Comments on: How many fsync / sec FusionIO can handle

By: Vadim

Vadim — Mon, 22 Mar 2010 19:45:37 +0000

VJ Kumar,

I blogged after that, that FusionIO fixed issue with durability in default mode, so you do not lose you data.

By: VJ Kumar

VJ Kumar — Fri, 05 Mar 2010 20:19:27 +0000

Vadim,

You haver earlier written: “As I understand FusionIO in default mode, like Intel SSD, is â€œlyingâ€ to application, and fsync() is not real, it still commit only to internal memory, not to final media. And FusionIO has â€œstrict_syncâ€ mode to guarantee fsync is trustworthy.”

Is it still the case ? If it is, obviously, one cannot use the device for transaction log storage.

By: VJ Kumar

VJ Kumar — Fri, 05 Mar 2010 20:08:22 +0000

Domas,

What is so funny ?

The numbers Vadim provided mirror my experience with sequential writes to any transaction log, not necessarily InnoDb’s.

If transaction log writes are disturbed by other IO, then the pattern becomes random and IOPS/sec will probably go down to a couple hundred per sec. Whether the file system is ext3/4 or xfs does not matter much (ignoring barriers and journaling for simplicity).

By: peter

peter — Fri, 05 Mar 2010 19:38:28 +0000

Guys,

I’d offer some background for this test. What we wanted to check if whenever FusionIO is good enough to keep transactional logs on it. For a lot of SSDs the fsync() rate they can do is slower compared to writes to the RAID cache (with BBU) It looks like it is not the case for FusionIO which is great.

By: Baron Schwartz

Baron Schwartz — Fri, 05 Mar 2010 05:56:50 +0000

Interesting. So apparently only XtraDB has “fixed” group commit? I was under the same impression as Vadim.

By: Mark Callaghan

Mark Callaghan — Fri, 05 Mar 2010 05:27:20 +0000

Group commit is fixed when the binlog is off. Otherwise, prepare_commit_mutex is locked in innobase_xa_prepare and unlocked in innobase_commit. The binlog write/sync is done between the two. I am not aware of any (or any good) documentation of this in MySQL or InnoDB docs and the plugin release notes that state that group commit is fixed might confuse people. Not having group commit can be a huge deal depending on your IO setup.

Subscribe to my feature request for this at http://bugs.mysql.com/bug.php?id=49326. Or if you have a support contract, let them know you need a fix.

By: Vadim

Vadim — Fri, 05 Mar 2010 05:05:12 +0000

Domas,

this is xfs with nobarrier.

Mark,

What about InnoDB-plugin ? it is supposed group commit is fixed there…

By: Mark Callaghan

Mark Callaghan — Fri, 05 Mar 2010 00:02:16 +0000

prepare_commit_mutex guarantees there is no group commit as soon as the binlog is enabled. What fraction of InnoDB deployments that care about performance run with the binlog off?

By: Domas

Domas — Thu, 04 Mar 2010 23:42:26 +0000

are you testing append performance on ext3?! hehehehehehehehehe

By: Gavin Towey

Gavin Towey — Thu, 04 Mar 2010 21:42:12 +0000

Andy:

fsync/s is not the same as throughput, which is not the same as TPM. Those tests measure different things.