Xaprb

The best overall method of performance optimization is optimization for response time. Users care about response time, not load average or cache hit ratios. The job of a system is to accept some request and do the required work, and deliver a result. The time elapsed between the request and the result is the response time.

Methods of Response Time Optimization

Not all optimization methods are created equal. Here are a few I see commonly.

• No method. Most people simply have no method of performance optimization at all. They just look for things that look “bad” and try to make them look “better.” In the MySQL database world, the classic example is trying to improve a cache hit ratio. This is utter folly, and doesn’t become any less stupid no matter how many times it is taught and repeated.
• Server Load Reduction. One step up from that is to try to understand what work the database is performing and discover what part of that work consumes the most response time, then improve that to lower the load on the database server. This is better, but still not a logical way to optimize response time for the end user. Imagine that you’ve built your scaling strategy around archiving and purging unnecessary data — a very sensible strategy. Most well-designed archiving and purging jobs are terribly inefficient, for a reason: they are designed to consume resources that are not needed by anything else, so they don’t interfere with anything else. Archiving a billion rows from a table is best done in nibbles, not in billion-row chunks. The nibbles are going to be slow. If you measure the entire system and find out where the response time goes, you’re almost guaranteed to find these jobs are a top offender. And yet they don’t matter at all, because they have no impact on the user’s response time. Server load reduction is a shotgun approach that sometimes yields results, because it’s easy to aim a shotgun.
• Method R, or Goal-Driven Performance Optimization. Two methods I know of that are based in sound thinking are Cary Millsap’s Method R and Peter Zaitsev’s Goal-Driven Performance Optimization. Cary wrote an excellent book about his method, and I recommend buying that book and reading it at least twice. These methods are guaranteed to truly optimize the system in question: they will produce the best possible performance improvements with the least possible cost, and they have a termination condition that is satisfied when further improvements are either not possible or cost more than they are worth. A system that has been subjected to one of these methods can be confidently called “fully optimized.”

Response time in queued systems

Method R looks at where a system consumes time and sorts the biggest consumers to the top in a profile, then works on those first. Amdahl’s Law guarantees that this is the best way to improve the system’s performance.

Although the approach is correct, it doesn’t mean it’s easy. It might be easy if a system is stable. But unstable systems, those suffering from queueing delay (usually characterized by response time with a high standard deviation, a.k.a. “spikes” or “blips”), are much harder to optimize. The queries that are performing badly can no longer be assumed to be the source of the performance problem. Instead, they might be “good” queries that are the victim of something else happening.

Unstable systems suffer from a) dependencies between requests, and b) statistical variations in request arrival time, which causes queueing. The classic case is lock contention. Suppose someone goes to your OLTP database and runs an ad-hoc query against the table of invoice line items, and locks the table. Normally that table has specific, fast, well-indexed queries against it, but as soon as the ad-hoc query locks it, the queries instantly pile up and “look bad.” The system becomes an unstable train wreck. Alas, database servers such as MySQL typically don’t give the DBA enough information to blame the problem on a source.

Internal contention inside the database software itself is another potential cause of queueing. I can no longer remember the number of times I’ve disabled the query cache in MySQL and solved a system’s random freezes. Typically, the only way to prove that the query cache mutex is the source of the problem is to take a backtrace in GDB. A complex piece of software without good instrumentation is pretty difficult to troubleshoot in conditions like this.

And that brings me back to Method R. I can see that the queries are suffering from unstable performance, but I cannot see directly how to optimize the system because it is un-instrumented. Unfortunately, falling back to system load optimization is often the best that can be done, in terms of maximum optimization with minimal cost. An expert can do this with as much rigor as possible, and hopefully with good knowledge of the system’s internals can find the source of the problem quickly, but it’s still a much harder problem.

And this is why it is a crime to write un-instrumented software: because when an un-instrumented system starts to get overloaded, it is very hard to determine the source of performance problems.

Optimizing Oracle Performance

Optimizing Oracle Performance. By Cary Millsap, O’Reilly 2003. Page count: about 375 pages with appendices. (Here’s a link to the publisher’s site.)

This is easily one of the best books I’ve ever read on performance optimization. I’ve just finished reading it for the second-and-a-half time in two weeks, and I very rarely read a book more than once. I’ve been telling a lot of people about it.

Despite the title, it is actually not about Oracle performance. It is a book on how to optimize a) any system, including a MySQL-based application b) the process of optimization itself. It is very, very good, and I highly recommend it to all database users. My bet is that most people will learn more by reading this book than by spending thousands of dollars on conferences and training, especially since a lot of what you’ll learn from those sources is wrong and harmful. So not only will you save money and time on learning, you’ll reap great rewards thereafter.

The book is organized into three sections. The first section explains a performance optimization methodology called Method R (for response time) that is designed to deterministically advance from identifying what needs to be optimized, to collecting diagnostic data, and then choosing the correct activities to optimize. Simply put, it is probably the clearest and most logical process I’ve ever seen for focusing on what matters. It is quite similar to our process at Percona, which we call Goal-Driven Performance Optimization. But Cary does a really good job explaining it. Cary told me that he wrote the book while simultaneously developing Method R training classes, and did dozens of classes before the book went to press. You can tell!

The second and third parts of the book are about putting the method into practice.

Cary shows many typical mistakes, such as focusing on ratios, working on things that “look bad,” ignorance of Amdahl’s Law, trying to draw conclusions about specific activities by examining system-wide information, and so on. He brings the focus back to response time again and again.

Another typical “tuning mistake” is not knowing when to quit. Cary shows how to know when further performance improvements, even if they’re possible, will not be economically justified. At the point where the gains don’t exceed the cost, you’re done. The system is optimized. Maybe it’s not performing as well as it could, but if it costs too much to get any more performance, it’s still optimal.

This is a book that backs up every assertion with references to source material, or even with a proof or derivation from first principles. It is significantly more rigorous than most books of this type. There’s a very good chapter on kernel timings that explains a ton of stuff about how processes work, measurement intrusion effect, quantization error, and so on.

There is a very good section on queueing theory, with excellent examples of using it to prove that a desired improvement is mathematically impossible and/or impractically expensive. There are also examples of how to use queueing theory to predict when a hardware upgrade will worsen your performance problems. (And lots of other examples of when “optimizing” something will actually make it worse, as for example improving performance of some task other than the one you’re interested in, thereby making it use and contend for more resources — and hurting performance of the task you wanted to improve.)

What is it NOT about? It’s not about how to write more efficient SQL — there is practically nothing in the book about SQL. It’s not about how everything inside Oracle works, though it does have one or two chapters that are mostly about how to apply the principles to Oracle through extended SQL trace files, a discussion that will transfer well to any other system that emits trace data. (Read these, or you’ll miss learning about quantization error and such!)

Highly, highly recommended.

Having written about what I think is cool about the upcoming MySQL Conference and the MySQL Camp, now I want to finish up with what I’d like to see at the Percona Performance Conference. Just to recap, this is a conference we created to serve those who want to learn about performance — not “learn about MySQL,” not “learn about database performance,” just learn about performance, period.

I want to see everything. I think this is going to be the single best conference I’ve ever been to. Even the way the conference is organized is exciting. For example, it’s running from early morning till late at night, nonstop. The sessions are also (mostly) only 25 minutes. This means if you decide a session isn’t all that interesting, you didn’t spend much time on it, and you don’t have long to wait for the next one.

So here is a small sample of the sessions:

• CouchDB: Behind the Buzz (Jan Lehnardt)
• Performance Instrumentation: Beyond What You Do Now (Cary Millsap)
• Hive: Distributed Data Warehousing with Hadoop (Ashish Thusoo and Prasad Chakka, Facebook)
• High Performance Erlang (Jan Henry Nystrom)

These are not just people who’ve learned about something and want to talk at you. These are the inventors, the originators, the gurus. It is truly the who’s who, and that’s just a few of them. If you aren’t familiar with those names, Google them and see. And after that, why not Google Theo Schlossnagle, Eric Burton, Monty Widenius, Andrew Aksyonoff, and a few others.

I hope to see you there. Bring your business cards and introduce yourself to me!