| 1 |
Hello, |
| 2 |
|
| 3 |
While browsing the list of kernel changes in 2.6.32 at |
| 4 |
kernelnewbies.org, I noticed this: |
| 5 |
|
| 6 |
http://kernelnewbies.org/Linux_2_6_32#head-e07f19bb79beab04cb9e68b942c9041b490b71db |
| 7 |
|
| 8 |
I'm posting this as a heads up as the article alleges that it can |
| 9 |
"cause noticeable performance issues". Moreover, it defaults to being |
| 10 |
swtiched on but is not exposed an option in the kernel build system. |
| 11 |
It can only be turned off via sysfs: |
| 12 |
|
| 13 |
echo 0 > /sys/class/block/<device name>/queue/iosched/low_latency |
| 14 |
|
| 15 |
Therefore, I would advise those who are using CONFIG_IOSCHED_CFQ - and |
| 16 |
who to intend to upgrade to 2.6.32 - to test the impact of the option |
| 17 |
upon their respective workloads. |
| 18 |
|
| 19 |
Perhaps it's just my interpretation, but I have detected an upstream |
| 20 |
trend of introducing changes which are either best suited to desktop |
| 21 |
users or which have had a tendency to introduce performance |
| 22 |
regressions that are of concern with respect to the server role |
| 23 |
(gentoo bug #247453 to name but one example). Hopefully this turns out |
| 24 |
to be a false alarm, but I thought that it would be worth mentioning |
| 25 |
in any case. |
| 26 |
|
| 27 |
Cheers, |
| 28 |
|
| 29 |
--Kerin |
Archives