Re: iperf: performance regression (was b44 driver problem?)

[Thomas Gleixner <tglx@linutronix.de>]

On Mon, 2007-06-04 at 09:59 -0700, Stephen Hemminger wrote:
> > > gettimeofday({1180973726, 982754}, NULL) = 0
> > > recv(4, "\0\0\0\0\0\0\0\1\0\0\23\211\0\0\0\0\0\0\0\0\377\377\364"..., 8192, 0) = 8192
> > > gettimeofday({1180973726, 983790}, NULL) = 0
> > 
> > Well, gettimeofday() is not affected by the highres code, but
> > 
> > > nanosleep({0, 0}, NULL) = 0
> > > nanosleep({0, 0}, NULL) = 0
> > 
> > is. The nanosleep call with a relative timeout of 0 returns immediately
> > with highres enabled, while it sleeps at least until the next tick
> > arrives when highres is off. Are there more of those stupid sleeps in
> > the code ?
> 
> GLIBC pthread_mutex does it, YES it is a problem!
> Looks like the old behavior is required for ABI compatibility.
>
> iperf server has several threads. One thread is using pthread_mutex_lock
> to wait for the other thread.  It looks like pthread_mutex_lock is using
> nanosleep as yield().

I doubt that. This is in the iperf code itself.

void thread_rest ( void ) {
#if defined( HAVE_THREAD )
#if defined( HAVE_POSIX_THREAD )
    // TODO add checks for sched_yield or pthread_yield and call that
    // if available
    usleep( 0 );

----------^^^^

It results in a nanosleep({0,0}, NULL)

	tglx