lkml.org 
[lkml]   [2004]   [Aug]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subjectkernbench on 512p
Date
I set out to benchmark the page cache round robin stuff I posted earlier to 
see whether round robining across 256 nodes would really hurt kernbench
performance. What I found was that things aren't scaling well enough to make
reasonable measurements.

time user system %cpu ctx sleeps
2.6.8.1-mm1+nodespan 108.622 1588 10536 11158 31193 397966
2.6.8.1-mm1+ns+rr 105.614 1163 8631 9302 27774 420888

2.6.8.1-mm1+nodespan is just 2.6.8.1-mm1 with my last node-span patch applied.
2.6.8.1-mm1+ns+rr is that plus the page cache round robin patch I posted
recently. I just extracted the 'optimal run' average for each of the above
lines.

Here's the kernel profile. It would be nice if the patch to show which lock
is contended got included. I think it was discussed awhile back, but I don't
have one against the newly merged profiling code.

[root@ascender root]# readprofile -m System.map | sort -nr | head -30
208218076 total 30.1677
90036167 ia64_pal_call_static 468938.3698
88140492 default_idle 229532.5312
10312592 ia64_save_scratch_fpregs 161134.2500
10306777 ia64_load_scratch_fpregs 161043.3906
8723555 ia64_spinlock_contention 90870.3646
121385 rcu_check_quiescent_state 316.1068
40464 file_move 180.6429
32374 file_kill 144.5268
25316 atomic_dec_and_lock 98.8906
24814 clear_page 155.0875
17709 file_ra_state_init 110.6813
17603 clear_page_tables 13.4169
16822 copy_page 65.7109
16683 del_timer_sync 32.5840
15098 zap_pte_range 7.8635
15076 __d_lookup 16.8259
13275 find_get_page 31.9111
12898 __copy_user 5.5214
10472 finish_task_switch 36.3611
10037 ia64_pfn_valid 52.2760
8610 get_zone_counts 22.4219
7905 current_kernel_time 41.1719
7155 __down_trylock 22.3594
7063 _pagebuf_find 4.9049
6995 xfs_ilock 13.6621
6965 tasklet_action 8.3714
6494 free_page_and_swap_cache 28.9911
5652 nr_free_pages 22.0781
5555 xfs_trans_push_ail 3.6165
This is a 64p profile only. If I set the prof_cpu_mask to include all 512
CPUs, the system livelocks. I reset the counter right after the warmup run,
partly through the half load run, and collected it after a few runs were
complete.

Thanks,
Jesse
-
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 14:05    [from the cache]
©2003-2011 Jasper Spaans