read_ahead (disabled) as steroid
Last week we were busy to align XtraDB performance with 5.4, now we have some results. Currently it is available as “hacks” to XtraDB (available on Lauchpad lp:~percona-dev/percona-xtradb/hacks-porting-tune if you are interested). Basically we took improvements from 5.4 and backported ones performance related to XtraDB.
Here are results for tpcc-like workload, 100W (~10GB) ( raw results and parameters are available here
http://spreadsheets.google.com/ccc?key=rxUEhM2dqbX0uAfq9j6WQ_w ). Box Dell PowerEdge R900 (Does Dell have referral program ? ), with RAID10 (8 disks) on ext3, 32GB of RAM.
As you see there almost no difference and you may say what’s the reason in XtraDB ? The most interesting reason is XtraDB based on InnoDB-plugin and contains its nice features like FAST INDEX CREATION and dynamic pages. And XtraDB has some parameters like “adaptive_checkpoint” and control of “read_ahead”. And if in the same benchmark you disable read-ahead (innodb_read_ahead=none), you can see improvement about 15%
Actually control of read_ahead is very simple patch and can easily be included in 5.4 or InnoDB-plugin.
For curiosity I run the same benchmarks on ext3 vs xfs on SSD card, the results are:
Interesting facts:
- xfs is 25% faster ext3
- the gaps on SSD is more visible than on RAID10
- in the same time gaps can be smoother with disabled read_ahead, however disabling it does not show such improvement like on RAID10
I should mention I have strange results on xtradb running it on xfs on RAID10, I have results about 2times slower than for ext3. I am not sure yet – is it xtradb or xfs problem, and why it appears only on RAID10, but not SSD.
del.icio.us
digg4 Comments
What did you backport from 5.4?
Comment :: May 12, 2009 @ 11:01 pm
Mark,
Here is comment from Yasufumi:
- “pause” (rep; nop) instruction in spin loop should increase performance of
mutex/rw_lock contention. It may affect 5~10% in throughput.
- log_buffer_flush_maybe_sync() will decrease the number of log sync
at innodb_flush_log_at_trx_commit=[0|2]. It may affect to performance much.
- The number of request slots for the each IO thread has been increased. (x4)
It might be short for some high-end environments.
- relax condition to flush log and contract ibuf at inner loop of srv_master_thread()
– if (n_pend_ios < 3 && (n_ios – n_ios_old < PCT_IO(5))) {
+ if (n_pend_ios < PCT_IO(3) && (n_ios – n_ios_old < PCT_IO(5))) {
Comment :: May 12, 2009 @ 11:27 pm
Should read_ahead always be disabled then to improve performance?
What are the pros and cons?
Comment :: May 13, 2009 @ 3:36 pm
Andy,
I do not think so, it depends on IO subsystem. For single HDD read_ahead may be useful.
Comment :: May 13, 2009 @ 8:00 pm