| Message ID | 20240529094435.11140-1-shameerali.kolothum.thodi@huawei.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Live migration acceleration with UADK | expand |
On Wed, May 29, 2024 at 10:44:20AM +0100, Shameer Kolothum via wrote: > Hi, > > This series adds support for UADK library based hardware acceleration > for live migration. UADK[0] is a general-purpose user space accelerator > framework that uses shared virtual addressing (SVA) to provide a unified > programming interface for hardware acceleration of cryptographic and > compression algorithms. > > UADK makes use of the UACCE(Unified/User-space-access-intended Accelerator > Framework) Linux kernel module which enables hardware accelerators from > different vendors that support SVA to adapt to UADK. Linux kernel from > v5.9 has support for UACCE and SVA on ARM64 platforms. > > Currently, HiSilicon Kunpeng hardware accelerators have been registered with > UACCE and the Zip accelerator on these platforms can be used for compression > which can free up CPU computing power and improve computing performance. > > This series is on top of Intel IAA accelerator live migration support > series[1] from Yuan Liu. Many thanks for doing this. Just looked at the IAA series too, I didn't read multifd-*.[ch] much on both sides but both the series look pretty clean to me. Thanks,
Hi, This series adds support for UADK library based hardware acceleration for live migration. UADK[0] is a general-purpose user space accelerator framework that uses shared virtual addressing (SVA) to provide a unified programming interface for hardware acceleration of cryptographic and compression algorithms. UADK makes use of the UACCE(Unified/User-space-access-intended Accelerator Framework) Linux kernel module which enables hardware accelerators from different vendors that support SVA to adapt to UADK. Linux kernel from v5.9 has support for UACCE and SVA on ARM64 platforms. Currently, HiSilicon Kunpeng hardware accelerators have been registered with UACCE and the Zip accelerator on these platforms can be used for compression which can free up CPU computing power and improve computing performance. This series is on top of Intel IAA accelerator live migration support series[1] from Yuan Liu. Many thanks for doing this. Initial tests were carried out on HiSilicon D06 platforms and the results are as below: Test setup: HiSilicon D06 boards connected over a 1Gbps n/w. Host Kernel: Host Kernel: 6.7.0 Mainline Kernel. Guest VM: 64 cpus, 16GB mem, hugepages, prealloc=on (80% Memory filled with random data) +--------+-------------+--------+--------+----------+------+ | | The number |total |downtime|pages per | CPU | | None | of channels |time(ms)|(ms) |second | Util | | Comp | | | | | | | +-------------+-----------------+----------+------+ |Network | 2| 114536| 79| 32849 | 18%| |BW: 1G +-------------+--------+--------+----------+------+ | | 4| 114327| 78| 34217 | 22%| | +-------------+--------+--------+----------+------+ | | 8| 114231| 107| 211840 | 24%| +--------+-------------+--------+--------+----------+------+ +--------+-------------+--------+--------+----------+------+ | | The number |total |downtime|pages per | CPU | | UADK | of channels |time(ms)|(ms) |second | Util | | Comp | | | | | | | +-------------+-----------------+----------+------+ |Network | 2| 77192 | 75| 182679 | 24%| |BW: 1G +-------------+--------+--------+----------+------+ | | 4| 77000 | 86| 185600 | 25%| | +-------------+--------+--------+----------+------+ | | 8| 76835 | 97| 330966 | 27%| +--------+-------------+--------+--------+----------+------+ +--------+-------------+--------+--------+----------+------+ | | The number |total |downtime|pages per | CPU | | ZLIB | of channels |time(ms)|(ms) |second | Util | | Comp | | | | | | | +-------------+-----------------+----------+------+ |Network | 2| 134664| 73| 42666 | 200%| |BW: 1G +-------------+--------+--------+----------+------+ | | 4| 71550 | 72| 181227 | 390%| | +-------------+--------+--------+----------+------+ | | 8| 67781 | 108| 200960 | 460%| +--------+-------------+--------+--------+----------+------+ +--------+-------------+--------+--------+----------+------+ | | The number |total |downtime|pages per | CPU | | ZSTD | of channels |time(ms)|(ms) |second | Util | | Comp | | | | | | | +-------------+-----------------+----------+------+ |Network | 2| 67822 | 73| 202772 | 160%| |BW: 1G +-------------+--------+--------+----------+------+ | | 4| 67460 | 107| 198400 | 180%| | +-------------+--------+--------+----------+------+ | | 8| 67422 | 83 | 349808 | 215%| +--------+-------------+--------+--------+----------+------+ From the above results, UADK has considerable CPU cycle savings compared to both Zlib/Zstd. Also compared with Qemu "multifd-compression none" mode UADK has an edge on migration "total time". Please take a look and let me know your feedback. Thanks, Shameer [0] https://github.com/Linaro/uadk/tree/master/docs [1] https://lore.kernel.org/qemu-devel/20240505165751.2392198-1-yuan1.liu@intel.com/ Shameer Kolothum (7): docs/migration: add uadk compression feature configure: Add uadk option migration/multifd: add uadk compression framework migration/multifd: Add UADK initialization migration/multifd: Add UADK based compression and decompression migration/multifd: Switch to no compression when no hardware support tests/migration-test: add uadk compression test docs/devel/migration/uadk-compression.rst | 144 +++++++++ hw/core/qdev-properties-system.c | 2 +- meson.build | 14 + meson_options.txt | 2 + migration/meson.build | 1 + migration/multifd-uadk.c | 368 ++++++++++++++++++++++ migration/multifd.h | 5 +- qapi/migration.json | 5 +- scripts/meson-buildoptions.sh | 3 + tests/qtest/migration-test.c | 23 ++ 10 files changed, 563 insertions(+), 4 deletions(-) create mode 100644 docs/devel/migration/uadk-compression.rst create mode 100644 migration/multifd-uadk.c