RDMA send performance bottleneck due to single send buffer
RDMA send performance bottleneck due to single send buffer: a task in LegoFlow-SWE (Harbor dataset). When using coro io::async write() with ib socket t for InfiniBand (ibverbs) RDMA communication, only one send buffer is available at a time. This serializes send operations, preventing overlapping…
The task
When using `coro_io::async_write()` with `ib_socket_t` for InfiniBand (ibverbs) RDMA communication, only one send buffer is available at a time. This serializes send operations, preventing overlapping of multiple outstanding sends and significantly limiting throughput under high-load scenarios.
Part of Lego-X/LegoFlow-SWE.