DaRPC registers only one memory region and creates single send and receive buffers (for post send and post receive) by using slices within this memory region (instead of creating a separate memory registration per send or receive buffer).
Advantages:
Higher locality when allocating a big block
Better for caching and TLB
Use less memory registration resources on the RDMA card
The change improves DaRPC IOPS in a 8 core server setup from
2.4Mio IOPS to 3.6Mio IOPS.
This is a first step in improving DaRPC performance.
DaRPC registers only one memory region and creates single send and receive buffers (for post send and post receive) by using slices within this memory region (instead of creating a separate memory registration per send or receive buffer).
Advantages:
The change improves DaRPC IOPS in a 8 core server setup from 2.4Mio IOPS to 3.6Mio IOPS.
This is a first step in improving DaRPC performance.
Please let me know your opinion.
Thanks Adrian