Re: [NFS] [CFT] Improved RPC congestion handling for 2.4.0 (and 2.2.18)

[Trond Myklebust]

>>>>> " " == H J Lu <[e-mail]> writes:

     > I got a report which indicates it may not be a good idea,
     > especially for UDP. Suppose you have a lousy LAN or NFS UDP
     > server for whatever reason, some NFS/UDP packets may get lost
     > very easily while a ping request may get through. In that case,
     > the rpc ping may slow down the NFS client over UDP
     > significantly.

Hi HJ,

Could you clarify this? Don't forget that we only send the ping after
a major timeout (usually after 3 or more resends).

IOW: If the ping gets through, then it'll have cost us 1 RPC request,
which is hardly a major contribution when talking about timescales of
the order of 5 seconds which is what that major timeout will have cost
(Don't forget that RPC timeout values increase geometrically).

I'm not touting this as a cure for incorrect rsize/wsize settings. I
use it rather to avoid false 'server is not responding' due to server
congestion. On the current RPC code, the latter can occur sometimes
even with a loopback NFS mount when, say, kflushd is busy hogging the
disk...

Cheers,
  Trond