mbox series

[v3,0/4] Implement using Intel QAT to offload ZLIB

Message ID 20240627223445.95096-1-yichen.wang@bytedance.com
Headers show
Series Implement using Intel QAT to offload ZLIB | expand

Message

Yichen Wang June 27, 2024, 10:34 p.m. UTC
v3:
- Rebase changes on top of master
- Merge two patches per Fabiano Rosas's comment
- Add versions into comments and documentations

v2:
- Rebase changes on top of recent multifd code changes.
- Use QATzip API 'qzMalloc' and 'qzFree' to allocate QAT buffers.
- Remove parameter tuning and use QATzip's defaults for better
  performance.
- Add parameter to enable QAT software fallback.

v1:
https://lists.nongnu.org/archive/html/qemu-devel/2023-12/msg03761.html

* Performance

We present updated performance results. For circumstantial reasons, v1
presented performance on a low-bandwidth (1Gbps) network.

Here, we present updated results with a similar setup as before but with
two main differences:

1. Our machines have a ~50Gbps connection, tested using 'iperf3'.
2. We had a bug in our memory allocation causing us to only use ~1/2 of
the VM's RAM. Now we properly allocate and fill nearly all of the VM's
RAM.

Thus, the test setup is as follows:

We perform multifd live migration over TCP using a VM with 64GB memory.
We prepare the machine's memory by powering it on, allocating a large
amount of memory (60GB) as a single buffer, and filling the buffer with
the repeated contents of the Silesia corpus[0]. This is in lieu of a more
realistic memory snapshot, which proved troublesome to acquire.

We analyze CPU usage by averaging the output of 'top' every second
during migration. This is admittedly imprecise, but we feel that it
accurately portrays the different degrees of CPU usage of varying
compression methods.

We present the latency, throughput, and CPU usage results for all of the
compression methods, with varying numbers of multifd threads (4, 8, and
16).

[0] The Silesia corpus can be accessed here:
https://sun.aei.polsl.pl//~sdeor/index.php?page=silesia

** Results

4 multifd threads:

    |---------------|---------------|----------------|---------|---------|
    |method         |time(sec)      |throughput(mbps)|send cpu%|recv cpu%|
    |---------------|---------------|----------------|---------|---------|
    |qatzip         | 23.13         | 8749.94        |117.50   |186.49   |
    |---------------|---------------|----------------|---------|---------|
    |zlib           |254.35         |  771.87        |388.20   |144.40   |
    |---------------|---------------|----------------|---------|---------|
    |zstd           | 54.52         | 3442.59        |414.59   |149.77   |
    |---------------|---------------|----------------|---------|---------|
    |none           | 12.45         |43739.60        |159.71   |204.96   |
    |---------------|---------------|----------------|---------|---------|

8 multifd threads:

    |---------------|---------------|----------------|---------|---------|
    |method         |time(sec)      |throughput(mbps)|send cpu%|recv cpu%|
    |---------------|---------------|----------------|---------|---------|
    |qatzip         | 16.91         |12306.52        |186.37   |391.84   |
    |---------------|---------------|----------------|---------|---------|
    |zlib           |130.11         | 1508.89        |753.86   |289.35   |
    |---------------|---------------|----------------|---------|---------|
    |zstd           | 27.57         | 6823.23        |786.83   |303.80   |
    |---------------|---------------|----------------|---------|---------|
    |none           | 11.82         |46072.63        |163.74   |238.56   |
    |---------------|---------------|----------------|---------|---------|

16 multifd threads:

    |---------------|---------------|----------------|---------|---------|
    |method         |time(sec)      |throughput(mbps)|send cpu%|recv cpu%|
    |---------------|---------------|----------------|---------|---------|
    |qatzip         |18.64          |11044.52        | 573.61  |437.65   |
    |---------------|---------------|----------------|---------|---------|
    |zlib           |66.43          | 2955.79        |1469.68  |567.47   |
    |---------------|---------------|----------------|---------|---------|
    |zstd           |14.17          |13290.66        |1504.08  |615.33   |
    |---------------|---------------|----------------|---------|---------|
    |none           |16.82          |32363.26        | 180.74  |217.17   |
    |---------------|---------------|----------------|---------|---------|

** Observations

- In general, not using compression outperforms using compression in a
  non-network-bound environment.
- 'qatzip' outperforms other compression workers with 4 and 8 workers,
  achieving a ~91% latency reduction over 'zlib' with 4 workers, and a
~58% latency reduction over 'zstd' with 4 workers.
- 'qatzip' maintains comparable performance with 'zstd' at 16 workers,
  showing a ~32% increase in latency. This performance difference
becomes more noticeable with more workers, as CPU compression is highly
parallelizable.
- 'qatzip' compression uses considerably less CPU than other compression
  methods. At 8 workers, 'qatzip' demonstrates a ~75% reduction in
compression CPU usage compared to 'zstd' and 'zlib'.
- 'qatzip' decompression CPU usage is less impressive, and is even
  slightly worse than 'zstd' and 'zlib' CPU usage at 4 and 16 workers.


Bryan Zhang (4):
  meson: Introduce 'qatzip' feature to the build system
  migration: Add migration parameters for QATzip
  migration: Introduce 'qatzip' compression method
  tests/migration: Add integration test for 'qatzip' compression method

 hw/core/qdev-properties-system.c |   6 +-
 meson.build                      |  10 +
 meson_options.txt                |   2 +
 migration/meson.build            |   1 +
 migration/migration-hmp-cmds.c   |   8 +
 migration/multifd-qatzip.c       | 382 +++++++++++++++++++++++++++++++
 migration/multifd.h              |   1 +
 migration/options.c              |  57 +++++
 migration/options.h              |   2 +
 qapi/migration.json              |  38 +++
 scripts/meson-buildoptions.sh    |   6 +
 tests/qtest/meson.build          |   4 +
 tests/qtest/migration-test.c     |  35 +++
 13 files changed, 551 insertions(+), 1 deletion(-)
 create mode 100644 migration/multifd-qatzip.c

Comments

Peter Xu July 2, 2024, 7:16 p.m. UTC | #1
On Thu, Jun 27, 2024 at 03:34:41PM -0700, Yichen Wang wrote:
> v3:
> - Rebase changes on top of master
> - Merge two patches per Fabiano Rosas's comment
> - Add versions into comments and documentations
> 
> v2:
> - Rebase changes on top of recent multifd code changes.
> - Use QATzip API 'qzMalloc' and 'qzFree' to allocate QAT buffers.
> - Remove parameter tuning and use QATzip's defaults for better
>   performance.
> - Add parameter to enable QAT software fallback.
> 
> v1:
> https://lists.nongnu.org/archive/html/qemu-devel/2023-12/msg03761.html
> 
> * Performance
> 
> We present updated performance results. For circumstantial reasons, v1
> presented performance on a low-bandwidth (1Gbps) network.
> 
> Here, we present updated results with a similar setup as before but with
> two main differences:
> 
> 1. Our machines have a ~50Gbps connection, tested using 'iperf3'.
> 2. We had a bug in our memory allocation causing us to only use ~1/2 of
> the VM's RAM. Now we properly allocate and fill nearly all of the VM's
> RAM.
> 
> Thus, the test setup is as follows:
> 
> We perform multifd live migration over TCP using a VM with 64GB memory.
> We prepare the machine's memory by powering it on, allocating a large
> amount of memory (60GB) as a single buffer, and filling the buffer with
> the repeated contents of the Silesia corpus[0]. This is in lieu of a more
> realistic memory snapshot, which proved troublesome to acquire.
> 
> We analyze CPU usage by averaging the output of 'top' every second
> during migration. This is admittedly imprecise, but we feel that it
> accurately portrays the different degrees of CPU usage of varying
> compression methods.
> 
> We present the latency, throughput, and CPU usage results for all of the
> compression methods, with varying numbers of multifd threads (4, 8, and
> 16).
> 
> [0] The Silesia corpus can be accessed here:
> https://sun.aei.polsl.pl//~sdeor/index.php?page=silesia
> 
> ** Results
> 
> 4 multifd threads:
> 
>     |---------------|---------------|----------------|---------|---------|
>     |method         |time(sec)      |throughput(mbps)|send cpu%|recv cpu%|
>     |---------------|---------------|----------------|---------|---------|
>     |qatzip         | 23.13         | 8749.94        |117.50   |186.49   |
>     |---------------|---------------|----------------|---------|---------|
>     |zlib           |254.35         |  771.87        |388.20   |144.40   |
>     |---------------|---------------|----------------|---------|---------|
>     |zstd           | 54.52         | 3442.59        |414.59   |149.77   |
>     |---------------|---------------|----------------|---------|---------|
>     |none           | 12.45         |43739.60        |159.71   |204.96   |
>     |---------------|---------------|----------------|---------|---------|
> 
> 8 multifd threads:
> 
>     |---------------|---------------|----------------|---------|---------|
>     |method         |time(sec)      |throughput(mbps)|send cpu%|recv cpu%|
>     |---------------|---------------|----------------|---------|---------|
>     |qatzip         | 16.91         |12306.52        |186.37   |391.84   |
>     |---------------|---------------|----------------|---------|---------|
>     |zlib           |130.11         | 1508.89        |753.86   |289.35   |
>     |---------------|---------------|----------------|---------|---------|
>     |zstd           | 27.57         | 6823.23        |786.83   |303.80   |
>     |---------------|---------------|----------------|---------|---------|
>     |none           | 11.82         |46072.63        |163.74   |238.56   |
>     |---------------|---------------|----------------|---------|---------|
> 
> 16 multifd threads:
> 
>     |---------------|---------------|----------------|---------|---------|
>     |method         |time(sec)      |throughput(mbps)|send cpu%|recv cpu%|
>     |---------------|---------------|----------------|---------|---------|
>     |qatzip         |18.64          |11044.52        | 573.61  |437.65   |
>     |---------------|---------------|----------------|---------|---------|
>     |zlib           |66.43          | 2955.79        |1469.68  |567.47   |
>     |---------------|---------------|----------------|---------|---------|
>     |zstd           |14.17          |13290.66        |1504.08  |615.33   |
>     |---------------|---------------|----------------|---------|---------|
>     |none           |16.82          |32363.26        | 180.74  |217.17   |
>     |---------------|---------------|----------------|---------|---------|
> 
> ** Observations
> 
> - In general, not using compression outperforms using compression in a
>   non-network-bound environment.
> - 'qatzip' outperforms other compression workers with 4 and 8 workers,
>   achieving a ~91% latency reduction over 'zlib' with 4 workers, and a
> ~58% latency reduction over 'zstd' with 4 workers.
> - 'qatzip' maintains comparable performance with 'zstd' at 16 workers,
>   showing a ~32% increase in latency. This performance difference
> becomes more noticeable with more workers, as CPU compression is highly
> parallelizable.
> - 'qatzip' compression uses considerably less CPU than other compression
>   methods. At 8 workers, 'qatzip' demonstrates a ~75% reduction in
> compression CPU usage compared to 'zstd' and 'zlib'.
> - 'qatzip' decompression CPU usage is less impressive, and is even
>   slightly worse than 'zstd' and 'zlib' CPU usage at 4 and 16 workers.

Thanks for the results update.

It looks like the docs/migration/ file is still missing.  It'll be great to
have it in the next version or separately.

So how it compares with QPL (which got merged already)?  They at least look
like both supported on an Intel platform, so an user whoever wants to
compress the RAM could start to look at both.  I'm utterly confused on why
Intel provides these two similar compressors.  It would be great to have
some answer and perhaps put into the doc.

I am honestly curious too on whether are you planning to use it in
production.  It looks like if the network resources are rich, no-comp is
mostly always better than qatzip, no matter on total migration time or cpu
consumption.  I'm pretty surprised that it'll take that much resources even
if the work should have been offloaded to the QAT chips iiuc.

I think it may not be a problem to merge this series even if it performs
slower at some criterias.. but I think we may still want to know when this
should be used, or the good reason this should be merged (if it's not about
it outperforms others).

Thanks,

> 
> 
> Bryan Zhang (4):
>   meson: Introduce 'qatzip' feature to the build system
>   migration: Add migration parameters for QATzip
>   migration: Introduce 'qatzip' compression method
>   tests/migration: Add integration test for 'qatzip' compression method
> 
>  hw/core/qdev-properties-system.c |   6 +-
>  meson.build                      |  10 +
>  meson_options.txt                |   2 +
>  migration/meson.build            |   1 +
>  migration/migration-hmp-cmds.c   |   8 +
>  migration/multifd-qatzip.c       | 382 +++++++++++++++++++++++++++++++
>  migration/multifd.h              |   1 +
>  migration/options.c              |  57 +++++
>  migration/options.h              |   2 +
>  qapi/migration.json              |  38 +++
>  scripts/meson-buildoptions.sh    |   6 +
>  tests/qtest/meson.build          |   4 +
>  tests/qtest/migration-test.c     |  35 +++
>  13 files changed, 551 insertions(+), 1 deletion(-)
>  create mode 100644 migration/multifd-qatzip.c
> 
> -- 
> Yichen Wang
>
Yuan Liu July 4, 2024, 3:15 a.m. UTC | #2
> -----Original Message-----
> From: Peter Xu <peterx@redhat.com>
> Sent: Wednesday, July 3, 2024 3:16 AM
> To: Wang, Yichen <yichen.wang@bytedance.com>
> Cc: Paolo Bonzini <pbonzini@redhat.com>; Daniel P. Berrangé
> <berrange@redhat.com>; Eduardo Habkost <eduardo@habkost.net>; Marc-André
> Lureau <marcandre.lureau@redhat.com>; Thomas Huth <thuth@redhat.com>;
> Philippe Mathieu-Daudé <philmd@linaro.org>; Fabiano Rosas
> <farosas@suse.de>; Eric Blake <eblake@redhat.com>; Markus Armbruster
> <armbru@redhat.com>; Laurent Vivier <lvivier@redhat.com>; qemu-
> devel@nongnu.org; Hao Xiang <hao.xiang@linux.dev>; Liu, Yuan1
> <yuan1.liu@intel.com>; Zou, Nanhai <nanhai.zou@intel.com>; Ho-Ren (Jack)
> Chuang <horenchuang@bytedance.com>
> Subject: Re: [PATCH v3 0/4] Implement using Intel QAT to offload ZLIB
> 
> On Thu, Jun 27, 2024 at 03:34:41PM -0700, Yichen Wang wrote:
> > v3:
> > - Rebase changes on top of master
> > - Merge two patches per Fabiano Rosas's comment
> > - Add versions into comments and documentations
> >
> > v2:
> > - Rebase changes on top of recent multifd code changes.
> > - Use QATzip API 'qzMalloc' and 'qzFree' to allocate QAT buffers.
> > - Remove parameter tuning and use QATzip's defaults for better
> >   performance.
> > - Add parameter to enable QAT software fallback.
> >
> > v1:
> > https://lists.nongnu.org/archive/html/qemu-devel/2023-12/msg03761.html
> >
> > * Performance
> >
> > We present updated performance results. For circumstantial reasons, v1
> > presented performance on a low-bandwidth (1Gbps) network.
> >
> > Here, we present updated results with a similar setup as before but with
> > two main differences:
> >
> > 1. Our machines have a ~50Gbps connection, tested using 'iperf3'.
> > 2. We had a bug in our memory allocation causing us to only use ~1/2 of
> > the VM's RAM. Now we properly allocate and fill nearly all of the VM's
> > RAM.
> >
> > Thus, the test setup is as follows:
> >
> > We perform multifd live migration over TCP using a VM with 64GB memory.
> > We prepare the machine's memory by powering it on, allocating a large
> > amount of memory (60GB) as a single buffer, and filling the buffer with
> > the repeated contents of the Silesia corpus[0]. This is in lieu of a
> more
> > realistic memory snapshot, which proved troublesome to acquire.
> >
> > We analyze CPU usage by averaging the output of 'top' every second
> > during migration. This is admittedly imprecise, but we feel that it
> > accurately portrays the different degrees of CPU usage of varying
> > compression methods.
> >
> > We present the latency, throughput, and CPU usage results for all of the
> > compression methods, with varying numbers of multifd threads (4, 8, and
> > 16).
> >
> > [0] The Silesia corpus can be accessed here:
> > https://sun.aei.polsl.pl//~sdeor/index.php?page=silesia
> >
> > ** Results
> >
> > 4 multifd threads:
> >
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> cpu%|
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |qatzip         | 23.13         | 8749.94        |117.50   |186.49
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |zlib           |254.35         |  771.87        |388.20   |144.40
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |zstd           | 54.52         | 3442.59        |414.59   |149.77
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |none           | 12.45         |43739.60        |159.71   |204.96
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >
> > 8 multifd threads:
> >
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> cpu%|
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |qatzip         | 16.91         |12306.52        |186.37   |391.84
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |zlib           |130.11         | 1508.89        |753.86   |289.35
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |zstd           | 27.57         | 6823.23        |786.83   |303.80
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |none           | 11.82         |46072.63        |163.74   |238.56
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >
> > 16 multifd threads:
> >
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> cpu%|
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |qatzip         |18.64          |11044.52        | 573.61  |437.65
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |zlib           |66.43          | 2955.79        |1469.68  |567.47
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |zstd           |14.17          |13290.66        |1504.08  |615.33
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >     |none           |16.82          |32363.26        | 180.74  |217.17
> |
> >     |---------------|---------------|----------------|---------|--------
> -|
> >
> > ** Observations
> >
> > - In general, not using compression outperforms using compression in a
> >   non-network-bound environment.
> > - 'qatzip' outperforms other compression workers with 4 and 8 workers,
> >   achieving a ~91% latency reduction over 'zlib' with 4 workers, and a
> > ~58% latency reduction over 'zstd' with 4 workers.
> > - 'qatzip' maintains comparable performance with 'zstd' at 16 workers,
> >   showing a ~32% increase in latency. This performance difference
> > becomes more noticeable with more workers, as CPU compression is highly
> > parallelizable.
> > - 'qatzip' compression uses considerably less CPU than other compression
> >   methods. At 8 workers, 'qatzip' demonstrates a ~75% reduction in
> > compression CPU usage compared to 'zstd' and 'zlib'.
> > - 'qatzip' decompression CPU usage is less impressive, and is even
> >   slightly worse than 'zstd' and 'zlib' CPU usage at 4 and 16 workers.
> 
> Thanks for the results update.
> 
> It looks like the docs/migration/ file is still missing.  It'll be great
> to
> have it in the next version or separately.
> 
> So how it compares with QPL (which got merged already)?  They at least
> look
> like both supported on an Intel platform, so an user whoever wants to
> compress the RAM could start to look at both.  I'm utterly confused on why
> Intel provides these two similar compressors.  It would be great to have
> some answer and perhaps put into the doc.

I would like to explain some of the reasons why we want to merge the 
two QAT and IAA solutions into the community.

1. Although Intel Xeon Sapphire Rapids supports both QAT and IAA, different 
   SKUs support different numbers of QAT and IAA, so some users do not have 
   both QAT and IAA at the same time.

2. QAT products include PCIe cards, which are compatible with older Xeon
   products and other non-Xeon products. And some users have already used QAT
   cards to accelerate live migration.

3. In addition to compression, QAT and IAA also support various other features 
   to better serve different workloads. Here is an introduction to the accelerators,
   including usage scenarios of QAT and IAA.
   https://www.intel.com/content/dam/www/central-libraries/us/en/documents/2022-12/storage-engines-4th-gen-xeon-brief.pdf

For users who have both QAT and IAA, we recommend the following for choosing a 
live migration solution:

1. If the number of QAT devices is equal to or greater than the number of IAA devices 
   and network bandwidth is limited, it is recommended to use the QATZip(QAT) solution.

2. In other scenarios, the QPL (IAA) solution can be used first.

> I am honestly curious too on whether are you planning to use it in
> production.  It looks like if the network resources are rich, no-comp is
> mostly always better than qatzip, no matter on total migration time or cpu
> consumption.  I'm pretty surprised that it'll take that much resources
> even
> if the work should have been offloaded to the QAT chips iiuc.
> 
> I think it may not be a problem to merge this series even if it performs
> slower at some criterias.. but I think we may still want to know when this
> should be used, or the good reason this should be merged (if it's not
> about
> it outperforms others).
> 
> Thanks,
> 
> >
> >
> > Bryan Zhang (4):
> >   meson: Introduce 'qatzip' feature to the build system
> >   migration: Add migration parameters for QATzip
> >   migration: Introduce 'qatzip' compression method
> >   tests/migration: Add integration test for 'qatzip' compression method
> >
> >  hw/core/qdev-properties-system.c |   6 +-
> >  meson.build                      |  10 +
> >  meson_options.txt                |   2 +
> >  migration/meson.build            |   1 +
> >  migration/migration-hmp-cmds.c   |   8 +
> >  migration/multifd-qatzip.c       | 382 +++++++++++++++++++++++++++++++
> >  migration/multifd.h              |   1 +
> >  migration/options.c              |  57 +++++
> >  migration/options.h              |   2 +
> >  qapi/migration.json              |  38 +++
> >  scripts/meson-buildoptions.sh    |   6 +
> >  tests/qtest/meson.build          |   4 +
> >  tests/qtest/migration-test.c     |  35 +++
> >  13 files changed, 551 insertions(+), 1 deletion(-)
> >  create mode 100644 migration/multifd-qatzip.c
> >
> > --
> > Yichen Wang
> >
> 
> --
> Peter Xu
Peter Xu July 4, 2024, 3:36 p.m. UTC | #3
On Thu, Jul 04, 2024 at 03:15:51AM +0000, Liu, Yuan1 wrote:
> > -----Original Message-----
> > From: Peter Xu <peterx@redhat.com>
> > Sent: Wednesday, July 3, 2024 3:16 AM
> > To: Wang, Yichen <yichen.wang@bytedance.com>
> > Cc: Paolo Bonzini <pbonzini@redhat.com>; Daniel P. Berrangé
> > <berrange@redhat.com>; Eduardo Habkost <eduardo@habkost.net>; Marc-André
> > Lureau <marcandre.lureau@redhat.com>; Thomas Huth <thuth@redhat.com>;
> > Philippe Mathieu-Daudé <philmd@linaro.org>; Fabiano Rosas
> > <farosas@suse.de>; Eric Blake <eblake@redhat.com>; Markus Armbruster
> > <armbru@redhat.com>; Laurent Vivier <lvivier@redhat.com>; qemu-
> > devel@nongnu.org; Hao Xiang <hao.xiang@linux.dev>; Liu, Yuan1
> > <yuan1.liu@intel.com>; Zou, Nanhai <nanhai.zou@intel.com>; Ho-Ren (Jack)
> > Chuang <horenchuang@bytedance.com>
> > Subject: Re: [PATCH v3 0/4] Implement using Intel QAT to offload ZLIB
> > 
> > On Thu, Jun 27, 2024 at 03:34:41PM -0700, Yichen Wang wrote:
> > > v3:
> > > - Rebase changes on top of master
> > > - Merge two patches per Fabiano Rosas's comment
> > > - Add versions into comments and documentations
> > >
> > > v2:
> > > - Rebase changes on top of recent multifd code changes.
> > > - Use QATzip API 'qzMalloc' and 'qzFree' to allocate QAT buffers.
> > > - Remove parameter tuning and use QATzip's defaults for better
> > >   performance.
> > > - Add parameter to enable QAT software fallback.
> > >
> > > v1:
> > > https://lists.nongnu.org/archive/html/qemu-devel/2023-12/msg03761.html
> > >
> > > * Performance
> > >
> > > We present updated performance results. For circumstantial reasons, v1
> > > presented performance on a low-bandwidth (1Gbps) network.
> > >
> > > Here, we present updated results with a similar setup as before but with
> > > two main differences:
> > >
> > > 1. Our machines have a ~50Gbps connection, tested using 'iperf3'.
> > > 2. We had a bug in our memory allocation causing us to only use ~1/2 of
> > > the VM's RAM. Now we properly allocate and fill nearly all of the VM's
> > > RAM.
> > >
> > > Thus, the test setup is as follows:
> > >
> > > We perform multifd live migration over TCP using a VM with 64GB memory.
> > > We prepare the machine's memory by powering it on, allocating a large
> > > amount of memory (60GB) as a single buffer, and filling the buffer with
> > > the repeated contents of the Silesia corpus[0]. This is in lieu of a
> > more
> > > realistic memory snapshot, which proved troublesome to acquire.
> > >
> > > We analyze CPU usage by averaging the output of 'top' every second
> > > during migration. This is admittedly imprecise, but we feel that it
> > > accurately portrays the different degrees of CPU usage of varying
> > > compression methods.
> > >
> > > We present the latency, throughput, and CPU usage results for all of the
> > > compression methods, with varying numbers of multifd threads (4, 8, and
> > > 16).
> > >
> > > [0] The Silesia corpus can be accessed here:
> > > https://sun.aei.polsl.pl//~sdeor/index.php?page=silesia
> > >
> > > ** Results
> > >
> > > 4 multifd threads:
> > >
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> > cpu%|
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |qatzip         | 23.13         | 8749.94        |117.50   |186.49
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |zlib           |254.35         |  771.87        |388.20   |144.40
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |zstd           | 54.52         | 3442.59        |414.59   |149.77
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |none           | 12.45         |43739.60        |159.71   |204.96
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >
> > > 8 multifd threads:
> > >
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> > cpu%|
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |qatzip         | 16.91         |12306.52        |186.37   |391.84
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |zlib           |130.11         | 1508.89        |753.86   |289.35
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |zstd           | 27.57         | 6823.23        |786.83   |303.80
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |none           | 11.82         |46072.63        |163.74   |238.56
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >
> > > 16 multifd threads:
> > >
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> > cpu%|
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |qatzip         |18.64          |11044.52        | 573.61  |437.65
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |zlib           |66.43          | 2955.79        |1469.68  |567.47
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |zstd           |14.17          |13290.66        |1504.08  |615.33
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >     |none           |16.82          |32363.26        | 180.74  |217.17
> > |
> > >     |---------------|---------------|----------------|---------|--------
> > -|
> > >
> > > ** Observations
> > >
> > > - In general, not using compression outperforms using compression in a
> > >   non-network-bound environment.
> > > - 'qatzip' outperforms other compression workers with 4 and 8 workers,
> > >   achieving a ~91% latency reduction over 'zlib' with 4 workers, and a
> > > ~58% latency reduction over 'zstd' with 4 workers.
> > > - 'qatzip' maintains comparable performance with 'zstd' at 16 workers,
> > >   showing a ~32% increase in latency. This performance difference
> > > becomes more noticeable with more workers, as CPU compression is highly
> > > parallelizable.
> > > - 'qatzip' compression uses considerably less CPU than other compression
> > >   methods. At 8 workers, 'qatzip' demonstrates a ~75% reduction in
> > > compression CPU usage compared to 'zstd' and 'zlib'.
> > > - 'qatzip' decompression CPU usage is less impressive, and is even
> > >   slightly worse than 'zstd' and 'zlib' CPU usage at 4 and 16 workers.
> > 
> > Thanks for the results update.
> > 
> > It looks like the docs/migration/ file is still missing.  It'll be great
> > to
> > have it in the next version or separately.
> > 
> > So how it compares with QPL (which got merged already)?  They at least
> > look
> > like both supported on an Intel platform, so an user whoever wants to
> > compress the RAM could start to look at both.  I'm utterly confused on why
> > Intel provides these two similar compressors.  It would be great to have
> > some answer and perhaps put into the doc.

Yuan,

> 
> I would like to explain some of the reasons why we want to merge the 
> two QAT and IAA solutions into the community.

Yes, these are very helpful information.  Please consider putting them into
the cover letter if there's a repost, and perhaps also in the doc/ files.

> 
> 1. Although Intel Xeon Sapphire Rapids supports both QAT and IAA, different 
>    SKUs support different numbers of QAT and IAA, so some users do not have 
>    both QAT and IAA at the same time.
> 
> 2. QAT products include PCIe cards, which are compatible with older Xeon
>    products and other non-Xeon products. And some users have already used QAT
>    cards to accelerate live migration.

Ah, this makes some sense to me.

So a previous question always haunted me, where I wondered why an user who
bought all these fancy and expensive processors with QAT, would still like
to not invest on a better network of 50G or more, but stick with 10Gs
ancient NICs and switches.

So what you're saying is logically in some old clusters with old chips and
old network solutions, it's possible that user buys these PCIe cards
separately so it may help with that old infra migrate VMs faster.  Is that
the case?

If so, we may still want some numbers showing how this performs in a
network-limited environment, and how that helps users to migrate.  Sorry if
there's some back-and-forth requirement asking for these numbers, but I
think these are still important information when an user would like to
decide whether to use these features.  Again, put that into docs/ if proper
would be nice too.

> 
> 3. In addition to compression, QAT and IAA also support various other features 
>    to better serve different workloads. Here is an introduction to the accelerators,
>    including usage scenarios of QAT and IAA.
>    https://www.intel.com/content/dam/www/central-libraries/us/en/documents/2022-12/storage-engines-4th-gen-xeon-brief.pdf

Thanks for the link.

However this doesn't look like a reason to support it in migration?  It
needs to help migration in some form or another, no matter how many
features it provides.. since migration may not consume them.

Two major (but pure..) questions:

  1) Why high cpu usage?

     I raised this question below [1] too but I didn't yet get an answer.
     Ror 8-chan multifd, it's ~390% (QAT) v.s. ~240% (nocomp), even if
     46Gbps bw for the latter... so when throttled it will be even lower?

     The paper you provided above has this upfront:

        When a CPU can offload storage functions to built-in accelerators,
        it frees up cores for business-critical workloads...

     Isn't that a major feature to be able to "offload" things?  Why cpu
     isn't freed even if the offload happened?

  2) TLS?

     I think I asked before, I apologize if any of you've already answered
     and if I forgot.. but have any of you looked into offload TLS (instead
     of compression) with the QATs?

Thanks,

> 
> For users who have both QAT and IAA, we recommend the following for choosing a 
> live migration solution:
> 
> 1. If the number of QAT devices is equal to or greater than the number of IAA devices 
>    and network bandwidth is limited, it is recommended to use the QATZip(QAT) solution.
> 
> 2. In other scenarios, the QPL (IAA) solution can be used first.
> 
> > I am honestly curious too on whether are you planning to use it in
> > production.  It looks like if the network resources are rich, no-comp is
> > mostly always better than qatzip, no matter on total migration time or cpu
> > consumption.  I'm pretty surprised that it'll take that much resources
> > even
> > if the work should have been offloaded to the QAT chips iiuc.

[1]

> > 
> > I think it may not be a problem to merge this series even if it performs
> > slower at some criterias.. but I think we may still want to know when this
> > should be used, or the good reason this should be merged (if it's not
> > about
> > it outperforms others).
> > 
> > Thanks,
> > 
> > >
> > >
> > > Bryan Zhang (4):
> > >   meson: Introduce 'qatzip' feature to the build system
> > >   migration: Add migration parameters for QATzip
> > >   migration: Introduce 'qatzip' compression method
> > >   tests/migration: Add integration test for 'qatzip' compression method
> > >
> > >  hw/core/qdev-properties-system.c |   6 +-
> > >  meson.build                      |  10 +
> > >  meson_options.txt                |   2 +
> > >  migration/meson.build            |   1 +
> > >  migration/migration-hmp-cmds.c   |   8 +
> > >  migration/multifd-qatzip.c       | 382 +++++++++++++++++++++++++++++++
> > >  migration/multifd.h              |   1 +
> > >  migration/options.c              |  57 +++++
> > >  migration/options.h              |   2 +
> > >  qapi/migration.json              |  38 +++
> > >  scripts/meson-buildoptions.sh    |   6 +
> > >  tests/qtest/meson.build          |   4 +
> > >  tests/qtest/migration-test.c     |  35 +++
> > >  13 files changed, 551 insertions(+), 1 deletion(-)
> > >  create mode 100644 migration/multifd-qatzip.c
> > >
> > > --
> > > Yichen Wang
> > >
> > 
> > --
> > Peter Xu
>
Yuan Liu July 5, 2024, 8:32 a.m. UTC | #4
> -----Original Message-----
> From: Peter Xu <peterx@redhat.com>
> Sent: Thursday, July 4, 2024 11:36 PM
> To: Liu, Yuan1 <yuan1.liu@intel.com>
> Cc: Wang, Yichen <yichen.wang@bytedance.com>; Paolo Bonzini
> <pbonzini@redhat.com>; Daniel P. Berrangé <berrange@redhat.com>; Eduardo
> Habkost <eduardo@habkost.net>; Marc-André Lureau
> <marcandre.lureau@redhat.com>; Thomas Huth <thuth@redhat.com>; Philippe
> Mathieu-Daudé <philmd@linaro.org>; Fabiano Rosas <farosas@suse.de>; Eric
> Blake <eblake@redhat.com>; Markus Armbruster <armbru@redhat.com>; Laurent
> Vivier <lvivier@redhat.com>; qemu-devel@nongnu.org; Hao Xiang
> <hao.xiang@linux.dev>; Zou, Nanhai <nanhai.zou@intel.com>; Ho-Ren (Jack)
> Chuang <horenchuang@bytedance.com>
> Subject: Re: [PATCH v3 0/4] Implement using Intel QAT to offload ZLIB
> 
> On Thu, Jul 04, 2024 at 03:15:51AM +0000, Liu, Yuan1 wrote:
> > > -----Original Message-----
> > > From: Peter Xu <peterx@redhat.com>
> > > Sent: Wednesday, July 3, 2024 3:16 AM
> > > To: Wang, Yichen <yichen.wang@bytedance.com>
> > > Cc: Paolo Bonzini <pbonzini@redhat.com>; Daniel P. Berrangé
> > > <berrange@redhat.com>; Eduardo Habkost <eduardo@habkost.net>; Marc-
> André
> > > Lureau <marcandre.lureau@redhat.com>; Thomas Huth <thuth@redhat.com>;
> > > Philippe Mathieu-Daudé <philmd@linaro.org>; Fabiano Rosas
> > > <farosas@suse.de>; Eric Blake <eblake@redhat.com>; Markus Armbruster
> > > <armbru@redhat.com>; Laurent Vivier <lvivier@redhat.com>; qemu-
> > > devel@nongnu.org; Hao Xiang <hao.xiang@linux.dev>; Liu, Yuan1
> > > <yuan1.liu@intel.com>; Zou, Nanhai <nanhai.zou@intel.com>; Ho-Ren
> (Jack)
> > > Chuang <horenchuang@bytedance.com>
> > > Subject: Re: [PATCH v3 0/4] Implement using Intel QAT to offload ZLIB
> > >
> > > On Thu, Jun 27, 2024 at 03:34:41PM -0700, Yichen Wang wrote:
> > > > v3:
> > > > - Rebase changes on top of master
> > > > - Merge two patches per Fabiano Rosas's comment
> > > > - Add versions into comments and documentations
> > > >
> > > > v2:
> > > > - Rebase changes on top of recent multifd code changes.
> > > > - Use QATzip API 'qzMalloc' and 'qzFree' to allocate QAT buffers.
> > > > - Remove parameter tuning and use QATzip's defaults for better
> > > >   performance.
> > > > - Add parameter to enable QAT software fallback.
> > > >
> > > > v1:
> > > > https://lists.nongnu.org/archive/html/qemu-devel/2023-
> 12/msg03761.html
> > > >
> > > > * Performance
> > > >
> > > > We present updated performance results. For circumstantial reasons,
> v1
> > > > presented performance on a low-bandwidth (1Gbps) network.
> > > >
> > > > Here, we present updated results with a similar setup as before but
> with
> > > > two main differences:
> > > >
> > > > 1. Our machines have a ~50Gbps connection, tested using 'iperf3'.
> > > > 2. We had a bug in our memory allocation causing us to only use ~1/2
> of
> > > > the VM's RAM. Now we properly allocate and fill nearly all of the
> VM's
> > > > RAM.
> > > >
> > > > Thus, the test setup is as follows:
> > > >
> > > > We perform multifd live migration over TCP using a VM with 64GB
> memory.
> > > > We prepare the machine's memory by powering it on, allocating a
> large
> > > > amount of memory (60GB) as a single buffer, and filling the buffer
> with
> > > > the repeated contents of the Silesia corpus[0]. This is in lieu of a
> > > more
> > > > realistic memory snapshot, which proved troublesome to acquire.
> > > >
> > > > We analyze CPU usage by averaging the output of 'top' every second
> > > > during migration. This is admittedly imprecise, but we feel that it
> > > > accurately portrays the different degrees of CPU usage of varying
> > > > compression methods.
> > > >
> > > > We present the latency, throughput, and CPU usage results for all of
> the
> > > > compression methods, with varying numbers of multifd threads (4, 8,
> and
> > > > 16).
> > > >
> > > > [0] The Silesia corpus can be accessed here:
> > > > https://sun.aei.polsl.pl//~sdeor/index.php?page=silesia
> > > >
> > > > ** Results
> > > >
> > > > 4 multifd threads:
> > > >
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> > > cpu%|
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |qatzip         | 23.13         | 8749.94        |117.50
> |186.49
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |zlib           |254.35         |  771.87        |388.20
> |144.40
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |zstd           | 54.52         | 3442.59        |414.59
> |149.77
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |none           | 12.45         |43739.60        |159.71
> |204.96
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >
> > > > 8 multifd threads:
> > > >
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> > > cpu%|
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |qatzip         | 16.91         |12306.52        |186.37
> |391.84
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |zlib           |130.11         | 1508.89        |753.86
> |289.35
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |zstd           | 27.57         | 6823.23        |786.83
> |303.80
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |none           | 11.82         |46072.63        |163.74
> |238.56
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >
> > > > 16 multifd threads:
> > > >
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |method         |time(sec)      |throughput(mbps)|send cpu%|recv
> > > cpu%|
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |qatzip         |18.64          |11044.52        | 573.61
> |437.65
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |zlib           |66.43          | 2955.79        |1469.68
> |567.47
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |zstd           |14.17          |13290.66        |1504.08
> |615.33
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >     |none           |16.82          |32363.26        | 180.74
> |217.17
> > > |
> > > >     |---------------|---------------|----------------|---------|----
> ----
> > > -|
> > > >
> > > > ** Observations
> > > >
> > > > - In general, not using compression outperforms using compression in
> a
> > > >   non-network-bound environment.
> > > > - 'qatzip' outperforms other compression workers with 4 and 8
> workers,
> > > >   achieving a ~91% latency reduction over 'zlib' with 4 workers, and
> a
> > > > ~58% latency reduction over 'zstd' with 4 workers.
> > > > - 'qatzip' maintains comparable performance with 'zstd' at 16
> workers,
> > > >   showing a ~32% increase in latency. This performance difference
> > > > becomes more noticeable with more workers, as CPU compression is
> highly
> > > > parallelizable.
> > > > - 'qatzip' compression uses considerably less CPU than other
> compression
> > > >   methods. At 8 workers, 'qatzip' demonstrates a ~75% reduction in
> > > > compression CPU usage compared to 'zstd' and 'zlib'.
> > > > - 'qatzip' decompression CPU usage is less impressive, and is even
> > > >   slightly worse than 'zstd' and 'zlib' CPU usage at 4 and 16
> workers.
> > >
> > > Thanks for the results update.
> > >
> > > It looks like the docs/migration/ file is still missing.  It'll be
> great
> > > to
> > > have it in the next version or separately.
> > >
> > > So how it compares with QPL (which got merged already)?  They at least
> > > look
> > > like both supported on an Intel platform, so an user whoever wants to
> > > compress the RAM could start to look at both.  I'm utterly confused on
> why
> > > Intel provides these two similar compressors.  It would be great to
> have
> > > some answer and perhaps put into the doc.
> 
> Yuan,
> 
> >
> > I would like to explain some of the reasons why we want to merge the
> > two QAT and IAA solutions into the community.
> 
> Yes, these are very helpful information.  Please consider putting them
> into
> the cover letter if there's a repost, and perhaps also in the doc/ files.
> 
> >
> > 1. Although Intel Xeon Sapphire Rapids supports both QAT and IAA,
> different
> >    SKUs support different numbers of QAT and IAA, so some users do not
> have
> >    both QAT and IAA at the same time.
> >
> > 2. QAT products include PCIe cards, which are compatible with older Xeon
> >    products and other non-Xeon products. And some users have already
> used QAT
> >    cards to accelerate live migration.
> 
> Ah, this makes some sense to me.
> 
> So a previous question always haunted me, where I wondered why an user who
> bought all these fancy and expensive processors with QAT, would still like
> to not invest on a better network of 50G or more, but stick with 10Gs
> ancient NICs and switches.
> 
> So what you're saying is logically in some old clusters with old chips and
> old network solutions, it's possible that user buys these PCIe cards
> separately so it may help with that old infra migrate VMs faster.  Is that
> the case?

Yes, users do not add a QAT card just for live migration. Users mainly use 
QAT-SRIOV technology to help cloud users offload compression and encryption.

> If so, we may still want some numbers showing how this performs in a
> network-limited environment, and how that helps users to migrate.  Sorry
> if
> there's some back-and-forth requirement asking for these numbers, but I
> think these are still important information when an user would like to
> decide whether to use these features.  Again, put that into docs/ if
> proper
> would be nice too.

Yes, I will provide some performance data at some specific 
bandwidths(100Mbps/1Gbps/10Gbps). And add documentation to explain 
the advantages of using QAT 

> >
> > 3. In addition to compression, QAT and IAA also support various other
> features
> >    to better serve different workloads. Here is an introduction to the
> accelerators,
> >    including usage scenarios of QAT and IAA.
> >    https://www.intel.com/content/dam/www/central-
> libraries/us/en/documents/2022-12/storage-engines-4th-gen-xeon-brief.pdf
> 
> Thanks for the link.
> 
> However this doesn't look like a reason to support it in migration?  It
> needs to help migration in some form or another, no matter how many
> features it provides.. since migration may not consume them.
> 
> Two major (but pure..) questions:
> 
>   1) Why high cpu usage?
> 
>      I raised this question below [1] too but I didn't yet get an answer.
>      Ror 8-chan multifd, it's ~390% (QAT) v.s. ~240% (nocomp), even if
>      46Gbps bw for the latter... so when throttled it will be even lower?
> 
>      The paper you provided above has this upfront:
> 
>         When a CPU can offload storage functions to built-in accelerators,
>         it frees up cores for business-critical workloads...
> 
>      Isn't that a major feature to be able to "offload" things?  Why cpu
>      isn't freed even if the offload happened?

Yes, it doesn't make sense, I will check this.

>   2) TLS?
> 
>      I think I asked before, I apologize if any of you've already answered
>      and if I forgot.. but have any of you looked into offload TLS
> (instead
>      of compression) with the QATs?

I'm sorry for not responding to the previous question about TLS. QAT has many 
related success cases (for example, OpenSSL). 
https://www.intel.com/content/dam/www/public/us/en/documents/solution-briefs/accelerating-openssl-brief.pdf

I will send a separate RFC or patch about this part because the software stacks of QAT 
compression and encryption are independent, so we discuss them separately.

> Thanks,
> 
> >
> > For users who have both QAT and IAA, we recommend the following for
> choosing a
> > live migration solution:
> >
> > 1. If the number of QAT devices is equal to or greater than the number
> of IAA devices
> >    and network bandwidth is limited, it is recommended to use the
> QATZip(QAT) solution.
> >
> > 2. In other scenarios, the QPL (IAA) solution can be used first.
> >
> > > I am honestly curious too on whether are you planning to use it in
> > > production.  It looks like if the network resources are rich, no-comp
> is
> > > mostly always better than qatzip, no matter on total migration time or
> cpu
> > > consumption.  I'm pretty surprised that it'll take that much resources
> > > even
> > > if the work should have been offloaded to the QAT chips iiuc.
> 
> [1]
> 
> > >
> > > I think it may not be a problem to merge this series even if it
> performs
> > > slower at some criterias.. but I think we may still want to know when
> this
> > > should be used, or the good reason this should be merged (if it's not
> > > about
> > > it outperforms others).
> > >
> > > Thanks,
> > >
> > > >
> > > >
> > > > Bryan Zhang (4):
> > > >   meson: Introduce 'qatzip' feature to the build system
> > > >   migration: Add migration parameters for QATzip
> > > >   migration: Introduce 'qatzip' compression method
> > > >   tests/migration: Add integration test for 'qatzip' compression
> method
> > > >
> > > >  hw/core/qdev-properties-system.c |   6 +-
> > > >  meson.build                      |  10 +
> > > >  meson_options.txt                |   2 +
> > > >  migration/meson.build            |   1 +
> > > >  migration/migration-hmp-cmds.c   |   8 +
> > > >  migration/multifd-qatzip.c       | 382
> +++++++++++++++++++++++++++++++
> > > >  migration/multifd.h              |   1 +
> > > >  migration/options.c              |  57 +++++
> > > >  migration/options.h              |   2 +
> > > >  qapi/migration.json              |  38 +++
> > > >  scripts/meson-buildoptions.sh    |   6 +
> > > >  tests/qtest/meson.build          |   4 +
> > > >  tests/qtest/migration-test.c     |  35 +++
> > > >  13 files changed, 551 insertions(+), 1 deletion(-)
> > > >  create mode 100644 migration/multifd-qatzip.c
> > > >
> > > > --
> > > > Yichen Wang
> > > >
> > >
> > > --
> > > Peter Xu
> >
> 
> --
> Peter Xu
Yichen Wang July 5, 2024, 6:28 p.m. UTC | #5
> On Jul 5, 2024, at 1:32 AM, Liu, Yuan1 <yuan1.liu@intel.com> wrote:
> 
>> -----Original Message-----
>> From: Peter Xu <peterx@redhat.com <mailto:peterx@redhat.com>>
>> Sent: Thursday, July 4, 2024 11:36 PM
>> To: Liu, Yuan1 <yuan1.liu@intel.com <mailto:yuan1.liu@intel.com>>
>> Cc: Wang, Yichen <yichen.wang@bytedance.com <mailto:yichen.wang@bytedance.com>>; Paolo Bonzini
>> <pbonzini@redhat.com <mailto:pbonzini@redhat.com>>; Daniel P. Berrangé <berrange@redhat.com <mailto:berrange@redhat.com>>; Eduardo
>> Habkost <eduardo@habkost.net <mailto:eduardo@habkost.net>>; Marc-André Lureau
>> <marcandre.lureau@redhat.com <mailto:marcandre.lureau@redhat.com>>; Thomas Huth <thuth@redhat.com <mailto:thuth@redhat.com>>; Philippe
>> Mathieu-Daudé <philmd@linaro.org <mailto:philmd@linaro.org>>; Fabiano Rosas <farosas@suse.de <mailto:farosas@suse.de>>; Eric
>> Blake <eblake@redhat.com <mailto:eblake@redhat.com>>; Markus Armbruster <armbru@redhat.com <mailto:armbru@redhat.com>>; Laurent
>> Vivier <lvivier@redhat.com <mailto:lvivier@redhat.com>>; qemu-devel@nongnu.org <mailto:qemu-devel@nongnu.org>; Hao Xiang
>> <hao.xiang@linux.dev <mailto:hao.xiang@linux.dev>>; Zou, Nanhai <nanhai.zou@intel.com <mailto:nanhai.zou@intel.com>>; Ho-Ren (Jack)
>> Chuang <horenchuang@bytedance.com <mailto:horenchuang@bytedance.com>>
>> Subject: Re: [PATCH v3 0/4] Implement using Intel QAT to offload ZLIB
>> 
>> On Thu, Jul 04, 2024 at 03:15:51AM +0000, Liu, Yuan1 wrote:
>>>> -----Original Message-----
>>>> From: Peter Xu <peterx@redhat.com>
>>>> Sent: Wednesday, July 3, 2024 3:16 AM
>>>> To: Wang, Yichen <yichen.wang@bytedance.com>
>>>> Cc: Paolo Bonzini <pbonzini@redhat.com>; Daniel P. Berrangé
>>>> <berrange@redhat.com>; Eduardo Habkost <eduardo@habkost.net>; Marc-
>> André
>>>> Lureau <marcandre.lureau@redhat.com>; Thomas Huth <thuth@redhat.com>;
>>>> Philippe Mathieu-Daudé <philmd@linaro.org>; Fabiano Rosas
>>>> <farosas@suse.de>; Eric Blake <eblake@redhat.com>; Markus Armbruster
>>>> <armbru@redhat.com>; Laurent Vivier <lvivier@redhat.com>; qemu-
>>>> devel@nongnu.org; Hao Xiang <hao.xiang@linux.dev>; Liu, Yuan1
>>>> <yuan1.liu@intel.com>; Zou, Nanhai <nanhai.zou@intel.com>; Ho-Ren
>> (Jack)
>>>> Chuang <horenchuang@bytedance.com>
>>>> Subject: Re: [PATCH v3 0/4] Implement using Intel QAT to offload ZLIB
>>>> 
>>>> On Thu, Jun 27, 2024 at 03:34:41PM -0700, Yichen Wang wrote:
>>>>> v3:
>>>>> - Rebase changes on top of master
>>>>> - Merge two patches per Fabiano Rosas's comment
>>>>> - Add versions into comments and documentations
>>>>> 
>>>>> v2:
>>>>> - Rebase changes on top of recent multifd code changes.
>>>>> - Use QATzip API 'qzMalloc' and 'qzFree' to allocate QAT buffers.
>>>>> - Remove parameter tuning and use QATzip's defaults for better
>>>>>  performance.
>>>>> - Add parameter to enable QAT software fallback.
>>>>> 
>>>>> v1:
>>>>> https://lists.nongnu.org/archive/html/qemu-devel/2023-
>> 12/msg03761.html
>>>>> 
>>>>> * Performance
>>>>> 
>>>>> We present updated performance results. For circumstantial reasons,
>> v1
>>>>> presented performance on a low-bandwidth (1Gbps) network.
>>>>> 
>>>>> Here, we present updated results with a similar setup as before but
>> with
>>>>> two main differences:
>>>>> 
>>>>> 1. Our machines have a ~50Gbps connection, tested using 'iperf3'.
>>>>> 2. We had a bug in our memory allocation causing us to only use ~1/2
>> of
>>>>> the VM's RAM. Now we properly allocate and fill nearly all of the
>> VM's
>>>>> RAM.
>>>>> 
>>>>> Thus, the test setup is as follows:
>>>>> 
>>>>> We perform multifd live migration over TCP using a VM with 64GB
>> memory.
>>>>> We prepare the machine's memory by powering it on, allocating a
>> large
>>>>> amount of memory (60GB) as a single buffer, and filling the buffer
>> with
>>>>> the repeated contents of the Silesia corpus[0]. This is in lieu of a
>>>> more
>>>>> realistic memory snapshot, which proved troublesome to acquire.
>>>>> 
>>>>> We analyze CPU usage by averaging the output of 'top' every second
>>>>> during migration. This is admittedly imprecise, but we feel that it
>>>>> accurately portrays the different degrees of CPU usage of varying
>>>>> compression methods.
>>>>> 
>>>>> We present the latency, throughput, and CPU usage results for all of
>> the
>>>>> compression methods, with varying numbers of multifd threads (4, 8,
>> and
>>>>> 16).
>>>>> 
>>>>> [0] The Silesia corpus can be accessed here:
>>>>> https://sun.aei.polsl.pl//~sdeor/index.php?page=silesia
>>>>> 
>>>>> ** Results
>>>>> 
>>>>> 4 multifd threads:
>>>>> 
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |method         |time(sec)      |throughput(mbps)|send cpu%|recv
>>>> cpu%|
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |qatzip         | 23.13         | 8749.94        |117.50
>> |186.49
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |zlib           |254.35         |  771.87        |388.20
>> |144.40
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |zstd           | 54.52         | 3442.59        |414.59
>> |149.77
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |none           | 12.45         |43739.60        |159.71
>> |204.96
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>> 
>>>>> 8 multifd threads:
>>>>> 
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |method         |time(sec)      |throughput(mbps)|send cpu%|recv
>>>> cpu%|
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |qatzip         | 16.91         |12306.52        |186.37
>> |391.84
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |zlib           |130.11         | 1508.89        |753.86
>> |289.35
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |zstd           | 27.57         | 6823.23        |786.83
>> |303.80
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |none           | 11.82         |46072.63        |163.74
>> |238.56
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>> 
>>>>> 16 multifd threads:
>>>>> 
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |method         |time(sec)      |throughput(mbps)|send cpu%|recv
>>>> cpu%|
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |qatzip         |18.64          |11044.52        | 573.61
>> |437.65
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |zlib           |66.43          | 2955.79        |1469.68
>> |567.47
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |zstd           |14.17          |13290.66        |1504.08
>> |615.33
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>>    |none           |16.82          |32363.26        | 180.74
>> |217.17
>>>> |
>>>>>    |---------------|---------------|----------------|---------|----
>> ----
>>>> -|
>>>>> 
>>>>> ** Observations
>>>>> 
>>>>> - In general, not using compression outperforms using compression in
>> a
>>>>>  non-network-bound environment.
>>>>> - 'qatzip' outperforms other compression workers with 4 and 8
>> workers,
>>>>>  achieving a ~91% latency reduction over 'zlib' with 4 workers, and
>> a
>>>>> ~58% latency reduction over 'zstd' with 4 workers.
>>>>> - 'qatzip' maintains comparable performance with 'zstd' at 16
>> workers,
>>>>>  showing a ~32% increase in latency. This performance difference
>>>>> becomes more noticeable with more workers, as CPU compression is
>> highly
>>>>> parallelizable.
>>>>> - 'qatzip' compression uses considerably less CPU than other
>> compression
>>>>>  methods. At 8 workers, 'qatzip' demonstrates a ~75% reduction in
>>>>> compression CPU usage compared to 'zstd' and 'zlib'.
>>>>> - 'qatzip' decompression CPU usage is less impressive, and is even
>>>>>  slightly worse than 'zstd' and 'zlib' CPU usage at 4 and 16
>> workers.
>>>> 
>>>> Thanks for the results update.
>>>> 
>>>> It looks like the docs/migration/ file is still missing.  It'll be
>> great
>>>> to
>>>> have it in the next version or separately.
>>>> 
>>>> So how it compares with QPL (which got merged already)?  They at least
>>>> look
>>>> like both supported on an Intel platform, so an user whoever wants to
>>>> compress the RAM could start to look at both.  I'm utterly confused on
>> why
>>>> Intel provides these two similar compressors.  It would be great to
>> have
>>>> some answer and perhaps put into the doc.
>> 
>> Yuan,
>> 
>>> 
>>> I would like to explain some of the reasons why we want to merge the
>>> two QAT and IAA solutions into the community.
>> 
>> Yes, these are very helpful information.  Please consider putting them
>> into
>> the cover letter if there's a repost, and perhaps also in the doc/ files.
>> 
>>> 
>>> 1. Although Intel Xeon Sapphire Rapids supports both QAT and IAA,
>> different
>>>   SKUs support different numbers of QAT and IAA, so some users do not
>> have
>>>   both QAT and IAA at the same time.
>>> 
>>> 2. QAT products include PCIe cards, which are compatible with older Xeon
>>>   products and other non-Xeon products. And some users have already
>> used QAT
>>>   cards to accelerate live migration.
>> 
>> Ah, this makes some sense to me.
>> 
>> So a previous question always haunted me, where I wondered why an user who
>> bought all these fancy and expensive processors with QAT, would still like
>> to not invest on a better network of 50G or more, but stick with 10Gs
>> ancient NICs and switches.
>> 
>> So what you're saying is logically in some old clusters with old chips and
>> old network solutions, it's possible that user buys these PCIe cards
>> separately so it may help with that old infra migrate VMs faster.  Is that
>> the case?
> 
> Yes, users do not add a QAT card just for live migration. Users mainly use 
> QAT-SRIOV technology to help cloud users offload compression and encryption.
> 
>> If so, we may still want some numbers showing how this performs in a
>> network-limited environment, and how that helps users to migrate.  Sorry
>> if
>> there's some back-and-forth requirement asking for these numbers, but I
>> think these are still important information when an user would like to
>> decide whether to use these features.  Again, put that into docs/ if
>> proper
>> would be nice too.
> 
> Yes, I will provide some performance data at some specific 
> bandwidths(100Mbps/1Gbps/10Gbps). And add documentation to explain 
> the advantages of using QAT 
Just want to add some information here. So in ByteDance, the current generation server is quipped with 2*100Gb NIC. We reserve 10Gbps for control plane purposes which includes live migration here. So it is not about we are using “good network”, it is about not normal to use full bandwidth for control plane purposes. Hence we do have a requirements for QAT/IAA in these cases.
>>> 
>>> 3. In addition to compression, QAT and IAA also support various other
>> features
>>>   to better serve different workloads. Here is an introduction to the
>> accelerators,
>>>   including usage scenarios of QAT and IAA.
>>>   https://www.intel.com/content/dam/www/central-
>> libraries/us/en/documents/2022-12/storage-engines-4th-gen-xeon-brief.pdf
>> 
>> Thanks for the link.
>> 
>> However this doesn't look like a reason to support it in migration?  It
>> needs to help migration in some form or another, no matter how many
>> features it provides.. since migration may not consume them.
>> 
>> Two major (but pure..) questions:
>> 
>>  1) Why high cpu usage?
>> 
>>     I raised this question below [1] too but I didn't yet get an answer.
>>     Ror 8-chan multifd, it's ~390% (QAT) v.s. ~240% (nocomp), even if
>>     46Gbps bw for the latter... so when throttled it will be even lower?
>> 
>>     The paper you provided above has this upfront:
>> 
>>        When a CPU can offload storage functions to built-in accelerators,
>>        it frees up cores for business-critical workloads...
>> 
>>     Isn't that a major feature to be able to "offload" things?  Why cpu
>>     isn't freed even if the offload happened?
> 
> Yes, it doesn't make sense, I will check this.
> 
>>  2) TLS?
>> 
>>     I think I asked before, I apologize if any of you've already answered
>>     and if I forgot.. but have any of you looked into offload TLS
>> (instead
>>     of compression) with the QATs?
> 
> I'm sorry for not responding to the previous question about TLS. QAT has many 
> related success cases (for example, OpenSSL). 
> https://www.intel.com/content/dam/www/public/us/en/documents/solution-briefs/accelerating-openssl-brief.pdf
> 
> I will send a separate RFC or patch about this part because the software stacks of QAT 
> compression and encryption are independent, so we discuss them separately.
> 
>> Thanks,
>> 
>>> 
>>> For users who have both QAT and IAA, we recommend the following for
>> choosing a
>>> live migration solution:
>>> 
>>> 1. If the number of QAT devices is equal to or greater than the number
>> of IAA devices
>>>   and network bandwidth is limited, it is recommended to use the
>> QATZip(QAT) solution.
>>> 
>>> 2. In other scenarios, the QPL (IAA) solution can be used first.
>>> 
>>>> I am honestly curious too on whether are you planning to use it in
>>>> production.  It looks like if the network resources are rich, no-comp
>> is
>>>> mostly always better than qatzip, no matter on total migration time or
>> cpu
>>>> consumption.  I'm pretty surprised that it'll take that much resources
>>>> even
>>>> if the work should have been offloaded to the QAT chips iiuc.
>> 
>> [1]
>> 
>>>> 
>>>> I think it may not be a problem to merge this series even if it
>> performs
>>>> slower at some criterias.. but I think we may still want to know when
>> this
>>>> should be used, or the good reason this should be merged (if it's not
>>>> about
>>>> it outperforms others).
>>>> 
>>>> Thanks,
>>>> 
>>>>> 
>>>>> 
>>>>> Bryan Zhang (4):
>>>>>  meson: Introduce 'qatzip' feature to the build system
>>>>>  migration: Add migration parameters for QATzip
>>>>>  migration: Introduce 'qatzip' compression method
>>>>>  tests/migration: Add integration test for 'qatzip' compression
>> method
>>>>> 
>>>>> hw/core/qdev-properties-system.c |   6 +-
>>>>> meson.build                      |  10 +
>>>>> meson_options.txt                |   2 +
>>>>> migration/meson.build            |   1 +
>>>>> migration/migration-hmp-cmds.c   |   8 +
>>>>> migration/multifd-qatzip.c       | 382
>> +++++++++++++++++++++++++++++++
>>>>> migration/multifd.h              |   1 +
>>>>> migration/options.c              |  57 +++++
>>>>> migration/options.h              |   2 +
>>>>> qapi/migration.json              |  38 +++
>>>>> scripts/meson-buildoptions.sh    |   6 +
>>>>> tests/qtest/meson.build          |   4 +
>>>>> tests/qtest/migration-test.c     |  35 +++
>>>>> 13 files changed, 551 insertions(+), 1 deletion(-)
>>>>> create mode 100644 migration/multifd-qatzip.c
>>>>> 
>>>>> --
>>>>> Yichen Wang
>>>>> 
>>>> 
>>>> --
>>>> Peter Xu
>>> 
>> 
>> --
>> Peter Xu
Peter Xu July 8, 2024, 3:46 p.m. UTC | #6
On Fri, Jul 05, 2024 at 11:28:25AM -0700, Yichen Wang wrote:
> Just want to add some information here. So in ByteDance, the current
> generation server is quipped with 2*100Gb NIC. We reserve 10Gbps for
> control plane purposes which includes live migration here. So it is not
> about we are using “good network”, it is about not normal to use full
> bandwidth for control plane purposes. Hence we do have a requirements for
> QAT/IAA in these cases.

Yes, this makes sense.

But then you may then also want to figure out the high cpu consumption of
those cards, to not interrupt more important workloads?

I saw there's a new version posted, I didn't see an explanation of the cpu
consumption issue mentioned.  Meanwhile I also see that the docs/ update is
missing.

Would you consider adding both by replying to the new version?