?

Log in

No account? Create an account

Peter Z's comments on PBXT, My own thoughts

« previous entry | next entry »
Apr. 9th, 2007 | 02:40 am

LJ ate part of the post, so I am updating it with the original content.

I've not been posting a lot about my studies on bottlenecks in the
server and my work on what I am seeing with different engines, but I
thought I would comment on Peter's benchmarks on PBXT:
http://www.mysqlperformanceblog.com/2007/04/08/pbxt-benchmarks/

This is a bit off what I have found so far:
Picture 2.png

Picture 3.png

Picture 4.png

Picture 5.png



In all tests but the final, the task was to split a load among many connections and see how the engine behaved. PBXT's write performance is quite good. Its update crashed on me before I got to see how well it would perform and the scan read was awful. On the other hand its performance with mixed IO is quite good. For an engine at this point in its integration, these are all promising numbers. I've got little doubt about PBXT being an engine that people will use in production.

All numbers were performed on 5.1, and I used PBXT as a loadable engine (aka it was not statically compiled in).

If you are curious about the script I used to generate these numbers you can grab it here:
http://hg.tangent.org/bbench?ca=tip;type=gz

Link | Leave a comment | Share

Comments {8}

??????

from: burtonator
date: Apr. 9th, 2007 03:42 pm (UTC)
Link

How is it possible that INNODB is faster than MEMORY here?

I assume this unit of work fit in the INNODB row cache and that memory tables have global locks that were slowing down performance?

Kevin

Reply | Thread

Re: ??????

from: burtonator
date: Apr. 9th, 2007 03:44 pm (UTC)
Link

nevermind..... I'm a complete idiot.... the y axis is time. Peters was operations per second..... Sorry.. I literally just woke up !

Note to self. Never again read benchmarks when you're asleep.

Reply | Parent | Thread

Brian "Krow" Aker

Re: ??????

from: krow
date: Apr. 9th, 2007 04:08 pm (UTC)
Link

When I read your comment I suddenly wondered if I had put in the wrong benchmark since I do have one where this was the case (it was a failing test).

Reply | Parent | Thread

peter_zaitsev

Re: ??????

from: peter_zaitsev
date: Apr. 10th, 2007 03:06 pm (UTC)
Link

A fun note, you can get HEAP to be slower than Innodb in some cases but not this general stuff. For example VARCHAR(50000) make HEAP table quite nasty and there are few other things.

Reply | Parent | Thread

peter_zaitsev

(no subject)

from: peter_zaitsev
date: Apr. 10th, 2007 03:08 pm (UTC)
Link

Thanks Brian,

I have not yet checked your benchmark script to comment a lot, just wanted to mention you should keep into account PBXT does not flush neither logs nor data to the disk at all.

Reply | Thread

Brian "Krow" Aker

(no subject)

from: krow
date: Apr. 10th, 2007 08:03 pm (UTC)
Link

I had suspected that it didn't... the numbers are just too good to be true. Did you look at the code to determine this, or...?

Reply | Parent | Thread

peter_zaitsev

(no subject)

from: peter_zaitsev
date: Apr. 10th, 2007 08:29 pm (UTC)
Link

I asked Paul about PBXT durability and he told me this.

Reply | Parent | Thread

Brian "Krow" Aker

(no subject)

from: krow
date: Apr. 10th, 2007 08:27 pm (UTC)
Link

One of things I struggle with is showing engines side by side (other then blackhole, since it just show upper level engine behavior with compare to the IO/CPU of an actual engine).

People really want an answer for what engine to use... I explain it from my own point of view, and design engines from this point of view.

What is frequently missing through is the "I need OLTP, what should I do". I think the answer is easy at the moment, everyone but Innodb has on training wheels, but I think this will become more difficult over time.

What have your thoughts on this been?

Reply | Parent | Thread