Message ID | 20240319164527.1873891-2-yuan1.liu@intel.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | Live Migration With IAA | expand |
On Wed, Mar 20, 2024 at 12:45:21AM +0800, Yuan Liu wrote: > add Intel Query Processing Library (QPL) compression method > introduction > > Signed-off-by: Yuan Liu <yuan1.liu@intel.com> > Reviewed-by: Nanhai Zou <nanhai.zou@intel.com> > --- > docs/devel/migration/features.rst | 1 + > docs/devel/migration/qpl-compression.rst | 231 +++++++++++++++++++++++ > 2 files changed, 232 insertions(+) > create mode 100644 docs/devel/migration/qpl-compression.rst > > diff --git a/docs/devel/migration/features.rst b/docs/devel/migration/features.rst > index d5ca7b86d5..bc98b65075 100644 > --- a/docs/devel/migration/features.rst > +++ b/docs/devel/migration/features.rst > @@ -12,3 +12,4 @@ Migration has plenty of features to support different use cases. > virtio > mapped-ram > CPR > + qpl-compression > diff --git a/docs/devel/migration/qpl-compression.rst b/docs/devel/migration/qpl-compression.rst > new file mode 100644 > index 0000000000..42c7969d30 > --- /dev/null > +++ b/docs/devel/migration/qpl-compression.rst > @@ -0,0 +1,231 @@ > +=============== > +QPL Compression > +=============== > +The Intel Query Processing Library (Intel ``QPL``) is an open-source library to > +provide compression and decompression features and it is based on deflate > +compression algorithm (RFC 1951). > + > +The ``QPL`` compression relies on Intel In-Memory Analytics Accelerator(``IAA``) > +and Shared Virtual Memory(``SVM``) technology, they are new features supported > +from Intel 4th Gen Intel Xeon Scalable processors, codenamed Sapphire Rapids > +processor(``SPR``). > + > +For more ``QPL`` introduction, please refer to: > + > +https://intel.github.io/qpl/documentation/introduction_docs/introduction.html There're a bunch of links in this page, please consider switching all of them to use the link formats of .rST: Please refer to `QPL introduction page <https://...>`_. > + > +QPL Compression Framework > +========================= > + > +:: > + > + +----------------+ +------------------+ > + | MultiFD Service| |accel-config tool | > + +-------+--------+ +--------+---------+ > + | | > + | | > + +-------+--------+ | Setup IAA > + | QPL library | | Resources > + +-------+---+----+ | > + | | | > + | +-------------+-------+ > + | Open IAA | > + | Devices +-----+-----+ > + | |idxd driver| > + | +-----+-----+ > + | | > + | | > + | +-----+-----+ > + +-----------+IAA Devices| > + Submit jobs +-----------+ > + via enqcmd > + > + > +Intel In-Memory Analytics Accelerator (Intel IAA) Introduction > +================================================================ > + > +Intel ``IAA`` is an accelerator that has been designed to help benefit > +in-memory databases and analytic workloads. There are three main areas > +that Intel ``IAA`` can assist with analytics primitives (scan, filter, etc.), > +sparse data compression and memory tiering. > + > +``IAA`` Manual Documentation: > + > +https://www.intel.com/content/www/us/en/content-details/721858/intel-in-memory-analytics-accelerator-architecture-specification > + > +IAA Device Enabling > +------------------- > + > +- Enabling ``IAA`` devices for platform configuration, please refer to: > + > +https://www.intel.com/content/www/us/en/content-details/780887/intel-in-memory-analytics-accelerator-intel-iaa.html > + > +- ``IAA`` device driver is ``Intel Data Accelerator Driver (idxd)``, it is > + recommended that the minimum version of Linux kernel is 5.18. > + > +- Add ``"intel_iommu=on,sm_on"`` parameter to kernel command line > + for ``SVM`` feature enabling. > + > +Here is an easy way to verify ``IAA`` device driver and ``SVM``, refer to: > + > +https://github.com/intel/idxd-config/tree/stable/test > + > +IAA Device Management > +--------------------- > + > +The number of ``IAA`` devices will vary depending on the Xeon product model. > +On a ``SPR`` server, there can be a maximum of 8 ``IAA`` devices, with up to > +4 devices per socket. > + > +By default, all ``IAA`` devices are disabled and need to be configured and > +enabled by users manually. > + > +Check the number of devices through the following command > + > +.. code-block:: shell > + > + # lspci -d 8086:0cfe > + # 6a:02.0 System peripheral: Intel Corporation Device 0cfe > + # 6f:02.0 System peripheral: Intel Corporation Device 0cfe > + # 74:02.0 System peripheral: Intel Corporation Device 0cfe > + # 79:02.0 System peripheral: Intel Corporation Device 0cfe > + # e7:02.0 System peripheral: Intel Corporation Device 0cfe > + # ec:02.0 System peripheral: Intel Corporation Device 0cfe > + # f1:02.0 System peripheral: Intel Corporation Device 0cfe > + # f6:02.0 System peripheral: Intel Corporation Device 0cfe > + > +IAA Device Configuration > +------------------------ > + > +The ``accel-config`` tool is used to enable ``IAA`` devices and configure > +``IAA`` hardware resources(work queues and engines). One ``IAA`` device > +has 8 work queues and 8 processing engines, multiple engines can be assigned > +to a work queue via ``group`` attribute. > + > +One example of configuring and enabling an ``IAA`` device. > + > +.. code-block:: shell > + > + # accel-config config-engine iax1/engine1.0 -g 0 > + # accel-config config-engine iax1/engine1.1 -g 0 > + # accel-config config-engine iax1/engine1.2 -g 0 > + # accel-config config-engine iax1/engine1.3 -g 0 > + # accel-config config-engine iax1/engine1.4 -g 0 > + # accel-config config-engine iax1/engine1.5 -g 0 > + # accel-config config-engine iax1/engine1.6 -g 0 > + # accel-config config-engine iax1/engine1.7 -g 0 > + # accel-config config-wq iax1/wq1.0 -g 0 -s 128 -p 10 -b 1 -t 128 -m shared -y user -n app1 -d user > + # accel-config enable-device iax1 > + # accel-config enable-wq iax1/wq1.0 > + > +.. note:: > + IAX is an early name for IAA > + > +- The ``IAA`` device index is 1, use ``ls -lh /sys/bus/dsa/devices/iax*`` > + command to query the ``IAA`` device index. > + > +- 8 engines and 1 work queue are configured in group 0, so all compression jobs > + submitted to this work queue can be processed by all engines at the same time. > + > +- Set work queue attributes including the work mode, work queue size and so on. > + > +- Enable the ``IAA1`` device and work queue 1.0 > + > +.. note:: > + Set work queue mode to shared mode, since ``QPL`` library only supports > + shared mode > + > +For more detailed configuration, please refer to: > + > +https://github.com/intel/idxd-config/tree/stable/Documentation/accfg > + > +IAA Resources Allocation For Migration > +-------------------------------------- > + > +There is no ``IAA`` resource configuration parameters for migration and > +``accel-config`` tool configuration cannot directly specify the ``IAA`` > +resources used for migration. > + > +``QPL`` will use all work queues that are enabled and set to shared mode, > +and use all engines assigned to the work queues with shared mode. > + > +By default, ``QPL`` will only use the local ``IAA`` device for compression > +job processing. The local ``IAA`` device means that the CPU of the job > +submission and the ``IAA`` device are on the same socket, so one CPU > +can submit the jobs to up to 4 ``IAA`` devices. > + > +Shared Virtual Memory(SVM) Introduction > +======================================= > + > +An ability for an accelerator I/O device to operate in the same virtual > +memory space of applications on host processors. It also implies the > +ability to operate from pageable memory, avoiding functional requirements > +to pin memory for DMA operations. > + > +When using ``SVM`` technology, users do not need to reserve memory for the > +``IAA`` device and perform pin memory operation. The ``IAA`` device can > +directly access data using the virtual address of the process. > + > +For more ``SVM`` technology, please refer to: > + > +https://docs.kernel.org/next/x86/sva.html > + > + > +How To Use QPL Compression In Migration > +======================================= > + > +1 - Installation of ``accel-config`` tool and ``QPL`` library We can drop "1 " and stick with: - item1 - item1.1 ... - item2 > + > + - Install ``accel-config`` tool from https://github.com/intel/idxd-config > + - Install ``QPL`` library from https://github.com/intel/qpl > + > +2 - Configure and enable ``IAA`` devices and work queues via ``accel-config`` > + > +3 - Build ``Qemu`` with ``--enable-qpl`` parameter > + > + E.g. configure --target-list=x86_64-softmmu --enable-kvm ``--enable-qpl`` > + > +4 - Start VMs with ``sudo`` command or ``root`` permission > + > + Use the ``sudo`` command or ``root`` privilege to start the source and > + destination virtual machines, since migration service needs permission > + to access ``IAA`` hardware resources. > + > +5 - Enable ``QPL`` compression during migration > + > + Set ``migrate_set_parameter multifd-compression qpl`` when migrating, the > + ``QPL`` compression does not support configuring the compression level, it > + only supports one compression level. > + > +The Difference Between QPL And ZLIB > +=================================== > + > +Although both ``QPL`` and ``ZLIB`` are based on the deflate compression > +algorithm, and ``QPL`` can support the header and tail of ``ZLIB``, ``QPL`` > +is still not fully compatible with the ``ZLIB`` compression in the migration. > + > +``QPL`` only supports 4K history buffer, and ``ZLIB`` is 32K by default. The > +``ZLIB`` compressed data that ``QPL`` may not decompress correctly and > +vice versa. > + > +``QPL`` does not support the ``Z_SYNC_FLUSH`` operation in ``ZLIB`` streaming > +compression, current ``ZLIB`` implementation uses ``Z_SYNC_FLUSH``, so each > +``multifd`` thread has a ``ZLIB`` streaming context, and all page compression > +and decompression are based on this stream. ``QPL`` cannot decompress such data > +and vice versa. > + > +The introduction for ``Z_SYNC_FLUSH``, please refer to: > + > +https://www.zlib.net/manual.html > + > +The Best Practices > +================== > + > +When the virtual machine's pages are not populated and the ``IAA`` device is > +used, I/O page faults occur, which can impact performance due to a large number > +of flush ``IOTLB`` operations. AFAIU the IOTLB issue is not expected and can be fixed, while per our discussion in the other thread, the DMA fault latency cannot. I think we can mention the possibility of IOTLB flush issue but that shouldn't be the major cause of such suggestion. Again, I think it'll be great to mention how slow a DMA page fault can be, it can be compared in migration performance difference, or just describe an average DMA page fault latency, v.s. a processor fault. That might explain why we suggest prefault the pages (comparing to a generic setup where the pages are always faulted by processors). > + > +Since the normal pages on the source side are all populated, ``IOTLB`` caused > +by I/O page fault will not occur. On the destination side, a large number > +of normal pages need to be loaded, so it is recommended to add ``-mem-prealloc`` > +parameter on the destination side. > -- > 2.39.3 >
> -----Original Message----- > From: Peter Xu <peterx@redhat.com> > Sent: Wednesday, March 27, 2024 1:58 AM > To: Liu, Yuan1 <yuan1.liu@intel.com> > Cc: farosas@suse.de; qemu-devel@nongnu.org; hao.xiang@bytedance.com; > bryan.zhang@bytedance.com; Zou, Nanhai <nanhai.zou@intel.com> > Subject: Re: [PATCH v5 1/7] docs/migration: add qpl compression feature > > On Wed, Mar 20, 2024 at 12:45:21AM +0800, Yuan Liu wrote: > > add Intel Query Processing Library (QPL) compression method > > introduction > > > > Signed-off-by: Yuan Liu <yuan1.liu@intel.com> > > Reviewed-by: Nanhai Zou <nanhai.zou@intel.com> > > --- > > docs/devel/migration/features.rst | 1 + > > docs/devel/migration/qpl-compression.rst | 231 +++++++++++++++++++++++ > > 2 files changed, 232 insertions(+) > > create mode 100644 docs/devel/migration/qpl-compression.rst > > > > diff --git a/docs/devel/migration/features.rst > b/docs/devel/migration/features.rst > > index d5ca7b86d5..bc98b65075 100644 > > --- a/docs/devel/migration/features.rst > > +++ b/docs/devel/migration/features.rst > > @@ -12,3 +12,4 @@ Migration has plenty of features to support different > use cases. > > virtio > > mapped-ram > > CPR > > + qpl-compression > > diff --git a/docs/devel/migration/qpl-compression.rst > b/docs/devel/migration/qpl-compression.rst > > new file mode 100644 > > index 0000000000..42c7969d30 > > --- /dev/null > > +++ b/docs/devel/migration/qpl-compression.rst > > @@ -0,0 +1,231 @@ > > +=============== > > +QPL Compression > > +=============== > > +The Intel Query Processing Library (Intel ``QPL``) is an open-source > library to > > +provide compression and decompression features and it is based on > deflate > > +compression algorithm (RFC 1951). > > + > > +The ``QPL`` compression relies on Intel In-Memory Analytics > Accelerator(``IAA``) > > +and Shared Virtual Memory(``SVM``) technology, they are new features > supported > > +from Intel 4th Gen Intel Xeon Scalable processors, codenamed Sapphire > Rapids > > +processor(``SPR``). > > + > > +For more ``QPL`` introduction, please refer to: > > + > > > +https://intel.github.io/qpl/documentation/introduction_docs/introduction. > html > > There're a bunch of links in this page, please consider switching all of > them to use the link formats of .rST: > > Please refer to `QPL introduction page <https://...>`_. Sure, thanks for the suggestion > > + > > +QPL Compression Framework > > +========================= > > + > > +:: > > + > > + +----------------+ +------------------+ > > + | MultiFD Service| |accel-config tool | > > + +-------+--------+ +--------+---------+ > > + | | > > + | | > > + +-------+--------+ | Setup IAA > > + | QPL library | | Resources > > + +-------+---+----+ | > > + | | | > > + | +-------------+-------+ > > + | Open IAA | > > + | Devices +-----+-----+ > > + | |idxd driver| > > + | +-----+-----+ > > + | | > > + | | > > + | +-----+-----+ > > + +-----------+IAA Devices| > > + Submit jobs +-----------+ > > + via enqcmd > > + > > + > > +Intel In-Memory Analytics Accelerator (Intel IAA) Introduction > > +================================================================ > > + > > +Intel ``IAA`` is an accelerator that has been designed to help benefit > > +in-memory databases and analytic workloads. There are three main areas > > +that Intel ``IAA`` can assist with analytics primitives (scan, filter, > etc.), > > +sparse data compression and memory tiering. > > + > > +``IAA`` Manual Documentation: > > + > > +https://www.intel.com/content/www/us/en/content-details/721858/intel- > in-memory-analytics-accelerator-architecture-specification > > + > > +IAA Device Enabling > > +------------------- > > + > > +- Enabling ``IAA`` devices for platform configuration, please refer to: > > + > > +https://www.intel.com/content/www/us/en/content-details/780887/intel- > in-memory-analytics-accelerator-intel-iaa.html > > + > > +- ``IAA`` device driver is ``Intel Data Accelerator Driver (idxd)``, it > is > > + recommended that the minimum version of Linux kernel is 5.18. > > + > > +- Add ``"intel_iommu=on,sm_on"`` parameter to kernel command line > > + for ``SVM`` feature enabling. > > + > > +Here is an easy way to verify ``IAA`` device driver and ``SVM``, refer > to: > > + > > +https://github.com/intel/idxd-config/tree/stable/test > > + > > +IAA Device Management > > +--------------------- > > + > > +The number of ``IAA`` devices will vary depending on the Xeon product > model. > > +On a ``SPR`` server, there can be a maximum of 8 ``IAA`` devices, with > up to > > +4 devices per socket. > > + > > +By default, all ``IAA`` devices are disabled and need to be configured > and > > +enabled by users manually. > > + > > +Check the number of devices through the following command > > + > > +.. code-block:: shell > > + > > + # lspci -d 8086:0cfe > > + # 6a:02.0 System peripheral: Intel Corporation Device 0cfe > > + # 6f:02.0 System peripheral: Intel Corporation Device 0cfe > > + # 74:02.0 System peripheral: Intel Corporation Device 0cfe > > + # 79:02.0 System peripheral: Intel Corporation Device 0cfe > > + # e7:02.0 System peripheral: Intel Corporation Device 0cfe > > + # ec:02.0 System peripheral: Intel Corporation Device 0cfe > > + # f1:02.0 System peripheral: Intel Corporation Device 0cfe > > + # f6:02.0 System peripheral: Intel Corporation Device 0cfe > > + > > +IAA Device Configuration > > +------------------------ > > + > > +The ``accel-config`` tool is used to enable ``IAA`` devices and > configure > > +``IAA`` hardware resources(work queues and engines). One ``IAA`` device > > +has 8 work queues and 8 processing engines, multiple engines can be > assigned > > +to a work queue via ``group`` attribute. > > + > > +One example of configuring and enabling an ``IAA`` device. > > + > > +.. code-block:: shell > > + > > + # accel-config config-engine iax1/engine1.0 -g 0 > > + # accel-config config-engine iax1/engine1.1 -g 0 > > + # accel-config config-engine iax1/engine1.2 -g 0 > > + # accel-config config-engine iax1/engine1.3 -g 0 > > + # accel-config config-engine iax1/engine1.4 -g 0 > > + # accel-config config-engine iax1/engine1.5 -g 0 > > + # accel-config config-engine iax1/engine1.6 -g 0 > > + # accel-config config-engine iax1/engine1.7 -g 0 > > + # accel-config config-wq iax1/wq1.0 -g 0 -s 128 -p 10 -b 1 -t 128 -m > shared -y user -n app1 -d user > > + # accel-config enable-device iax1 > > + # accel-config enable-wq iax1/wq1.0 > > + > > +.. note:: > > + IAX is an early name for IAA > > + > > +- The ``IAA`` device index is 1, use ``ls -lh > /sys/bus/dsa/devices/iax*`` > > + command to query the ``IAA`` device index. > > + > > +- 8 engines and 1 work queue are configured in group 0, so all > compression jobs > > + submitted to this work queue can be processed by all engines at the > same time. > > + > > +- Set work queue attributes including the work mode, work queue size > and so on. > > + > > +- Enable the ``IAA1`` device and work queue 1.0 > > + > > +.. note:: > > + Set work queue mode to shared mode, since ``QPL`` library only > supports > > + shared mode > > + > > +For more detailed configuration, please refer to: > > + > > +https://github.com/intel/idxd-config/tree/stable/Documentation/accfg > > + > > +IAA Resources Allocation For Migration > > +-------------------------------------- > > + > > +There is no ``IAA`` resource configuration parameters for migration and > > +``accel-config`` tool configuration cannot directly specify the ``IAA`` > > +resources used for migration. > > + > > +``QPL`` will use all work queues that are enabled and set to shared > mode, > > +and use all engines assigned to the work queues with shared mode. > > + > > +By default, ``QPL`` will only use the local ``IAA`` device for > compression > > +job processing. The local ``IAA`` device means that the CPU of the job > > +submission and the ``IAA`` device are on the same socket, so one CPU > > +can submit the jobs to up to 4 ``IAA`` devices. > > + > > +Shared Virtual Memory(SVM) Introduction > > +======================================= > > + > > +An ability for an accelerator I/O device to operate in the same virtual > > +memory space of applications on host processors. It also implies the > > +ability to operate from pageable memory, avoiding functional > requirements > > +to pin memory for DMA operations. > > + > > +When using ``SVM`` technology, users do not need to reserve memory for > the > > +``IAA`` device and perform pin memory operation. The ``IAA`` device can > > +directly access data using the virtual address of the process. > > + > > +For more ``SVM`` technology, please refer to: > > + > > +https://docs.kernel.org/next/x86/sva.html > > + > > + > > +How To Use QPL Compression In Migration > > +======================================= > > + > > +1 - Installation of ``accel-config`` tool and ``QPL`` library > > We can drop "1 " and stick with: > > - item1 > - item1.1 > ... > - item2 > Yes, it is better. > > + > > + - Install ``accel-config`` tool from https://github.com/intel/idxd- > config > > + - Install ``QPL`` library from https://github.com/intel/qpl > > + > > +2 - Configure and enable ``IAA`` devices and work queues via ``accel- > config`` > > + > > +3 - Build ``Qemu`` with ``--enable-qpl`` parameter > > + > > + E.g. configure --target-list=x86_64-softmmu --enable-kvm ``--enable- > qpl`` > > + > > +4 - Start VMs with ``sudo`` command or ``root`` permission > > + > > + Use the ``sudo`` command or ``root`` privilege to start the source > and > > + destination virtual machines, since migration service needs > permission > > + to access ``IAA`` hardware resources. > > + > > +5 - Enable ``QPL`` compression during migration > > + > > + Set ``migrate_set_parameter multifd-compression qpl`` when migrating, > the > > + ``QPL`` compression does not support configuring the compression > level, it > > + only supports one compression level. > > + > > +The Difference Between QPL And ZLIB > > +=================================== > > + > > +Although both ``QPL`` and ``ZLIB`` are based on the deflate compression > > +algorithm, and ``QPL`` can support the header and tail of ``ZLIB``, > ``QPL`` > > +is still not fully compatible with the ``ZLIB`` compression in the > migration. > > + > > +``QPL`` only supports 4K history buffer, and ``ZLIB`` is 32K by > default. The > > +``ZLIB`` compressed data that ``QPL`` may not decompress correctly and > > +vice versa. > > + > > +``QPL`` does not support the ``Z_SYNC_FLUSH`` operation in ``ZLIB`` > streaming > > +compression, current ``ZLIB`` implementation uses ``Z_SYNC_FLUSH``, so > each > > +``multifd`` thread has a ``ZLIB`` streaming context, and all page > compression > > +and decompression are based on this stream. ``QPL`` cannot decompress > such data > > +and vice versa. > > + > > +The introduction for ``Z_SYNC_FLUSH``, please refer to: > > + > > +https://www.zlib.net/manual.html > > + > > +The Best Practices > > +================== > > + > > +When the virtual machine's pages are not populated and the ``IAA`` > device is > > +used, I/O page faults occur, which can impact performance due to a > large number > > +of flush ``IOTLB`` operations. > > AFAIU the IOTLB issue is not expected and can be fixed, while per our > discussion in the other thread, the DMA fault latency cannot. > > I think we can mention the possibility of IOTLB flush issue but that > shouldn't be the major cause of such suggestion. Again, I think it'll be > great to mention how slow a DMA page fault can be, it can be compared in > migration performance difference, or just describe an average DMA page > fault latency, v.s. a processor fault. That might explain why we suggest > prefault the pages (comparing to a generic setup where the pages are > always > faulted by processors). Get it, I will explain this issue in depth and add some comparative data, including I/O page fault Vs. process fault performance comparison and the performance impact of I/O page fault on the entire live migration. Help developers and users better understand why -mem- prealloc needs to be added currently. > > + > > +Since the normal pages on the source side are all populated, ``IOTLB`` > caused > > +by I/O page fault will not occur. On the destination side, a large > number > > +of normal pages need to be loaded, so it is recommended to add ``-mem- > prealloc`` > > +parameter on the destination side. > > -- > > 2.39.3 > > > > -- > Peter Xu
diff --git a/docs/devel/migration/features.rst b/docs/devel/migration/features.rst index d5ca7b86d5..bc98b65075 100644 --- a/docs/devel/migration/features.rst +++ b/docs/devel/migration/features.rst @@ -12,3 +12,4 @@ Migration has plenty of features to support different use cases. virtio mapped-ram CPR + qpl-compression diff --git a/docs/devel/migration/qpl-compression.rst b/docs/devel/migration/qpl-compression.rst new file mode 100644 index 0000000000..42c7969d30 --- /dev/null +++ b/docs/devel/migration/qpl-compression.rst @@ -0,0 +1,231 @@ +=============== +QPL Compression +=============== +The Intel Query Processing Library (Intel ``QPL``) is an open-source library to +provide compression and decompression features and it is based on deflate +compression algorithm (RFC 1951). + +The ``QPL`` compression relies on Intel In-Memory Analytics Accelerator(``IAA``) +and Shared Virtual Memory(``SVM``) technology, they are new features supported +from Intel 4th Gen Intel Xeon Scalable processors, codenamed Sapphire Rapids +processor(``SPR``). + +For more ``QPL`` introduction, please refer to: + +https://intel.github.io/qpl/documentation/introduction_docs/introduction.html + +QPL Compression Framework +========================= + +:: + + +----------------+ +------------------+ + | MultiFD Service| |accel-config tool | + +-------+--------+ +--------+---------+ + | | + | | + +-------+--------+ | Setup IAA + | QPL library | | Resources + +-------+---+----+ | + | | | + | +-------------+-------+ + | Open IAA | + | Devices +-----+-----+ + | |idxd driver| + | +-----+-----+ + | | + | | + | +-----+-----+ + +-----------+IAA Devices| + Submit jobs +-----------+ + via enqcmd + + +Intel In-Memory Analytics Accelerator (Intel IAA) Introduction +================================================================ + +Intel ``IAA`` is an accelerator that has been designed to help benefit +in-memory databases and analytic workloads. There are three main areas +that Intel ``IAA`` can assist with analytics primitives (scan, filter, etc.), +sparse data compression and memory tiering. + +``IAA`` Manual Documentation: + +https://www.intel.com/content/www/us/en/content-details/721858/intel-in-memory-analytics-accelerator-architecture-specification + +IAA Device Enabling +------------------- + +- Enabling ``IAA`` devices for platform configuration, please refer to: + +https://www.intel.com/content/www/us/en/content-details/780887/intel-in-memory-analytics-accelerator-intel-iaa.html + +- ``IAA`` device driver is ``Intel Data Accelerator Driver (idxd)``, it is + recommended that the minimum version of Linux kernel is 5.18. + +- Add ``"intel_iommu=on,sm_on"`` parameter to kernel command line + for ``SVM`` feature enabling. + +Here is an easy way to verify ``IAA`` device driver and ``SVM``, refer to: + +https://github.com/intel/idxd-config/tree/stable/test + +IAA Device Management +--------------------- + +The number of ``IAA`` devices will vary depending on the Xeon product model. +On a ``SPR`` server, there can be a maximum of 8 ``IAA`` devices, with up to +4 devices per socket. + +By default, all ``IAA`` devices are disabled and need to be configured and +enabled by users manually. + +Check the number of devices through the following command + +.. code-block:: shell + + # lspci -d 8086:0cfe + # 6a:02.0 System peripheral: Intel Corporation Device 0cfe + # 6f:02.0 System peripheral: Intel Corporation Device 0cfe + # 74:02.0 System peripheral: Intel Corporation Device 0cfe + # 79:02.0 System peripheral: Intel Corporation Device 0cfe + # e7:02.0 System peripheral: Intel Corporation Device 0cfe + # ec:02.0 System peripheral: Intel Corporation Device 0cfe + # f1:02.0 System peripheral: Intel Corporation Device 0cfe + # f6:02.0 System peripheral: Intel Corporation Device 0cfe + +IAA Device Configuration +------------------------ + +The ``accel-config`` tool is used to enable ``IAA`` devices and configure +``IAA`` hardware resources(work queues and engines). One ``IAA`` device +has 8 work queues and 8 processing engines, multiple engines can be assigned +to a work queue via ``group`` attribute. + +One example of configuring and enabling an ``IAA`` device. + +.. code-block:: shell + + # accel-config config-engine iax1/engine1.0 -g 0 + # accel-config config-engine iax1/engine1.1 -g 0 + # accel-config config-engine iax1/engine1.2 -g 0 + # accel-config config-engine iax1/engine1.3 -g 0 + # accel-config config-engine iax1/engine1.4 -g 0 + # accel-config config-engine iax1/engine1.5 -g 0 + # accel-config config-engine iax1/engine1.6 -g 0 + # accel-config config-engine iax1/engine1.7 -g 0 + # accel-config config-wq iax1/wq1.0 -g 0 -s 128 -p 10 -b 1 -t 128 -m shared -y user -n app1 -d user + # accel-config enable-device iax1 + # accel-config enable-wq iax1/wq1.0 + +.. note:: + IAX is an early name for IAA + +- The ``IAA`` device index is 1, use ``ls -lh /sys/bus/dsa/devices/iax*`` + command to query the ``IAA`` device index. + +- 8 engines and 1 work queue are configured in group 0, so all compression jobs + submitted to this work queue can be processed by all engines at the same time. + +- Set work queue attributes including the work mode, work queue size and so on. + +- Enable the ``IAA1`` device and work queue 1.0 + +.. note:: + Set work queue mode to shared mode, since ``QPL`` library only supports + shared mode + +For more detailed configuration, please refer to: + +https://github.com/intel/idxd-config/tree/stable/Documentation/accfg + +IAA Resources Allocation For Migration +-------------------------------------- + +There is no ``IAA`` resource configuration parameters for migration and +``accel-config`` tool configuration cannot directly specify the ``IAA`` +resources used for migration. + +``QPL`` will use all work queues that are enabled and set to shared mode, +and use all engines assigned to the work queues with shared mode. + +By default, ``QPL`` will only use the local ``IAA`` device for compression +job processing. The local ``IAA`` device means that the CPU of the job +submission and the ``IAA`` device are on the same socket, so one CPU +can submit the jobs to up to 4 ``IAA`` devices. + +Shared Virtual Memory(SVM) Introduction +======================================= + +An ability for an accelerator I/O device to operate in the same virtual +memory space of applications on host processors. It also implies the +ability to operate from pageable memory, avoiding functional requirements +to pin memory for DMA operations. + +When using ``SVM`` technology, users do not need to reserve memory for the +``IAA`` device and perform pin memory operation. The ``IAA`` device can +directly access data using the virtual address of the process. + +For more ``SVM`` technology, please refer to: + +https://docs.kernel.org/next/x86/sva.html + + +How To Use QPL Compression In Migration +======================================= + +1 - Installation of ``accel-config`` tool and ``QPL`` library + + - Install ``accel-config`` tool from https://github.com/intel/idxd-config + - Install ``QPL`` library from https://github.com/intel/qpl + +2 - Configure and enable ``IAA`` devices and work queues via ``accel-config`` + +3 - Build ``Qemu`` with ``--enable-qpl`` parameter + + E.g. configure --target-list=x86_64-softmmu --enable-kvm ``--enable-qpl`` + +4 - Start VMs with ``sudo`` command or ``root`` permission + + Use the ``sudo`` command or ``root`` privilege to start the source and + destination virtual machines, since migration service needs permission + to access ``IAA`` hardware resources. + +5 - Enable ``QPL`` compression during migration + + Set ``migrate_set_parameter multifd-compression qpl`` when migrating, the + ``QPL`` compression does not support configuring the compression level, it + only supports one compression level. + +The Difference Between QPL And ZLIB +=================================== + +Although both ``QPL`` and ``ZLIB`` are based on the deflate compression +algorithm, and ``QPL`` can support the header and tail of ``ZLIB``, ``QPL`` +is still not fully compatible with the ``ZLIB`` compression in the migration. + +``QPL`` only supports 4K history buffer, and ``ZLIB`` is 32K by default. The +``ZLIB`` compressed data that ``QPL`` may not decompress correctly and +vice versa. + +``QPL`` does not support the ``Z_SYNC_FLUSH`` operation in ``ZLIB`` streaming +compression, current ``ZLIB`` implementation uses ``Z_SYNC_FLUSH``, so each +``multifd`` thread has a ``ZLIB`` streaming context, and all page compression +and decompression are based on this stream. ``QPL`` cannot decompress such data +and vice versa. + +The introduction for ``Z_SYNC_FLUSH``, please refer to: + +https://www.zlib.net/manual.html + +The Best Practices +================== + +When the virtual machine's pages are not populated and the ``IAA`` device is +used, I/O page faults occur, which can impact performance due to a large number +of flush ``IOTLB`` operations. + +Since the normal pages on the source side are all populated, ``IOTLB`` caused +by I/O page fault will not occur. On the destination side, a large number +of normal pages need to be loaded, so it is recommended to add ``-mem-prealloc`` +parameter on the destination side.