Throughput problems with NFS

[Matthew Dillon]

:On Sunday, October 12, 2003, at 12:20 AM, Matthew Dillon wrote:
:>     I don't understand.  An unacked segment should not present a 
:> problem.
:>
:>     You may have to make that reserve scale by MTU.
:
:This happens because Linux hangs onto the ack for the last segment of a
:32K+header send for a while. Then the NFS server will not put anymore
:data in the socket because of an soreserve with a size of 32kB+header, 
:so
:it waits for about 39mS until Linux finally sends the ack for the last
:segment.
:
:-DR

    Ah, ok now I see.  It sounds like soreserve should nominally be 2.5x
    or 3x the NFS io size in order to properly stream requests, or you
    won't be able to fill the wire.

					-Matt
					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>