root@sbcs.sunysb.edu (SBCS System Staff) (02/19/90)
I have a port of Berkeley TCP-IP that I am interested in tuning. One area that is causing some trouble is the management of the fragment assembly queue in a low/no memory situation. The current Berkeley code simply flushes all the fragment queues when the system is out of network buffers. One possible solution would be to LRU the queues out, but the claim I heard was this will unduly penalize long delay networks. Another solution recommended was to just randomly discard queues when out of memory. Does anyone have a better algorithm than the two discussed here? It would seem that something along the lines of LRU modified with information about average packet arrival rate would be more appropriate (but is it worth the extra work). Rick Spanbauer State U of NY/Stony Brook