Subject: Re: NFS/RPC and server clusters
To: Thor Lancelot Simon <email@example.com>
From: David Laight <firstname.lastname@example.org>
Date: 10/15/2003 23:16:42
> 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.
However on a local lan segment packets don't get dropped.
So UDP has the advantage that ACKs are only sent when needed - and
typically when the lan is idle.
On a HDX network this gets bad...
If a system sends a large TCP message it will go out as back to back
ethernet frames. The receiving systems tcp stack will (probably)
generate TCP ACKs for every other frame. None of these can be transmitted
because the media is busy, so the MAC starts doing longer and longer
timeouts. Then the response will also get queued. Eventually the MAC
will find the LAN idle and send out a back to back stream of ACKs
followed by the response. The original system now has to process
all the ACK packets before getting to the response itself.
David Laight: email@example.com