Subject: Re: NFS/RPC and server clusters
To: None <firstname.lastname@example.org>
From: Jonathan Stone <jonathan@DSG.Stanford.EDU>
Date: 10/15/2003 13:30:49
>Actually, there is. If you're using large RPC requests, TCP will
>segment them neatly into multiple underlying IP packets; UDP will
>generate gigantic packets that must then be fragmented and reassembled.
>Aside from the inherent inefficiency, this also defeats checksum offload
>on most network adapters that can do checksum offload. Oops!
All well, and true, and good. However, a datapoint:
I regularly set up NFS throughput saturation tests, using multiple
i386 machines (1.4GHz Tualatin or faster P4 Xeons) to export MFS
filesystems to NFS clients. Aggregate write throughput is over
100Mbyte/sec. All NICs are either bge or wm. Checksum assist is
In that regime (and using a hashed IP reassembly queue, which I hope
to commit RSN), NFS over UDP is still faster than TCP.