UCS: stats performance enhancements (contig. allocation)

[alex--m]

Signed-off-by: Alex Margolin [email protected]

What

Make both the stats nodes and the counters are more or less contiguous in (virtual) memory.

Why ?

An effort to reduce the average latency of accessing stats, and also to put all the counters in a single buffer so it could be RDMA-ed instead of UDP-ed.

How ?

Both filter and regular stats nodes are allocated from a memory pool (the first chunk should cover common UCX initialization), and the counters are allocated in a flexible array (so there's one buffer containing all of them).

[swx-jenkins3]

Can one of the admins verify this patch?

[yosefe]

ok to test

[shuki-zanyovka]

ok to test

Hi @yosefe ,

Can anyone review this PR?

Regards and thanks, Shuki