diff options
| -rw-r--r-- | Documentation/infiniband/tag_matching.txt | 64 |
1 files changed, 64 insertions, 0 deletions
diff --git a/Documentation/infiniband/tag_matching.txt b/Documentation/infiniband/tag_matching.txt new file mode 100644 index 000000000000..d2a3bf819226 --- /dev/null +++ b/Documentation/infiniband/tag_matching.txt @@ -0,0 +1,64 @@ +Tag matching logic + +The MPI standard defines a set of rules, known as tag-matching, for matching +source send operations to destination receives. The following parameters must +match the following source and destination parameters: +* Communicator +* User tag - wild card may be specified by the receiver +* Source rank – wild car may be specified by the receiver +* Destination rank – wild +The ordering rules require that when more than one pair of send and receive +message envelopes may match, the pair that includes the earliest posted-send +and the earliest posted-receive is the pair that must be used to satisfy the +matching operation. However, this doesn’t imply that tags are consumed in +the order they are created, e.g., a later generated tag may be consumed, if +earlier tags can’t be used to satisfy the matching rules. + +When a message is sent from the sender to the receiver, the communication +library may attempt to process the operation either after or before the +corresponding matching receive is posted. If a matching receive is posted, +this is an expected message, otherwise it is called an unexpected message. +Implementations frequently use different matching schemes for these two +different matching instances. + +To keep MPI library memory footprint down, MPI implementations typically use +two different protocols for this purpose: + +1. The Eager protocol- the complete message is sent when the send is +processed by the sender. A completion send is received in the send_cq +notifying that the buffer can be reused. + +2. The Rendezvous Protocol - the sender sends the tag-matching header, +and perhaps a portion of data when first notifying the receiver. When the +corresponding buffer is posted, the responder will use the information from +the header to initiate an RDMA READ operation directly to the matching buffer. +A fin message needs to be received in order for the buffer to be reused. + +Tag matching implementation + +There are two types of matching objects used, the posted receive list and the +unexpected message list. The application posts receive buffers through calls +to the MPI receive routines in the posted receive list and posts send messages +using the MPI send routines. The head of the posted receive list may be +maintained by the hardware, with the software expected to shadow this list. + +When send is initiated and arrives at the receive side, if there is no +pre-posted receive for this arriving message, it is passed to the software and +placed in the unexpected message list. Otherwise the match is processed, +including rendezvous processing, if appropriate, delivering the data to the +specified receive buffer. This allows overlapping receive-side MPI tag +matching with computation. + +When a receive-message is posted, the communication library will first check +the software unexpected message list for a matching receive. If a match is +found, data is delivered to the user buffer, using a software controlled +protocol. The UCX implementation uses either an eager or rendezvous protocol, +depending on data size. If no match is found, the entire pre-posted receive +list is maintained by the hardware, and there is space to add one more +pre-posted receive to this list, this receive is passed to the hardware. +Software is expected to shadow this list, to help with processing MPI cancel +operations. In addition, because hardware and software are not expected to be +tightly synchronized with respect to the tag-matching operation, this shadow +list is used to detect the case that a pre-posted receive is passed to the +hardware, as the matching unexpected message is being passed from the hardware +to the software. |