diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-05-06 01:02:30 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-05-06 01:02:30 +0000 |
commit | 76cb841cb886eef6b3bee341a2266c76578724ad (patch) | |
tree | f5892e5ba6cc11949952a6ce4ecbe6d516d6ce58 /Documentation/mic | |
parent | Initial commit. (diff) | |
download | linux-76cb841cb886eef6b3bee341a2266c76578724ad.tar.xz linux-76cb841cb886eef6b3bee341a2266c76578724ad.zip |
Adding upstream version 4.19.249.upstream/4.19.249
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'Documentation/mic')
-rw-r--r-- | Documentation/mic/mic_overview.txt | 81 | ||||
-rw-r--r-- | Documentation/mic/scif_overview.txt | 98 |
2 files changed, 179 insertions, 0 deletions
diff --git a/Documentation/mic/mic_overview.txt b/Documentation/mic/mic_overview.txt new file mode 100644 index 000000000..074adbdf8 --- /dev/null +++ b/Documentation/mic/mic_overview.txt @@ -0,0 +1,81 @@ +An Intel MIC X100 device is a PCIe form factor add-in coprocessor +card based on the Intel Many Integrated Core (MIC) architecture +that runs a Linux OS. It is a PCIe endpoint in a platform and therefore +implements the three required standard address spaces i.e. configuration, +memory and I/O. The host OS loads a device driver as is typical for +PCIe devices. The card itself runs a bootstrap after reset that +transfers control to the card OS downloaded from the host driver. The +host driver supports OSPM suspend and resume operations. It shuts down +the card during suspend and reboots the card OS during resume. +The card OS as shipped by Intel is a Linux kernel with modifications +for the X100 devices. + +Since it is a PCIe card, it does not have the ability to host hardware +devices for networking, storage and console. We provide these devices +on X100 coprocessors thus enabling a self-bootable equivalent +environment for applications. A key benefit of our solution is that it +leverages the standard virtio framework for network, disk and console +devices, though in our case the virtio framework is used across a PCIe +bus. A Virtio Over PCIe (VOP) driver allows creating user space +backends or devices on the host which are used to probe virtio drivers +for these devices on the MIC card. The existing VRINGH infrastructure +in the kernel is used to access virtio rings from the host. The card +VOP driver allows card virtio drivers to communicate with their user +space backends on the host via a device page. Ring 3 apps on the host +can add, remove and configure virtio devices. A thin MIC specific +virtio_config_ops is implemented which is borrowed heavily from +previous similar implementations in lguest and s390. + +MIC PCIe card has a dma controller with 8 channels. These channels are +shared between the host s/w and the card s/w. 0 to 3 are used by host +and 4 to 7 by card. As the dma device doesn't show up as PCIe device, +a virtual bus called mic bus is created and virtual dma devices are +created on it by the host/card drivers. On host the channels are private +and used only by the host driver to transfer data for the virtio devices. + +The Symmetric Communication Interface (SCIF (pronounced as skiff)) is a +low level communications API across PCIe currently implemented for MIC. +More details are available at scif_overview.txt. + +The Coprocessor State Management (COSM) driver on the host allows for +boot, shutdown and reset of Intel MIC devices. It communicates with a COSM +"client" driver on the MIC cards over SCIF to perform these functions. + +Here is a block diagram of the various components described above. The +virtio backends are situated on the host rather than the card given better +single threaded performance for the host compared to MIC, the ability of +the host to initiate DMA's to/from the card using the MIC DMA engine and +the fact that the virtio block storage backend can only be on the host. + + +----------+ | +----------+ + | Card OS | | | Host OS | + +----------+ | +----------+ + | + +-------+ +--------+ +------+ | +---------+ +--------+ +--------+ + | Virtio| |Virtio | |Virtio| | |Virtio | |Virtio | |Virtio | + | Net | |Console | |Block | | |Net | |Console | |Block | + | Driver| |Driver | |Driver| | |backend | |backend | |backend | + +---+---+ +---+----+ +--+---+ | +---------+ +----+---+ +--------+ + | | | | | | | + | | | |User | | | + | | | |------|------------|--+------|------- + +---------+---------+ |Kernel | + | | | + +---------+ +---+----+ +------+ | +------+ +------+ +--+---+ +-------+ + |MIC DMA | | VOP | | SCIF | | | SCIF | | COSM | | VOP | |MIC DMA| + +---+-----+ +---+----+ +--+---+ | +--+---+ +--+---+ +------+ +----+--+ + | | | | | | | + +---+-----+ +---+----+ +--+---+ | +--+---+ +--+---+ +------+ +----+--+ + |MIC | | VOP | |SCIF | | |SCIF | | COSM | | VOP | | MIC | + |HW Bus | | HW Bus| |HW Bus| | |HW Bus| | Bus | |HW Bus| |HW Bus | + +---------+ +--------+ +--+---+ | +--+---+ +------+ +------+ +-------+ + | | | | | | | + | +-----------+--+ | | | +---------------+ | + | |Intel MIC | | | | |Intel MIC | | + | |Card Driver | | | | |Host Driver | | + +---+--------------+------+ | +----+---------------+-----+ + | | | + +-------------------------------------------------------------+ + | | + | PCIe Bus | + +-------------------------------------------------------------+ diff --git a/Documentation/mic/scif_overview.txt b/Documentation/mic/scif_overview.txt new file mode 100644 index 000000000..0a280d986 --- /dev/null +++ b/Documentation/mic/scif_overview.txt @@ -0,0 +1,98 @@ +The Symmetric Communication Interface (SCIF (pronounced as skiff)) is a low +level communications API across PCIe currently implemented for MIC. Currently +SCIF provides inter-node communication within a single host platform, where a +node is a MIC Coprocessor or Xeon based host. SCIF abstracts the details of +communicating over the PCIe bus while providing an API that is symmetric +across all the nodes in the PCIe network. An important design objective for SCIF +is to deliver the maximum possible performance given the communication +abilities of the hardware. SCIF has been used to implement an offload compiler +runtime and OFED support for MPI implementations for MIC coprocessors. + +==== SCIF API Components ==== +The SCIF API has the following parts: +1. Connection establishment using a client server model +2. Byte stream messaging intended for short messages +3. Node enumeration to determine online nodes +4. Poll semantics for detection of incoming connections and messages +5. Memory registration to pin down pages +6. Remote memory mapping for low latency CPU accesses via mmap +7. Remote DMA (RDMA) for high bandwidth DMA transfers +8. Fence APIs for RDMA synchronization + +SCIF exposes the notion of a connection which can be used by peer processes on +nodes in a SCIF PCIe "network" to share memory "windows" and to communicate. A +process in a SCIF node initiates a SCIF connection to a peer process on a +different node via a SCIF "endpoint". SCIF endpoints support messaging APIs +which are similar to connection oriented socket APIs. Connected SCIF endpoints +can also register local memory which is followed by data transfer using either +DMA, CPU copies or remote memory mapping via mmap. SCIF supports both user and +kernel mode clients which are functionally equivalent. + +==== SCIF Performance for MIC ==== +DMA bandwidth comparison between the TCP (over ethernet over PCIe) stack versus +SCIF shows the performance advantages of SCIF for HPC applications and runtimes. + + Comparison of TCP and SCIF based BW + + Throughput (GB/sec) + 8 + PCIe Bandwidth ****** + + TCP ###### + 7 + ************************************** SCIF %%%%%% + | %%%%%%%%%%%%%%%%%%% + 6 + %%%% + | %% + | %%% + 5 + %% + | %% + 4 + %% + | %% + 3 + %% + | % + 2 + %% + | %% + | % + 1 + + + ###################################### + 0 +++---+++--+--+-+--+--+-++-+--+-++-+--+-++-+- + 1 10 100 1000 10000 100000 + Transfer Size (KBytes) + +SCIF allows memory sharing via mmap(..) between processes on different PCIe +nodes and thus provides bare-metal PCIe latency. The round trip SCIF mmap +latency from the host to an x100 MIC for an 8 byte message is 0.44 usecs. + +SCIF has a user space library which is a thin IOCTL wrapper providing a user +space API similar to the kernel API in scif.h. The SCIF user space library +is distributed @ https://software.intel.com/en-us/mic-developer + +Here is some pseudo code for an example of how two applications on two PCIe +nodes would typically use the SCIF API: + +Process A (on node A) Process B (on node B) + +/* get online node information */ +scif_get_node_ids(..) scif_get_node_ids(..) +scif_open(..) scif_open(..) +scif_bind(..) scif_bind(..) +scif_listen(..) +scif_accept(..) scif_connect(..) +/* SCIF connection established */ + +/* Send and receive short messages */ +scif_send(..)/scif_recv(..) scif_send(..)/scif_recv(..) + +/* Register memory */ +scif_register(..) scif_register(..) + +/* RDMA */ +scif_readfrom(..)/scif_writeto(..) scif_readfrom(..)/scif_writeto(..) + +/* Fence DMAs */ +scif_fence_signal(..) scif_fence_signal(..) + +mmap(..) mmap(..) + +/* Access remote registered memory */ + +/* Close the endpoints */ +scif_close(..) scif_close(..) |