Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

rdma: add separate bounce buffer freelist for data (eager) messages #614

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

rdma: add separate bounce buffer freelist for data (eager) messages #614

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
6913bd4
fix(rdma): send close message on ctrl rail
rauteric Jan 10, 2025
5c7b4cc
rdma: add `rx_buff_req_alloc` method to ep_rail
rauteric Jan 16, 2025
6748fdb
rdma: add separate request types for eager/ctrl rx buffers
rauteric Jan 9, 2025
ac8a54a
rdma: use separate freelists for ctrl and eager rx buffers
rauteric Jan 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 16 additions & 6 deletions include/nccl_ofi_rdma.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,8 +80,10 @@ typedef enum nccl_net_ofi_rdma_req_type {
NCCL_OFI_RDMA_RECV_SEGMS,
/* Eager local copy request. Subrequest of NCCL_OFI_RDMA_RECV */
NCCL_OFI_RDMA_EAGER_COPY,
/* Rx buff post request */
NCCL_OFI_RDMA_RX_BUFF,
/* Ctrl rx buff post request */
NCCL_OFI_RDMA_CTRL_RX_BUFF,
/* Eager rx buff post request */
NCCL_OFI_RDMA_EAGER_RX_BUFF,
/* Flush request */
NCCL_OFI_RDMA_FLUSH,
/* Connect message send request */
Expand Down Expand Up @@ -690,6 +692,10 @@ struct nccl_net_ofi_ep_rail {
size_t max_rx_buff_posted;
/* Mutex for rx buffer operations */
pthread_mutex_t rx_buff_mutex;

/* Allocate a receive buffer request for this rail (eager or ctrl) */
nccl_net_ofi_rdma_req_t* (*rx_buff_req_alloc)(nccl_net_ofi_rdma_ep_t *ep,
rauteric marked this conversation as resolved.
Show resolved Hide resolved
nccl_net_ofi_ep_rail_t *rail);
};

/*
Expand Down Expand Up @@ -728,12 +734,16 @@ struct nccl_net_ofi_rdma_ep {
/* Pending requests queue */
nccl_ofi_deque_t *pending_reqs_queue;

/* Free list of rx buffers */
nccl_ofi_freelist_t *rx_buff_fl;
/* Free list of ctrl rx buffers */
nccl_ofi_freelist_t *ctrl_rx_buff_fl;
/* Free list of eager rx buffers */
nccl_ofi_freelist_t *eager_rx_buff_fl;
/* Free list of rx buffer requests */
nccl_ofi_freelist_t *rx_buff_reqs_fl;
/* Size of rx buffers */
size_t rx_buff_size;
/* Size of ctrl rx buffers */
size_t ctrl_rx_buff_size;
/* Size of eager rx buffers */
size_t eager_rx_buff_size;

/* true if the current endpoint is a endpoint_per_communicator
receive communicator */
Expand Down
Loading
Loading