[httperf] file descriptors

Rick Jones rick.jones2 at hp.com
Tue Jul 29 16:56:58 PDT 2008


Mark Nottingham wrote:
> Oops, looked in the wrong place, this is
> 
> #define    EADDRNOTAVAIL    99    /* Cannot assign requested address */
> 
> Hmm...
>   http://www.ussg.iu.edu/hypermail/linux/kernel/9910.3/0843.html

Does httperf pick its own local port numbers, or is it leaving that 
decision to the stack?  Given the "requirement" for random ISN's these 
days, the "Get out of TIME_WAIT and go to ESTABLISHED" path is pretty 
much toast, so when you have one client IP connecting to one server IP 
and well-known port number, once the connection "churn" rate goes above:

sizeof(client_port_space_used)/lengthof(TIME_WAIT)

you will get attemtps to re-use a four-tuple which is still in TIME_WAIT.

IFF all the TIME_WAIT are on the client and IFF the client is _not_ part 
of what is being benchmarked, then one might consider shortening the 
length of TIME_WAIT, but only after increasing the client port space 
used to say 5000-65535 and determinine that using additional IP 
addresses on the client(s) is infeasible.

Whatever is done, _don't_ start using abortive closes.

rick jones


More information about the httperf mailing list