performance optimization: combating the evil

Posted by anton
on Monday, December 31, 2007

face the enemy

We should forget about small efficiencies, say about 97% of the time: premature optimization is the root of all evil1

continuing the topic of my previous post (do not deploy black boxes, allow for measurement, metrics, monitoring), i offer you my two rules of performance optimization (heavily optimized for enterprise environment2).

rule #1: don’t do it

i’ve seen more crimes committed against software in the name of performance optimization than for any other reason3.

don’t bother with performance optimization – deliver working software first.

instead of spending time salivating over sexy distributing caching algorithms and debating the merits of lock striping approaches, implement by feature and most likely you will find out that performance is good enough.

a fitting quote from Refactoring:

The secret to fast software, in all but hard real-time contexts, is to write tunable software first and then to tune it for sufficient speed.

i know it is hard to resist the glimmering image of a performance superhero squeezing out a dramatic 100000x speed increase. we are all guilty of dreaming about it. face it – you are not writing hard real-time systems. you’ll just have to buckle up and stick to implementing business functionality without getting to play with those exciting computational problems.

finally, know your requirements upfront – what latency can you actually tolerate, does it even matter? what throughput do you need? make sure you have the actual numbers; both average and worst case scenarios. guess what, in many cases it turns out that you do not need to optimize to begin with.

after all, performance optimization is always about trade offs – leave your options open.

rule #2: don’t do it blindly

measure, then optimize. most of the software spends most of its time in just a fraction of the code – the “hot spot” – find it and optimize it away.

having a well-factored program leads to hotspots that are easier to isolate and optimize.

way too often i see people diving in and tweaking things left and right, just because they think they know where the problems are.

at best you will waste your time, but most likely you will actually make things worse.

if you live in the java platform world, you are in luck – there are so many tools out there for modern JVMs – use them!

notes

rules are meant to be broken, but i’d rather overreact upfront to discourage frivolous optimization.

yes, yes – you have to have basic knowledge so that you do not do stupid things all over the place and end up bleeding to death from a thousand cuts. luckily, in enterprise software most of these things are very basic – a few language rules, and a few design rules – good software developer will follow them automatically.

why performance optimization is so alluring?

here’s my take on it – enterprise software is boring. you’ve done it a few times, and you do not feel like cranking out the same stuff over and over again.

so you start creating complexity to entertain yourself, to give your mind something to chew on: you fall in love with design patterns, you build beautiful multi-tier distributed designs with transactional semantics all over, and you fiddle with performance optimization on every step.

most of us have gone through it; it is like a childhood disease that you suffer from in order to become immune. most of us survived and gained valuable insight in the process. it does take a bit of self-reflection and experience to realize this though.

(those that did not survive ascended to the stratosphere and became raving zombies – and we all know what to do with zombies).

i think the key is to understand that although enterprise software is boring, the vast majority of all software projects still fail. this is where the true complexity is – figuring out how to deliver working software that customers actually use. not to mention doing it on time, on budget, and without burning your team.

this is much harder and at the first sight a lot less sexy, but there are still plenty of technical challenges to work through in order to create well-engineered systems. it is just your definition of “well-engineered” that has to change.

once you re-adjust your focus, the work is cut out for you.

it is tempting to pitch “enterprise” software against the opposite “swing” championed by the pragmatic programmers, rails, and the whole community around it. and of course, it is not the technology but the mindset.

1 donald knuth paraphrasing hoare

2 with apologies to m. a. jackson

3 with apologies to w. a. wulf

Comments

Leave a response