ODP is slower than regular RDMA even with memory copy overhead
Example numbers:
- 3950000 random read iops without ODP vs 240000 iops with ODP
- 1447000 random write iops without ODP vs 101000 iops with ODP
Reference: https://tkygtr6.github.io/pub/ISPASS21_slides.pdf
129K to leave extra space for the header
The problem with 8x 1M buffers is that the following happens with,
for example, 2 OSDs and 4M T1Q1 write:
- Server posts 8 receives
- Client posts 8 sends
- WRs are processed by the RDMA stack, but the OSD doesn't have the time
to handle them and doesn't refill buffers
- Client posts 1 more send
- RNR retransmission happens and performance drops to zero
Overall it seems that RDMA support should be reworked to use real 'RDMA'
operations i.e. operations writing into remote memory. This has an
additional advantage of avoiding a copy at the receive side of the OSD.
This is the simplest and, as usual, the best implementation :)
100% zero-copy implementation is also possible (see rdma-zerocopy branch),
but it requires to create A LOT of queues (~128 per client) to use QPN as a 'tag'
because of the lack of receive tags and the server may simply run out of queues.
Hardware limit is 262144 on Mellanox ConnectX-4 which amounts to only 2048
'connections' per host. And even with that amount of queues it's still less optimal
than the non-zerocopy one.
In fact, newest hardware like Mellanox ConnectX-5 does have Tag Matching
support, but it's still unsuitable for us because it doesn't support scatter/gather
(tm_caps.max_sge=1).
Basic naive implementation works, but it's highly non-optimal as
RNR retransmissions occur all the time. RDMA expects the receiver
to always have place for incoming WRs...
The bug reproduced if fio was temporarily stopped with SIGSTOP
during write test and then resumed after 10 seconds. In this case
"pings" were failed for all clients and fio process crashed with
'use-after-free' in keepalive_timer. It happened because it called
stop_client while having a live iterator to the map.
Previous implementation didn't respect write ordering and could lead
to corrupted data when restarting writes after an OSD outage
Also rework cluster_client queueing logic and add tests for it to verify the correct behaviour
Vitaliy Filippov