Re: starting to look at qemu savevm performance, a first regression detected

[Claudio Fontana]

On 3/7/22 12:31 PM, Dr. David Alan Gilbert wrote:
> * Claudio Fontana (cfontana@suse.de) wrote:
>> On 3/7/22 11:32 AM, Dr. David Alan Gilbert wrote:
>>> * Claudio Fontana (cfontana@suse.de) wrote:
>>>> On 3/5/22 2:20 PM, Claudio Fontana wrote:
>>>>>
>>>>> Hello all,
>>>>>
>>>>> I have been looking at some reports of bad qemu savevm performance in 
>>>>> large VMs (around 20+ Gb),
>>>>> when used in libvirt commands like:
>>>>>
>>>>>
>>>>> virsh save domain /dev/null
>>>>>
>>>>>
>>>>>
>>>>> I have written a simple test to run in a Linux centos7-minimal-2009 
>>>>> guest, which allocates and touches 20G mem.
>>>>>
>>>>> With any qemu version since around 2020, I am not seeing more than 580 
>>>>> Mb/Sec even in the most ideal of situations.
>>>>>
>>>>> This drops to around 122 Mb/sec after commit: 
>>>>> cbde7be900d2a2279cbc4becb91d1ddd6a014def .
>>>>>
>>>>> Here is the bisection for this particular drop in throughput:
>>>>>
>>>>> commit cbde7be900d2a2279cbc4becb91d1ddd6a014def (HEAD, refs/bisect/bad)
>>>>> Author: Daniel P. Berrangé <berrange@redhat.com>
>>>>> Date:   Fri Feb 19 18:40:12 2021 +0000
>>>>>
>>>>>     migrate: remove QMP/HMP commands for speed, downtime and cache size
>>>>>     
>>>>>     The generic 'migrate_set_parameters' command handle all types of 
>>>>> param.
>>>>>     
>>>>>     Only the QMP commands were documented in the deprecations page, but 
>>>>> the
>>>>>     rationale for deprecating applies equally to HMP, and the replacements
>>>>>     exist. Furthermore the HMP commands are just shims to the QMP 
>>>>> commands,
>>>>>     so removing the latter breaks the former unless they get 
>>>>> re-implemented.
>>>>>     
>>>>>     Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
>>>>>     Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
>>>>>
>>>>>
>>>>> git bisect start
>>>>> # bad: [5c8463886d50eeb0337bd121ab877cf692731e36] Merge remote-tracking 
>>>>> branch 'remotes/kraxel/tags/kraxel-20220304-pull-request' into staging
>>>>> git bisect bad 5c8463886d50eeb0337bd121ab877cf692731e36
>>>>> # good: [6cdf8c4efa073eac7d5f9894329e2d07743c2955] Update version for 
>>>>> 4.2.1 release
>>>>> git bisect good 6cdf8c4efa073eac7d5f9894329e2d07743c2955
>>>>> # good: [b0ca999a43a22b38158a222233d3f5881648bb4f] Update version for 
>>>>> v4.2.0 release
>>>>> git bisect good b0ca999a43a22b38158a222233d3f5881648bb4f
>>>>> # skip: [e2665f314d80d7edbfe7f8275abed7e2c93c0ddc] target/mips: Alias MSA 
>>>>> vector registers on FPU scalar registers
>>>>> git bisect skip e2665f314d80d7edbfe7f8275abed7e2c93c0ddc
>>>>> # good: [4762c82cbda22b1036ce9dd2c5e951ac0ed0a7d3] tests/docker: Install 
>>>>> static libc package in CentOS 7
>>>>> git bisect good 4762c82cbda22b1036ce9dd2c5e951ac0ed0a7d3
>>>>> # bad: [d4127349e316b5c78645f95dba5922196ac4cc23] Merge remote-tracking 
>>>>> branch 'remotes/berrange-gitlab/tags/crypto-and-more-pull-request' into 
>>>>> staging
>>>>> git bisect bad d4127349e316b5c78645f95dba5922196ac4cc23
>>>>> # bad: [d90f154867ec0ec22fd719164b88716e8fd48672] Merge remote-tracking 
>>>>> branch 'remotes/dg-gitlab/tags/ppc-for-6.1-20210504' into staging
>>>>> git bisect bad d90f154867ec0ec22fd719164b88716e8fd48672
>>>>> # good: [dd5af6ece9b101d29895851a7441d848b7ccdbff] tests/docker: add a 
>>>>> test-tcg for building then running check-tcg
>>>>> git bisect good dd5af6ece9b101d29895851a7441d848b7ccdbff
>>>>> # bad: [90ec1cff768fcbe1fa2870d2018f378376f4f744] target/riscv: Adjust 
>>>>> privilege level for HLV(X)/HSV instructions
>>>>> git bisect bad 90ec1cff768fcbe1fa2870d2018f378376f4f744
>>>>> # good: [373969507a3dc7de2d291da7e1bd03acf46ec643] migration: Replaced 
>>>>> qemu_mutex_lock calls with QEMU_LOCK_GUARD
>>>>> git bisect good 373969507a3dc7de2d291da7e1bd03acf46ec643
>>>>> # good: [4083904bc9fe5da580f7ca397b1e828fbc322732] Merge remote-tracking 
>>>>> branch 'remotes/rth-gitlab/tags/pull-tcg-20210317' into staging
>>>>> git bisect good 4083904bc9fe5da580f7ca397b1e828fbc322732
>>>>> # bad: [009ff89328b1da3ea8ba316bf2be2125bc9937c5] vl: allow passing JSON 
>>>>> to -object
>>>>> git bisect bad 009ff89328b1da3ea8ba316bf2be2125bc9937c5
>>>>> # bad: [50243407457a9fb0ed17b9a9ba9fc9aee09495b1] qapi/qom: Drop 
>>>>> deprecated 'props' from object-add
>>>>> git bisect bad 50243407457a9fb0ed17b9a9ba9fc9aee09495b1
>>>>> # bad: [1b507e55f8199eaad99744613823f6929e4d57c6] Merge remote-tracking 
>>>>> branch 'remotes/berrange-gitlab/tags/dep-many-pull-request' into staging
>>>>> git bisect bad 1b507e55f8199eaad99744613823f6929e4d57c6
>>>>> # bad: [24e13a4dc1eb1630eceffc7ab334145d902e763d] chardev: reject use of 
>>>>> 'wait' flag for socket client chardevs
>>>>> git bisect bad 24e13a4dc1eb1630eceffc7ab334145d902e763d
>>>>> # good: [8becb36063fb14df1e3ae4916215667e2cb65fa2] monitor: remove 
>>>>> 'query-events' QMP command
>>>>> git bisect good 8becb36063fb14df1e3ae4916215667e2cb65fa2
>>>>> # bad: [8af54b9172ff3b9bbdbb3191ed84994d275a0d81] machine: remove 
>>>>> 'query-cpus' QMP command
>>>>> git bisect bad 8af54b9172ff3b9bbdbb3191ed84994d275a0d81
>>>>> # bad: [cbde7be900d2a2279cbc4becb91d1ddd6a014def] migrate: remove QMP/HMP 
>>>>> commands for speed, downtime and cache size
>>>>> git bisect bad cbde7be900d2a2279cbc4becb91d1ddd6a014def
>>>>> # first bad commit: [cbde7be900d2a2279cbc4becb91d1ddd6a014def] migrate: 
>>>>> remove QMP/HMP commands for speed, downtime and cache size
>>>>>
>>>>>
>>>>> Are there some obvious settings / options I am missing to regain the 
>>>>> savevm performance after this commit?
>>>>
>>>> Answering myself: 
>>>
>>> <oops we seem to have split this thread into two>
>>>
>>>> this seems to be due to a resulting different default xbzrle cache size 
>>>> (probably interactions between libvirt/qemu versions?).
>>>>
>>>> When forcing the xbzrle cache size to a larger value, the performance is 
>>>> back.
>>>
>>> That's weird that 'virsh save' is ending up using xbzrle.
>>
>> virsh save (or qemu savevm..) seems to me like it uses a subset of the 
>> migration code and migration parameters but not all..
>>
>>>
>>>>>
>>>>> I have seen projects attempting to improve other aspects of performance 
>>>>> (snapshot performance, etc), is there something going on to improve the 
>>>>> transfer of RAM in savevm too?
>>>>
>>>>
>>>> Still I would think that we should be able to do better than 600ish Mb/s , 
>>>> any ideas, prior work on this,
>>>> to improve savevm performance, especially looking at RAM regions transfer 
>>>> speed?
>>>
>>> My normal feeling is ~10Gbps for a live migrate over the wire; I rarely
>>> try virsh save though.
>>> If you're using xbzrle that might explain it; it's known to eat cpu -
>>> but I'd never expect it to have been used with 'virsh save'.
>>
>> some valgrind shows it among the top cpu eaters;

well.. I was confused.

The usage of xbzrle is just on constantly calling migrate_use_xbzrle() and 
XBZRLE_cache_lock and XBZRLE_cache_unlock() as well as some 
xbzrle_cache_zero_page(),
which likely do not do anything useful, as ->ram_bulk_stage is not changed by 
anything so it should be true.


>>
>> I wonder why we are able to do more than 2x better for actual live 
>> migration, compared with virsh save /dev/null ...
> 
> What speed do you get if you force xbzrle off?


no substantial difference.


> 
> Dave
> 
>> Thanks,
>>
>> Claudio
>>