give us block alignment, potentially unlock 
better performance via splice that way?

This is all just ideas, don't have any data yet to back this up.

Thanks,

Claudio