[LU-16011] LNet selftest should use a preallocated bulk on the server side.

• Type: Improvement
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.16.0
• Assignee: Alexey Lyashkov
• Reporter: Alexey Lyashkov

Description

Cray testing have show an LST limited with 35Gb/s 1M bulk transfers with 3 rail server and 16 HDR clients. During profile, I found a LST is cpu bound with page cache operations, because server side don't have an preallocated bulk's for transfers.
Once preallocation done - server able to achieve a wire speed - 73 Gb/s with very low cpu power.

LNet Bandwidth of servers]
[R] Avg: 11.28    MB/s  Min: 11.28    MB/s  Max: 11.28    MB/s
[W] Avg: 73949.30 MB/s  Min: 73949.30 MB/s  Max: 73949.30 MB/s