NFS performance depends on low latency reception of small TCP segments

[Rick Macklem]

Hi,

I have seen recent discussion related to NET_EPOCH.
This is way out of my expertise, but...

NFS traffic is basically bi-directional small messages.
If a change increases the latency of reception of a small
message (TCP segment) which is not followed by further
TCP segments in the same direction...
--> A significant NFS performance hit could be observed.

I don't have hardware that can do reasonable benchmarking,
so if some else can test for NFS performance regressions
when making TCP stack changes, it would be appreciated.
--> The time it takes to do a fairly large build like a kernel
       or buildworld over NFS on fast hardware with fast networking
       would probably do it.

Just wanted the people doing TCP stack work to be aware of this, rick