Comments on: iiBench Contest – Who Can Insert 1B Rows into MYSQL the Fastest?

By: Bradley C. Kuszmaul

Bradley C. Kuszmaul — Sat, 06 Dec 2008 02:32:19 +0000

To do well by that metric, all I have to is slow down my first 100M insertions

Yes, our research is all about fast insertions into indexes. But I don’t want this to be a contest of MyISAM vs. some unspecified Tokuteknology. That doesn’t seem fair.

I’m interested in two things: (1) how fast *can* MyISAM or InnoDB go? (2) What can we do to improve iiBench to both reflect real-world insertion problems and to head off optimizations such as FakeAmelia’s (which is a great hack for this contest, but doesn’t necessarily achieve our intent that useful indexes should be maintained).

By: Mark Callaghan

Mark Callaghan — Sat, 06 Dec 2008 01:55:36 +0000

Another result metric that can be used to judge how well a storage supports this workload is (time to insert first 100M rows / time to insert last 100M rows). I will have numbers soon. InnoDB, with a few patches and better my.cnf parameters and without RAID5, does better although it still degrades a lot over time. I will guess that Tokutek does great on this test. Where are your results for that?

By: Bradley C. Kuszmaul

Bradley C. Kuszmaul — Fri, 05 Dec 2008 21:02:54 +0000

That kind of hardware variant seems fair to me.

By: Nils

Nils — Fri, 05 Dec 2008 17:33:08 +0000

Hmm well I don’t have that kind of hardware lying around… I got a spare box which has 8 SATA 10k RPM disks with a 3ware controller, maybe I’ll try on that one.

By: Bradley C. Kuszmaul

Bradley C. Kuszmaul — Fri, 05 Dec 2008 17:07:33 +0000

The choice of using RAID5 was kind of arbitrary. One of our beta customers was benchmarking us against MyISAM on a RAID 5, so we did the same thing.

It would be interesting to try another RAID configuration. If you can do better than we did, it might form the basis for a winning entry!

By: Hakan Kücükyilmaz

Hakan Kücükyilmaz — Fri, 05 Dec 2008 16:53:43 +0000

The RAID5 kills your write performance!

By: Nils

Nils — Thu, 04 Dec 2008 19:43:22 +0000

Why did you choose RAID-5 over RAID-10 (RAID0 is probably not an option)?

By: Mark Callaghan

Mark Callaghan — Thu, 04 Dec 2008 19:20:36 +0000

This is an interesting test that will find problems in other storage engines. So we can make things faster, but InnoDB will never be optimal for this workload. Some feedback:
* your test script would be much easier to use were it implemented in Perl or Python
* post the my.cnf settings you used for your test

I would like to see another test that does both inserts and updates. Some interesting workloads require a mix. Insert performance was in the news here — http://www.dbms2.com/2008/12/02/data-warehouse-load-speeds-in-the-spotlight