Wednesday, May 27, 2009

MySQL 5.4 testing - Failure and a W/O

I had this idea around 5.4. It is meant to scale better on multi-core machines compared to, say, standard 5.1. This has been proven also it seems. 5.4 is still in Beta, but testing has shown that scalability is definitively better.

What I wanted to know was how well 5.4 would work on a lower spec box compared to 5.1. One reason for me wanting to test this is that I am currently in my summer house, and I have no multi-core 64-bit machines around. All I have in terms of Linux boxes is an old Dell Laptop with Gentoo on it, all this will as similar setup as possible between 5.1 and 5.4.

So, I download 5.4 sources, as there are no 32-bit Linux binaries yet, and I do a simple configure and make. And.. I fail. Out of the box, some of the inline optimizations in innodb (the config I built with was max-no-ndb) will screw up gcc when using -O3 optimization (OK, I admit, I'm not 100% sure that this is the cause, but it sure looks like it, and some googling seems to confirm it, but I'm not a compiler engineer, so who knows).

So today, when I was planning to do some benchmarking and some other fun things, I tried to track this one down, and I did find a W/O for this issue, which is to configure with CFLAGS=-O2, i.e.:
./configure --with-plugins=max-no-ndb CFLAGS="-O2"
This will actually do the trick. And once I figured that out, I had to report it as a bug (#45131) and try with a few different compilers.

I did an emerge on gcc for my gentoo, and got gcc 4.1.2 instead of 4.1.1, and had the same problem. I also tested on another box running SuSE and gcc 3.3.1, and with the same problem. But as I said, at least there is a fix.

Tomorrow, I might do the actual low-end 32-bit machine performance test, I'll keep you posted.

/Karlsson

4 comments:

Mikael Ronstrom said...

Weird that you get this problem. I also came across this problem in some pushbuilds so disabled inlining on 32-bit gcc since there was some issue with this (also a similar problem on Mac OS X).

Karlsson said...

Also note that this was on several platforms, who machines (both 32-bit x86) and three versions of gcc.

Karlsson said...

This bug is now confirmed as a real bug. It was actually a bug in the code it seems: 45131

Mikael Ronstrom said...

It turns out that the fixes I did was only disabling the 32-bit gcc inlining on Mac OS X and Solaris. As you showed it was a problem with 32-bit inlining in general. But as showed this was due to a bug in the atomics code, so with that code fixed I can also remove the disabling of inlining of 32-bit on Mac OS X and Solaris. Great, thx for blogging about this and filing a bug.