Re: 2.6.8.1-mm3 lockmeter on 512p w/kernbench

[Greg Edwards <edwardsg@sgi.com>]

On Fri, Aug 20, 2004 at 02:04:32PM -0400, Jesse Barnes wrote:
| More lockstats.  dcache is obviously still there, but for some reason the rcu 
| stuff is gone (I didn't apply Manfred's patches).  I must have done some 
| stuff prior to collecting the lockstat data last time that caused it.

...[snip]...

| SPINLOCKS         HOLD            WAIT
|   UTIL  CON    MEAN(  MAX )   MEAN(  MAX )(% CPU)     TOTAL NOWAIT SPIN RJECT  NAME
| 
|         3.1%  2.2us( 102ms)   12ms( 488ms)(35.1%) 320195677 96.9%  3.1% 0.00%  *TOTAL*

...[snip]...

|  31.2% 95.2%   71us(  10ms)   37ms( 146ms)(30.2%)   2828650  4.8% 95.2%    0%  rcu_check_quiescent_state+0xf0

Looks like the RCU contention is still there.

I ran some bench marks at 512p on a 2.6.5 SUSE kernel with and without
Manfred's patches 4 and 5 added (the SUSE kernel already has the 3
accepted RCU patches), and the results were very promising.  I think
we'll want to pursue getting these other two RCU patches accepted.

I'll to get some lockmeter numbers on a latest -mm kernel next week.

Greg