[PATCH v2 0/6] migration: Multi-thread compression method support

Zeyu Jin posted 6 patches 3 years, 5 months ago
Only 0 patches received!
docs/multi-thread-compression.txt |  31 ++-
hw/core/qdev-properties-system.c  |  11 +
include/hw/qdev-properties.h      |   4 +
migration/migration.c             |  56 ++++-
migration/migration.h             |   1 +
migration/qemu-file.c             |  62 +----
migration/qemu-file.h             |   4 +-
migration/ram.c                   | 381 +++++++++++++++++++++++++-----
monitor/hmp-cmds.c                |  12 +
qapi/migration.json               |  26 +-
10 files changed, 465 insertions(+), 123 deletions(-)
[PATCH v2 0/6] migration: Multi-thread compression method support
Posted by Zeyu Jin 3 years, 5 months ago
Currently we have both multi-thread compression and multifd to optimize
live migration in Qemu. Mulit-thread compression deals with the situation
where network bandwith is limited but cpu resource adequate. Multifd instead
aims to take full advantage of network bandwith. Moreover it supports both
zlib and zstd compression on each channel.

In this patch series, we did some code refactoring on multi-thread compression
live migration and bring zstd compression method support for it.

Below is the test result of multi-thread compression live migration
with different compress methods. Test result shows that zstd outperforms
zlib by about 70%.

 Migration Configuration:
 Guest 8U 32G
 compress-threads   8
 decompress-threads 2
 compress-level 1
 bandwidth-limit 100Mbps

 Test Result:
 +---------------------+--------------+-------------+
 |  compress method    |   zlib       |    zstd     |
 +---------------------+--------------+-------------+
 |  total time (ms)    |   75256      |    44187    |
 +---------------------+--------------+-------------+
 |  downtime(ms)       |   128        |    81       |
 +---------------------+--------------+-------------+
 |  transferred ram(kB)|   1576866    |    736117   |
 +---------------------+--------------+-------------+
 |  throughput(mbps)   |   172.06     |    137.16   |
 +---------------------+--------------+-------------+
 |  total ram(kB)      |   33685952   |    33685952 |
 +---------------------+--------------+-------------+

Zeyu Jin (6):
  migration: Add multi-thread compress method
  migration: Refactoring multi-thread compress migration
  migration: Add multi-thread compress ops
  migration: Add zstd support in multi-thread compression
  migration: Add compress_level sanity check
  doc: Update multi-thread compression doc

 docs/multi-thread-compression.txt |  31 ++-
 hw/core/qdev-properties-system.c  |  11 +
 include/hw/qdev-properties.h      |   4 +
 migration/migration.c             |  56 ++++-
 migration/migration.h             |   1 +
 migration/qemu-file.c             |  62 +----
 migration/qemu-file.h             |   4 +-
 migration/ram.c                   | 381 +++++++++++++++++++++++++-----
 monitor/hmp-cmds.c                |  12 +
 qapi/migration.json               |  26 +-
 10 files changed, 465 insertions(+), 123 deletions(-)

-- 
2.27.0


Re: [PATCH v2 0/6] migration: Multi-thread compression method support
Posted by Eric Blake 3 years, 4 months ago
On 11/27/20 3:36 AM, Zeyu Jin wrote:

Meta-comment: you appear to be having problems threading your series;
I've now seen three separate cover letters (RFC v1, v2 with no subject,
v2 with subject) and two series where each patch was a separate thread.
It is difficult to follow which messages are related when reading in a
mail client that sorts by most-recently-active thread first.  You may
want to investigate why your threading is not working, although I'd wait
to send v3 until you have actual changes to incorporate.

> Currently we have both multi-thread compression and multifd to optimize
> live migration in Qemu. Mulit-thread compression deals with the situation
> where network bandwith is limited but cpu resource adequate. Multifd instead

Not that typos in the cover letter matter, but this should be 'bandwidth'

> aims to take full advantage of network bandwith. Moreover it supports both
> zlib and zstd compression on each channel.
> 
> In this patch series, we did some code refactoring on multi-thread compression
> live migration and bring zstd compression method support for it.
> 
> Below is the test result of multi-thread compression live migration
> with different compress methods. Test result shows that zstd outperforms
> zlib by about 70%.
> 

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org


Re: [PATCH v2 0/6] migration: Multi-thread compression method support
Posted by Zeyu Jin 3 years, 4 months ago
On 2020/12/1 0:42, Eric Blake wrote:
> On 11/27/20 3:36 AM, Zeyu Jin wrote:
> 
> Meta-comment: you appear to be having problems threading your series;
> I've now seen three separate cover letters (RFC v1, v2 with no subject,
> v2 with subject) and two series where each patch was a separate thread.
> It is difficult to follow which messages are related when reading in a
> mail client that sorts by most-recently-active thread first.  You may
> want to investigate why your threading is not working, although I'd wait
> to send v3 until you have actual changes to incorporate.
> 

Thank you for noticing that. It`s my mistake when sending patches.
Everything will be fine in v3.

>> Currently we have both multi-thread compression and multifd to optimize
>> live migration in Qemu. Mulit-thread compression deals with the situation
>> where network bandwith is limited but cpu resource adequate. Multifd instead
> 
> Not that typos in the cover letter matter, but this should be 'bandwidth'
>

Yes, I will fix that.

>> aims to take full advantage of network bandwith. Moreover it supports both
>> zlib and zstd compression on each channel.
>>
>> In this patch series, we did some code refactoring on multi-thread compression
>> live migration and bring zstd compression method support for it.
>>
>> Below is the test result of multi-thread compression live migration
>> with different compress methods. Test result shows that zstd outperforms
>> zlib by about 70%.
>>
>