Obviously, the Power7 architecture is pound-for-pound more powerful than any Intel platform — Sparc isn’t even a consideration.

I’d love to see the same tests run on the new 750 series Power7 CPU’s and hardware from IBM.

My guess is that we’d see another full CLASS of hardware / performance benchmarks unlike any yet seen….

What filesystem did you use here ? I remember EXT3 performs surprisingly bad with sync_binlog=1

http://www.mysqlperformanceblog.com/2009/01/21/beware-ext3-and-sync-binlog-do-not-play-well-together/

In any case it looks like sync_binlog has very bad overhead. in innodb_flush_log_at_trx comit=1 you should have 2 fsync per commit in single client (because of XA) with binlog it should be 3 as I understand – the fact performance gap is much larger is different.

Does group commit works on stage of writiting to binary log or is it only on transactional log commit ?

1) The huge drop in performance (sometimes over 50%) due to binlog is a surprise. According to the MySQL Manual “Running a server with binary logging enabled makes performance about 1% slower.” (http://dev.mysql.com/doc/refman/5.1/en/binary-log.html)

What caused the huge discrepancy between the 50% you measured and the 1% according to MySQL manual?

2) If the binlog IO is the bottleneck, would putting binlog on FusionIO (instead of RAID 10 HDD) speed things up?

3) What about innodb_support_xa? Will enabling that cause a further significant drop in performance?

time dd if=/dev/zero of=a.dat oflag=sync count=100000
100000+0 records in
100000+0 records out
51200000 bytes (51 MB) copied, 27.9191 s, 1.8 MB/s

real 0m27.931s
user 0m0.047s
sys 0m3.906s

About 3.5K writes per sec.

BTW, I was wondering how does XtraDB and the like behaves over different RAID-10 stripe sizes and the fact that actually a lot of Linux distros don’t have their default partition boundaries aligned on raid stripe sizes (due to 63 sector offset from a msdos partition table). Have you ever considered (or have any plans) in making those tests? It would be great to see how XtraDB behaves on unaligned vs aligned systems.

I’ve found some interesting links:
http://www.linux.com/archive/articles/140734
http://www.linuxfoundation.org/collaborate/workgroups/linux-raid/raid_setup
http://kbase.redhat.com/faq/docs/DOC-2893 { unfortunately didn’t find anything like this for XFS }

Best regards,
Raine

Few questions :
- should innodb_flush_log_at_trx_commit = 2 change anything on this ?
- if you use replication, by default tutorials put you on the green curve. Would you advise moving to “trx_commit=1 & binlog” or even “trx_commit=0 & binlog” ?

Thanks,
Nicolas