mo@prisma.UUCP (Mike O'Dell) (06/04/89)
Ok, so we don't do keep-alives at the TCP level. This is the best reason I've heard for a session protocol.
mckenney@distek4.uucp (Paul E. McKenney) (06/06/89)
In article <8906040508.AA20495@uunet.uu.net> mo@prisma.UUCP (Mike O'Dell) writes: >Ok, so we don't do keep-alives at the TCP level. This is the best reason >I've heard for a session protocol. Correct me if I am wrong, but if the transport layer cannot determine a good keepalive ``timeout'' from the RTT and window-size information available to it, then a session layer hasn't a ghost of a chance of coming up with anything reasonable. Has anyone tried to derive keepalive timeouts from RTT and window information? For a simple example, one might set the timeout to (3000*RTT). This would yield a timeout of about 25 minutes for typical NSFnet RTTs (about half a second). On low-baudrate lines, the longer packet transmission times would force the RTT higher (a 1200-baud line requires two-thirds of a second to turn a minimum-size TCP packet around). Thanx, Paul