Re: another issue with error recovery...

[Michael Krause <krause@cup.hp.com>]

At 07:04 AM 6/27/00 +0300, julian_satran@il.ibm.com wrote:


>As far as I know that is a recognized issue in the TCP world.  Let's hope
>it will we fixed.

The same problem exists in other technologies as well, e.g. InfiniBand.  At 
3 GB/s, the sequence number space gets used up in a couple hundred 
milliseconds using the smallest packet size.  However, is this really a 
realistic problem for either technology?

If one is only pounding on a single connection and has an application that 
can generate the workload, then it is a problem and one needs to expand the 
sliding window to much larger values (e.g. a TCP jumbo frame).  However, if 
one is using multiple connections sharing the same physical links, then the 
problem is further mitigated because there are multiple windows being 
operated upon and neither is being consumed as quickly.

Note: The problem will only get worse.  10 GbE will be here second half of 
next year; 100 GbE will be here 5+ years after that.  All of the transports 
including those still be developed, will not work well at those operating 
speeds.  One would need to expand the windows (however implemented) to 
accommodate much larger values - measured in GBs instead of MBs - to have 
any measure of success.

Mike