lkml.org 
[lkml]   [2012]   [Sep]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Subject[patch] block: make struct block_device cacheline_aligned
Date
Hi,

When testing against a pcie ssd or a ramdisk, making the block device
structure cacheline_aligned provided a significant increase in
performance:

vanilla
READ WRITE CPU
Job Name BW IOPS msec BW IOPS msec usr sys csw
write1 0 0 0 748522 187130 44864 16.34 60.65 3799440.00
read1 690615 172653 48602 0 0 0 13.45 61.42 4044720.00
randwrite1 0 0 0 716406 179101 46839 29.03 52.79 3151140.00
randread1 683466 170866 49108 0 0 0 25.92 54.67 3081610.00
readwrite1 377518 94379 44450 377645 94410 44450 15.49 64.32 3139240.00
randrw1 355815 88953 47178 355733 88933 47178 27.96 54.24 2944570.00

patched
READ WRITE CPU
Job Name BW IOPS msec BW IOPS msec usr sys csw
write1 0 0 0 871355 217838 38508 17.49 42.46 1642870.00
read1 1418560 354639 23675 0 0 0 14.96 54.75 337489.00
randwrite1 0 0 0 736970 184242 45633 30.62 35.25 1409440.00
randread1 1065440 266359 31544 0 0 0 32.67 43.74 255394.00
readwrite1 657940 164484 25867 657848 164461 25867 18.54 50.55 619474.00
randrw1 491940 122985 34245 492014 123003 34245 34.44 41.05 418999.00

%diff
READ WRITE CPU
Job Name BW IOPS msec BW IOPS msec usr sys csw
write1 0 0 0 16 16 -14 7.04 -29.99 -56.76
read1 105 105 -51 0 0 0 11.23 -10.86 -91.66
randwrite1 0 0 0 0 0 0 5.48 -33.23 -55.27
randread1 55 55 -35 0 0 0 26.04 -19.99 -91.71
readwrite1 74 74 -41 74 74 -41 19.69 -21.41 -80.27
randrw1 38 38 -27 38 38 -27 23.18 -24.32 -85.77

BW=bandwidth in KB/s
IOPS = I/Os per second
msec = # of miliseconds the run took (lower is better)
usr = % user time
sys = % system time
csw = # of context switches

The test is doing asynchronous direct I/O to the block device using 4
processes each driving a queue depth of 1024 to a different part of the
disk. The rows, in order, are sequential write, sequential read,
random write, random read, 50% mix of sequential reads and sequential
writes, 50% mix of random reads and random writes. The block size in
all cases is 4k.

I'd appreciate it if others could verify an increase in performance with
this patch. Thanks to Mikulas for initially suggesting that the cache
size/alignment was relevant to performance.

Cheers,
Jeff

Signed-off-by: Jeff Moyer <jmoyer@redhat.com>

diff --git a/include/linux/fs.h b/include/linux/fs.h
index aa11047..87ce6ca 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -724,7 +724,7 @@ struct block_device {
int bd_fsfreeze_count;
/* Mutex for freeze */
struct mutex bd_fsfreeze_mutex;
-};
+} __cacheline_aligned;

/*
* Radix-tree tags, for tagging dirty and writeback pages within the pagecache

\
 
 \ /
  Last update: 2012-09-19 16:21    [W:0.060 / U:0.048 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site