lkml.org 
[lkml]   [2002]   [Oct]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] pooling NUMA scheduler with initial load balancing
On Tue, Oct 08, 2002 at 07:33:06PM +0200, Erich Focht wrote:
> Aaargh, you got the wrong second patch :-( Sorry for that...
>
> Thanks for the hints, I cleaned up the first patch, too. No
> CONFIG_NUMA_SCHED any more, switched to MAX_NUMNODES, including
> asm/numa.h from asm/topology.h, so no need for you to see it.


diff -urNp a/arch/i386/kernel/smpboot.c b/arch/i386/kernel/smpboot.c
--- a/arch/i386/kernel/smpboot.c Fri Sep 27 23:49:54 2002
+++ b/arch/i386/kernel/smpboot.c Tue Oct 8 11:37:56 2002
@@ -1194,6 +1194,11 @@ int __devinit __cpu_up(unsigned int cpu)
void __init smp_cpus_done(unsigned int max_cpus)
{
zap_low_mappings();
+#ifdef CONFIG_NUMA
+ pooldata_lock();
+ bld_pools();
+ pooldata_unlock();
+#endif

All callers of bld_pools() need the pooldata lock - taking
it inside that function makes the code a little more readable..
Also I'd suggest to rename bld_pools() to build_pools() ;)

- cache_decay_ticks = 10; /* XXX base this on PAL info and cache-bandwidth estimate */
+ cache_decay_ticks = 8; /* XXX base this on PAL info and cache-bandwidth estimate */
Could you explain this change? And it's affect on non-NUMA IA64
machines?

/**
+ * atomic_inc_return - increment atomic variable and return new value
+ * @v: pointer of type atomic_t
+ *
+ * Atomically increments @v by 1 and return it's new value. Note that
+ * the guaranteed useful range of an atomic_t is only 24 bits.
+ */
+static inline int atomic_inc_return(atomic_t *v){
+ atomic_inc(v);
+ return v->counter;
+}
Who do you guarantee this is atomic? Please make it fit
Documentation/CodyingStyle, btw..

+int numpools, pool_ptr[MAX_NUMNODES+1], pool_cpus[NR_CPUS], pool_nr_cpus[MAX_NUMNODES];
+unsigned long pool_mask[MAX_NUMNODES];

Hmm, shouldn't those [MAX_NUMNODES] arrays be in some per-node array
to avoid cacheline-bouncing?

+void pooldata_lock(void)
+{
+ int i;
+ retry:
+ while (atomic_read(&pool_lock));
+ if (atomic_inc_return(&pool_lock) > 1) {
+ atomic_dec(&pool_lock);
+ goto retry;
+ }
Why not a simple spin_lock()?

+ /*
+ * Wait a while, any loops using pool data should finish
+ * in between. This is VERY ugly and should be replaced
+ * by some real RCU stuff. [EF]
+ */
+ for (i=0; i<100; i++)
+ udelay(1000);
Urgg. I'd suggest you switch to RCU now and make your patch apply
ontop of it - another reason to apply the RCU core patch..

+void pooldata_unlock(void)
+{
+ atomic_dec(&pool_lock);
+}
Dito for spin_unlock.

+ /* avoid deadlock by timer interrupt on own cpu */
+ if (atomic_read(&pool_lock)) return;
spin_trylock..

All in all your code doesn't seem to be very cachelign-friendly,
lots of global bouncing. Do you have any numbers on what your
patch changes for normal SMP configurations?
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:29    [W:0.094 / U:0.120 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site