blob: 0a280d986731dfc67943cb6c3ec0b616a0cbdfc4 [file] [log] [blame]
Sudeep Dutt7df20f22015-04-29 05:32:28 -07001The Symmetric Communication Interface (SCIF (pronounced as skiff)) is a low
2level communications API across PCIe currently implemented for MIC. Currently
3SCIF provides inter-node communication within a single host platform, where a
4node is a MIC Coprocessor or Xeon based host. SCIF abstracts the details of
5communicating over the PCIe bus while providing an API that is symmetric
6across all the nodes in the PCIe network. An important design objective for SCIF
7is to deliver the maximum possible performance given the communication
8abilities of the hardware. SCIF has been used to implement an offload compiler
9runtime and OFED support for MPI implementations for MIC coprocessors.
10
11==== SCIF API Components ====
12The SCIF API has the following parts:
131. Connection establishment using a client server model
142. Byte stream messaging intended for short messages
153. Node enumeration to determine online nodes
164. Poll semantics for detection of incoming connections and messages
175. Memory registration to pin down pages
186. Remote memory mapping for low latency CPU accesses via mmap
197. Remote DMA (RDMA) for high bandwidth DMA transfers
208. Fence APIs for RDMA synchronization
21
22SCIF exposes the notion of a connection which can be used by peer processes on
23nodes in a SCIF PCIe "network" to share memory "windows" and to communicate. A
24process in a SCIF node initiates a SCIF connection to a peer process on a
25different node via a SCIF "endpoint". SCIF endpoints support messaging APIs
26which are similar to connection oriented socket APIs. Connected SCIF endpoints
27can also register local memory which is followed by data transfer using either
28DMA, CPU copies or remote memory mapping via mmap. SCIF supports both user and
29kernel mode clients which are functionally equivalent.
30
31==== SCIF Performance for MIC ====
32DMA bandwidth comparison between the TCP (over ethernet over PCIe) stack versus
33SCIF shows the performance advantages of SCIF for HPC applications and runtimes.
34
35 Comparison of TCP and SCIF based BW
36
37 Throughput (GB/sec)
38 8 + PCIe Bandwidth ******
39 + TCP ######
40 7 + ************************************** SCIF %%%%%%
41 | %%%%%%%%%%%%%%%%%%%
42 6 + %%%%
43 | %%
44 | %%%
45 5 + %%
46 | %%
47 4 + %%
48 | %%
49 3 + %%
50 | %
51 2 + %%
52 | %%
53 | %
54 1 +
55 + ######################################
56 0 +++---+++--+--+-+--+--+-++-+--+-++-+--+-++-+-
57 1 10 100 1000 10000 100000
58 Transfer Size (KBytes)
59
60SCIF allows memory sharing via mmap(..) between processes on different PCIe
61nodes and thus provides bare-metal PCIe latency. The round trip SCIF mmap
62latency from the host to an x100 MIC for an 8 byte message is 0.44 usecs.
63
64SCIF has a user space library which is a thin IOCTL wrapper providing a user
65space API similar to the kernel API in scif.h. The SCIF user space library
66is distributed @ https://software.intel.com/en-us/mic-developer
67
68Here is some pseudo code for an example of how two applications on two PCIe
69nodes would typically use the SCIF API:
70
71Process A (on node A) Process B (on node B)
72
73/* get online node information */
74scif_get_node_ids(..) scif_get_node_ids(..)
75scif_open(..) scif_open(..)
76scif_bind(..) scif_bind(..)
77scif_listen(..)
78scif_accept(..) scif_connect(..)
79/* SCIF connection established */
80
81/* Send and receive short messages */
82scif_send(..)/scif_recv(..) scif_send(..)/scif_recv(..)
83
84/* Register memory */
85scif_register(..) scif_register(..)
86
87/* RDMA */
88scif_readfrom(..)/scif_writeto(..) scif_readfrom(..)/scif_writeto(..)
89
90/* Fence DMAs */
91scif_fence_signal(..) scif_fence_signal(..)
92
93mmap(..) mmap(..)
94
95/* Access remote registered memory */
96
97/* Close the endpoints */
98scif_close(..) scif_close(..)