Running NFS in long fat network (LFN) or gigabit WAN requires that the
TCP windows be sized appropriately.
An example of this is a recent customer request where they expected
parity with Linux ie. 20MB read throughput over 20ms RTT
networks. This can only be achieved with tcp window set at more
than 400K.
As it stands, both NFS server and client gets their window
sized from the tcp system default tcp_recv_hiwat etc. It may not be
appropriate to have all network connections use the larger values
required for an LFN.
The subject of this RFE is to request an NFS specific way to tune the
window sizes either using /etc/default/nfs, mount option, share
option or other means.
xxxxx@xxxxx.com 2003-11-14
<Edited the above comment replacing socket buffers with window size>
Work Around
[Unconfirmed] Use route(1M) to set the buffer sizes on a per route basis:
-recvpipe
Receive pipe size in bytes.
-sendpipe
Send pipe size in bytes.