On Mar 25, 2008, at 5:06 PM, Akins, Brian wrote:
No - it does not; so you get that speed increase (which is very
noticable on a swapspace/ram-disk*).
I'd like to see some numbers on that. I just did a quick test on
Linux and
saw no real improvement (testing our hacked-to-heck version).
I'll see if I can get something decent - am measuring anyways.
Meanwhile - in your test - did most of your disk-IO get cached
regardless ? Using a config:
# create with
#
# M=`mdconfig -a -s 500M`
# newfs /dev/$M; mount /dev/$M /mnt
# mkdir /mnt/_cache /local0/_cache
# chown www:www /mnt/_cache /local0/_cache
#
CacheDirLevels 5
CacheDirLength 3
CacheIgnoreNoLastMod on
CacheStoreNoStore on
CacheStorePrivate On
TransferLog /dev/null
ErrorLog /dev/null
Listen 127.0.0.1:1081
<VirtualHost *:1081>
CacheRoot /mnt/_cache
CacheEnable disk /
</VirtualHost>
Listen 127.0.0.1:1082
<VirtualHost *:1082>
CacheRoot /local0/_cache
CacheEnable disk /
</VirtualHost>
include "conf/tset-woollight.com-no-override"
And ensuring that there is enough traffic to 'hide' the disk - and a
reply of a log file across about 300M of data; 10% responsible for 80%
of the traffic; remainder spread flat (not log/log) I get about:
Requests per second: 5146.06 [#/sec] (mean)
versus
Requests per second: 2031.03 [#/sec] (mean)
However - drop the disk traffic and keep the footprint to 10Mbyte and
I get:
Requests per second: 5121.30 [#/sec] (mean)
versus
Requests per second: 4837.41 [#/sec] (mean)
with an standard deviation which is soo large that the latter two are
about the same. Which supports your point perfectly.
Though if anyone can find me some more serious hardware* - I'd love to
do this proper; as I am struggling getting a fixed mod_memm_cache and
mod_memcached_cache to be taxed hard enough to actually measure/
profile sensibly (right now I saturate network(switches), irq2cpu
affinity of the network cards and IO - and just cannot get any
meaningful number).
Thanks,
Dw
*: anyone - feel free to ping me privately if you have a spare 3 tier
1Gb/s network with a perfectly full duplex
perfectly switched network fabric, a handful of 2x2 core/ 8Gb+ /
fast scsi/sas machines (or a very decent blade
chassis) and a nearby linux/freebsd capable sysadmin. Have SSH key
which travels well.
