Mythbusters: How to configure InnoDB buffer poll on large MySQL servers

Yesterday I wrote about the dangers in using top on systems with 100+ GB of RAM, not to mention future systems with 1+ TB. A related topic is, how should I configure MySQL on such a large system?

There is a classic rule of thumb that on a dedicated MySQL server one should allocate 80% of memory to the InnoDB buffer pool. On a 128GB system that is 102.4 GB. This means that I would leave 25.6 GB of RAM "unused". So surely on these large systems, this old piece of advice cannot hold anymore. If the database was previously running on a server that in total had less than that altogether, it seems wrong to leave so much memory just unused. Let's label the old rule of thumb tentatively a "myth" and ask mythbusters to figure out a new MySQL configuration for us...

We all know that a megabyte in binary system is not the same as one million bytes (in decimal system). But have you actually cared much about it? I have to admit I haven't. I know there is a small rounding error, but by and large I always treated 2^10 = 1 kB = 1024 bytes and 10^3 = 1 kB = 1000 as the same thing. (Update: Opening sentence was edited to remove units MB and MiB since it seems even I managed to use them backwards! The math in this article is correct. The rest of the article uses MB, GB and TB mostly to refer to binary magnitudes, which is apparently incorrect. See comments for wikipedia links and discussion.)

More importantly, when you move into larger numbers, rounding errors usually become even less important. Unfortunately, in this case they become bigger:

There was a guy on TV (and unfortunately I've forgotten his name) who does crisis management and peace negotiations. You know, mediating when two countries or tribes have been at war. At the opening of peace negotiations he would usually tell the following story:

A week ago Baron wrote a blog post which can only be described as the final nail in the coffin for MMM. At MySQL AB we never used or recommended MMM as a High Availability solution. I never really asked about details about that, but surely one reason was that it is based on using the MySQL replication. At MySQL/Sun we recommended against asynchronous replication as a HA solution so that was the end of it as far as MMM was concerned. Instead we recommended DRBD, shared disk or MySQL Cluster based solutions. Of course, to replicate across continents (geographical redundancy) you will mostly just use asynchronous replication, also MySQL Cluster used the standard MySQL replication for that purpose.

Yesterday I posted results from some MySQL benchmarks I had been doing on a large server. In this post I'd like to list 4 important fixes that were done to avoid bad performance:

DRBD and semi-sync benchmarks on a 2x8 132 GB server

I recently had the opportunity to run some benchmarks against a relatively large server, to learn how it was behaving in its specific configuration. I got some interesting results that I'll share here.

In March I posted a series of blog posts on my paternity leave MepSQL project, which I called MepSQL. There was still one piece created in the MepSQL buildsystem that I didn't publish or blog about. Since it is generally useful, I wanted to generalize and polish it and publish it separately. I finally had that done last week, when I also found that somebody else, namely alestic.com already published a similar solution 2 years ago. So yesterday I ported my BuildBot setup to use that system instead and am happy to publish it at the Open DB Camp 2011 in Sardinia.

Ok, so let's go back a little... What is the problem we are solving?

Let's refresh our memory with a picture (and you can also go back and read about it):

Some more recent readers of this blog perhaps never realized that it is a spin-off of a book I wrote years ago. I still write occasionally about open source business models and community stuff, but perhaps that is drowned by increasing amount of hard core database topics.

Reader Andrew Moore - who happens to be my collague at Nokia and a great DBA - has submitted a Kindle version of the book. I've now added it to the download page.

Thanks Andrew!