Test config changes for Veritas CIO and Linux IO Scheduler
May 3, 2011 1 Comment
Just a short note to report on the impact that some config changes have made on IO times in a specific environment for a specific workload.
I mentioned previously that I’m working on an upgrade of an application from 22.214.171.124 to 126.96.36.199
- There are changes in pretty much every area – new hardware, different OS, etc, etc.
- We’re using a full-volume UAT environment within the new set-up to compare new against old (production) and that will form the main basis for performance changes to the application required for this upgrade.
- It’s pretty much an apples vs oranges comparison and not helped by the fact that UAT runs on tier 2 storage to be compared against the current tier 1 storage – UAT IO is slow.
- In summary, not exactly best practice upgrade approach – but that’s just how it is sometimes in the real world.
Anyway, anyway, anyway… we’ve been waiting for recommendations and official go-ahead from the database engineering group who run the tests and control the builds of the machines and the following config changes have been made to the following:
- Veritas VxFS CIO
- Linux Deadline scheduler
Ideally such changes would be made individually to gauge their individual impact, however, as mentioned, it’s not always like that is it?.
And on the UAT environments above, based on a before-flashback-after run of the main application workload, the following IO times were observed:
|Wait Event||Average wait time (before)||Average wait time (after)|
|db file sequential read||10 ms||7 ms|
|db file scattered read||21 ms||11 ms|
|db file parallel read||38 ms||60 ms|
|direct path read||95 ms||64 ms|
|direct path read temp||32 ms||9 ms|
|direct path write temp||34 ms||8 ms|
|log file sync||9 ms||3 ms|
I’m not convinced that we have a level of control over the whole environment and time to deliver change incrementally to read too much into a single before/after comparison of the same workload, however these initial findings were better than I expected.
(Bottom line is that it’s still apples vs oranges)
So… it may be that someone finds this useful.