> > 	the numbers i've posted (hoge.c) could be misleading, as (1) it only
> > 	measures the algorithm itself, not the impact to RPC code as a
> > whole,
> True - that is why I was careful not to overreact to your timings.
> Can you post the results of lat_rpc for comparison?

	i have already.

> > 	(2) it runs on userland so kernel-userland overhead like gettimeofday()
> > 	call which makes it inaccurate.
> True but that overhead is relatively constant in your before/after
> case so the perturbation in your measurement should not be that high,
> certainly not by 3 orders of magnitude which is what your numbers were
> indicating.

	in "before" case it is
		getpid() ^ tv.tv_sec ^ tv.tv_usec
	therefore there's gettimeofday(2) in every loop (3 system calls per
	a loop).

itojun