[ofa-general] RDS flow control

[Talpey, Thomas]

At 04:45 PM 5/20/2008, Jon Mason wrote:
>With proper flow control, there should no longer be a need for 
>rnr_retry (as there
>should always be a posted recv buffer waiting for the incoming data).  I did a
>quick test and removed it and everything seemed to be happy on my 
>rds-stress run.

I'd be interested in any extended load-testing of operation with
rnr_retry==0 that you might be able to do. The NFS/RDMA client
sets it to zero, for the same reason (the rpcrdma protocol exchanges
credits).

But at the NFS Connectathon last week we were seeing spontaneous
connection loss, that went away when we set rnr_retry to 7 (infinity).
However, it also did not appear when it was set to 1, and later we
were able to pass again at zero.

Very strange, I'm still trying to figure if it's an upper layer issue
or some lower layer timing quirk. The switch we were using there
was a bit flaky.

Tom.