openmpi/opal/mca/btl/usnic/README.md

Design notes on usnic BTL

nomenclature

• fragment - something the PML asks us to send or put, any size
• segment - something we can put on the wire in a single packet
• chunk - a piece of a fragment that fits into one segment

a segment can contain either an entire fragment or a chunk of a fragment

each segment and fragment has associated descriptor.

Each segment data structure has a block of registered memory associated with it which matches MTU for that segment

• ACK - acks get special small segments with only enough memory for an ACK

• non-ACK segments always have a parent fragment

• fragments are either large (> MTU) or small (<= MTU)

• a small fragment has a segment descriptor embedded within it since it always needs exactly one.

• a large fragment has no permanently associated segments, but allocates them as needed.

channels

A channel is a queue pair with an associated completion queue each channel has its own MTU and r/w queue entry counts

There are 2 channels, command and data:

• command queue is generally for higher priority fragments
• data queue is for standard data traffic
• command queue should possibly be called "priority" queue

command queue is shorter and has a smaller MTU that the data queue. this makes the command queue a lot faster than the data queue, so we hijack it for sending very small fragments (<= tiny_mtu, currently 768 bytes)

command queue is used for ACKs and tiny fragments. data queue is used for everything else.

PML fragments marked priority should perhaps use command queue

sending

Normally, all send requests are simply enqueued and then actually posted to the NIC by the routine opal_btl_usnic_module_progress_sends(). "fastpath" tiny sends are the exception.

Each module maintains a queue of endpoints that are ready to send. An endpoint is ready to send if all of the following are met:

1. the endpoint has fragments to send
2. the endpoint has send credits
3. the endpoint's send window is "open" (not full of un-ACKed segments)

Each module also maintains a list of segments that need to be retransmitted. Note that the list of pending retrans is per-module, not per-endpoint.

Send progression first posts any pending retransmissions, always using the data channel. (reason is that if we start getting heavy congestion and there are lots of retransmits, it becomes more important than ever to prioritize ACKs, clogging command channel with retrans data makes things worse, not better)

Next, progression loops sending segments to the endpoint at the top of the endpoints_with_sends queue. When an endpoint exhausts its send credits or fills its send window or runs out of segments to send, it removes itself from the endpoint_with_sends list. Any pending ACKs will be picked up and piggy-backed on these sends.

Finally, any endpoints that still need ACKs whose timer has expired will be sent explicit ACK packets.

fragment sending

The middle part of the progression loop handles both small (single-segment) and large (multi-segment) sends.

For small fragments, the verbs descriptor within the embedded segment is updated with length, BTL header is updated, then we call opal_btl_usnic_endpoint_send_segment() to send the segment. After posting, we make a PML callback if needed.

For large fragments, a little more is needed. segments froma large fragment have a slightly larger BTL header which contains a fragment ID, and offset, and a size. The fragment ID is allocated when the first chunk the fragment is sent. A segment gets allocated, next blob of data is copied into this segment, segment is posted. If last chunk of fragment sent, perform callback if needed, then remove fragment from endpoint send queue.

opal_btl_usnic_endpoint_send_segment()

This is common posting code for large or small segments. It assigns a sequence number to a segment, checks for an ACK to piggy-back, posts the segment to the NIC, and then starts the retransmit timer by checking the segment into hotel. Send credits are consumed here.

send dataflow

PML control messages with no user data are sent via:

• desc = usnic_alloc(size)
• usnic_send(desc)

user messages less than eager limit and 1st part of larger

messages are sent via:

• desc = usnic_prepare_src(convertor, size)
• usnic_send(desc)

larger msgs:

• desc = usnic_prepare_src(convertor, size)
• usnic_put(desc)

usnic_alloc() currently asserts the length is "small", allocates and fills in a small fragment. src pointer will point to start of associated registered mem + sizeof BTL header, and PML will put its data there.

usnic_prepare_src() allocated either a large or small fragment based on size The fragment descriptor is filled in to have 2 SG entries, 1st pointing to place where PML should construct its header. If the data convertor says data is contiguous, 2nd SG entry points to user buffer, else it is null and sf_convertor is filled in with address of convertor.

usnic_send()

If the fragment being sent is small enough, has contiguous data, and "very few" command queue send WQEs have been consumed, usnic_send() does a fastpath send. This means it posts the segment immediately to the NIC with INLINE flag set.

If all of the conditions for fastpath send are not met, and this is a small fragment, the user data is copied into the associated registered memory at this time and the SG list in the descriptor is collapsed to one entry.

After the checks above are done, the fragment is enqueued to be sent via opal_btl_usnic_endpoint_enqueue_frag()

usnic_put()

Do a fast version of what happens in prepare_src() (can take shortcuts because we know it will always be a contiguous buffer / no convertor needed). PML gives us the destination address, which we save on the fragment (which is the sentinel value that the underlying engine uses to know that this is a PUT and not a SEND), and the fragment is enqueued for processing.

opal_btl_usnic_endpoint_enqueue_frag()

This appends the fragment to the "to be sent" list of the endpoint and conditionally adds the endpoint to the list of endpoints with data to send via opal_btl_usnic_check_rts()

receive dataflow

BTL packets has one of 3 types in header: frag, chunk, or ack.

• A frag packet is a full PML fragment.

• A chunk packet is a piece of a fragment that needs to be reassembled.

• An ack packet is header only with a sequence number being ACKed.

• Both frag and chunk packets go through some of the same processing.

• Both may carry piggy-backed ACKs which may need to be processed.

• Both have sequence numbers which must be processed and may result in dropping the packet and/or queueing an ACK to the sender.

frag packets may be either regular PML fragments or PUT segments. If the "put_addr" field of the BTL header is set, this is a PUT and the data is copied directly to the user buffer. If this field is NULL, the segment is passed up to the PML. The PML is expected to do everything it needs with this packet in the callback, including copying data out if needed. Once the callback is complete, the receive buffer is recycled.

chunk packets are parts of a larger fragment. If an active fragment receive for the matching fragment ID cannot be found, and new fragment info descriptor is allocated. If this is not a PUT (put_addr == NULL), we malloc() data to reassemble the fragment into. Each subsequent chunk is copied either into this reassembly buffer or directly into user memory. When the last chunk of a fragment arrives, a PML callback is made for non-PUTs, then the fragment info descriptor is released.

fast receive optimization

In order to optimize latency of small packets, the component progress routine implements a fast path for receives. If the first completion is a receive on the priority queue, then it is handled by a routine called opal_btl_usnic_recv_fast() which does nothing but validates that the packet is OK to be received (sequence number OK and not a DUP) and then delivers it to the PML. This packet is recorded in the channel structure, and all bookeeping for the packet is deferred until the next time component_progress is called again.

This fast path cannot be taken every time we pass through component_progress because there will be other completions that need processing, and the receive bookeeping for one fast receive must be complete before allowing another fast receive to occur, as only one recv segment can be saved for deferred processing at a time. This is handled by maintaining a variable in opal_btl_usnic_recv_fast() called fastpath_ok which is set to false every time the fastpath is taken. A call into the regular progress routine will set this flag back to true.

reliability:

• every packet has sequence #

• each endpoint has a "send window" , currently 4096 entries.

• once a segment is sent, it is saved in window array until ACK is received

• ACKs acknowledge all packets <= specified sequence #

• rcvr only ACKs a sequence # when all packets up to that sequence have arrived

• each pkt has dflt retrans timer of 100ms

• packet will be scheduled for retrans if timer expires

Once a segment is sent, it always has its retransmit timer started. This is accomplished by opal_hotel_checkin(). Any time a segment is posted to the NIC for retransmit, it is checked out of the hotel (timer stopped). So, a send segment is always in one of 4 states:

• on free list, unallocated
• on endpoint to-send list in the case of segment associated with small fragment
• posted to NIC and in hotel awaiting ACK
• on module re-send list awaiting retransmission

rcvr:

• if a pkt with seq >= expected seq is received, schedule ack of largest in-order sequence received if not already scheduled. dflt time is 50us
• if a packet with seq < expected seq arrives, we send an ACK immediately, as this indicates a lost ACK

sender:

• duplicate ACK triggers immediate retrans if one is not pending for that segment

Reordering induced by two queues and piggy-backing:

ACKs can be reordered-

• not an issue at all, old ACKs are simply ignored

Sends can be reordered-

• (small send can jump far ahead of large sends)
• large send followed by lots of small sends could trigger many retrans of the large sends. smalls would have to be paced pretty precisely to keep command queue empty enough and also beat out the large sends. send credits limit how many larges can be queued on the sender, but there could be many on the receiver

RDMA emulation

We emulate the RDMA PUT because it's more efficient than regular send: it allows the receive to copy directly to the target buffer (vs. making an intermediate copy out of the bounce buffer).

It would actually be better to morph this PUT into a GET -- GET would be slightly more efficient. In short, when the target requests the actual RDMA data, with PUT, the request has to go up to the PML, which will then invoke PUT on the source's BTL module. With GET, the target issues the GET, and the source BTL module can reply without needing to go up the stack to the PML.

Once we start supporting RDMA in hardware:

• we need to provide module.btl_register_mem and module.btl_deregister_mem functions (see openib for an example)
• we need to put something meaningful in btl_usnic_frag.h:mca_btl_base_registration_handle_t.
• we need to set module.btl_registration_handle_size to sizeof(struct mca_btl_base_registration_handle_t).
• module.btl_put / module.btl_get will receive the mca_btl_base_registration_handle_t from the peer as a cookie.

Also, module.btl_put / module.btl_get do not need to make descriptors (this was an optimization added in BTL 3.0). They are now called with enough information to do whatever they need to do. module.btl_put still makes a descriptor and submits it to the usnic sending engine so as to utilize a common infrastructure for send and put.

But it doesn't necessarily have to be that way -- we could optimize out the use of the descriptors. Have not investigated how easy/hard that would be.

libfabric abstractions:

• fi_fabric: corresponds to a VIC PF
• fi_domain: corresponds to a VIC VF
• fi_endpoint: resources inside the VIC VF (basically a QP)

MPI_THREAD_MULTIPLE support

In order to make usnic btl thread-safe, the mutex locks are issued to protect the critical path. ie; libfabric routines, book keeping, etc.

The said lock is btl_usnic_lock. It is a RECURSIVE lock, meaning that the same thread can take the lock again even if it already has the lock to allow the callback function to post another segment right away if we know that the current segment is completed inline. (So we can call send in send without deadlocking)

These two functions taking care of hotel checkin/checkout and we have to protect that part. So we take the mutex lock before we enter the function.

• opal_btl_usnic_check_rts()
• opal_btl_usnic_handle_ack()

We also have to protect the call to libfabric routines

• opal_btl_usnic_endpoint_send_segment() (fi_send)
• opal_btl_usnic_recv_call() (fi_recvmsg)

have to be protected as well.

Also cclient connection checking (opal_btl_usnic_connectivity_ping) has to be protected. This happens only in the beginning but cclient communicate with cagent through opal_fd_read/write() and if two or more clients do opal_fd_write() at the same time, the data might be corrupt.

With this concept, many functions in btl/usnic that make calls to the listed functions are protected by OPAL_THREAD_LOCK macro which will only be active if the user specify MPI_Init_thread() with MPI_THREAD_MULTIPLE support.