Terrible PA8800 SMP Performance

Terrible PA8800 SMP Performance

[John David Anglin]

I've been doing GCC builds on a four processor PA8800 currently
running 2.6.24-rc6.  With make -j 6 bootstrap, this machine is
several hours slower than gsyprf11 doing the same GCC build and
check.  gsyprf11 is a two processor PA8700 and I typically
use make -j 4 on it.  The PA8800 is also slower than my c3750.

I've mentioned this before but some tests that never timeout even
on slow machines randomly timeout on the PA8800.

Any thoughts on what's wrong?

Dave
-- 
J. David Anglin                                  dave.anglin@nrc-cnrc.gc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6602)

Re: Terrible PA8800 SMP Performance

[Kyle McMartin]

On Sun, Jan 20, 2008 at 02:55:18PM -0500, John David Anglin wrote:
> I've been doing GCC builds on a four processor PA8800 currently
> running 2.6.24-rc6.  With make -j 6 bootstrap, this machine is
> several hours slower than gsyprf11 doing the same GCC build and
> check.  gsyprf11 is a two processor PA8700 and I typically
> use make -j 4 on it.  The PA8800 is also slower than my c3750.
> 

flushing 32MB of cache on every fork and exec? :/

cheers,
	kyle

Re: Terrible PA8800 SMP Performance

[John David Anglin]

> flushing 32MB of cache on every fork and exec? :/

That's what I was thinking, although the actual situation may be worse.
It appears we may flush the entire cache (32MB I-cache and 32MB D-cache)
on every cpu.

static void cacheflush_h_tmp_function(void *dummy)
{
        flush_cache_all_local();
}

void flush_cache_all(void)
{
        on_each_cpu(cacheflush_h_tmp_function, NULL, 1, 1);
}

void flush_cache_mm(struct mm_struct *mm)
{
#ifdef CONFIG_SMP
        flush_cache_all();
#else
	flush_cache_all_local();
#endif
}

Dave
-- 
J. David Anglin                                  dave.anglin@nrc-cnrc.gc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6602)