An inconvenient truth
One of the things every software developer has to learn the hard way is that performance optimizations are a bad thing most of the time. The lesson is counter-intuitive because it is in conflict with several fundamental motivations of our artist/engineer attitude:
- We want our code to be as fast as possible or even faster. (We really like the prospect of making the impossible happen!)
- We don’t want to waste resources if it can be avoided. (Digital resources have always been scarce. Development in abundance wasn’t taught!)
- We want to be clever and one-up everybody with the latest trick/hack. (This ego boost driven development attitude is a major problem on its own!)
But we can’t deny that clever guys exist since forever and that their wisdom should be considered. Here’s one clever guy fourty years ago:
premature optimization is the root of all evil.
said Donald Knuth in 1974. A typical computer had RAM in the lower kilobyte range these days and clocked with around a megahertz. The Intel 8080 was introduced in this year. No sign of abundance all around.
Coming from this insight, i teach the “three rules of performance optimization” to my students:
- Not Yet
A little disclaimer: Performance optimizations are measured in milliseconds. You are still responsible for complexity optimizations and should engage them. Complexity is measured in O(n).
I blogged about the rules some time ago, so I won’t repeat the whole meaning behind them. Today, I want to tell a story related to rule three (“Measure”) and why it’s generally a good idea to stay clear of optimizations if not absolutely necessary.
An accidental observation
In one of our long-running projects, we store data in a 24/7 manner since more than ten years. The project started on Pentium IV boxes with 120 GB harddisks. Soon enough, the available disk space vanished rapidely. Our customer wanted to optimize the storage efficiency. We told him that we can always trade storage space versus computation time or the other way around (the infamous space/time tradeoff), but if we compress the data in the system’s archive to save space, it will result in slower archive access. Our customer understood the tradeoff and decided he wanted storage efficiency over access speed for the archive.
So we added a compression step when certain data types were stored in the archive. Because the data was text based (XML and other formats), the compression rate was at 90% and even higher, meaning that we could fit ten times more data on the disk than before. We certainly met the customer’s goal of store efficiency. But what about the access speed? We needed to add the decompression step for certain data types at the place where our system loads from the archive. After that, we hoped that the speed didn’t suffer too badly. We measured the access times before and after the change and couldn’t believe our eyes: The access time was nearly ten times faster than before. There was no tradeoff, we actually shrank memory consumption and computation time at once and in the same ballpark figure. We felt like heros.
Discovering false premises
Why did this happen? Our explanation is that at some specific point in time, the CPUs became too fast for a tradeoff. In the good old days, there really was a tradeoff: if you needed to compress parts of your harddisk to save space, loading these parts became slower. Because the CPU had to perform extra work on top of the loading, you had to wait longer. Loading more data directly from disk was faster than loading less data from disk and decompressing it. When the CPU became fast enough to actually decompress during the I/O delay of the harddisk, that’s when the raw amount of data that needed to be transferred from disk determined the loading speed. And since uncompressed data is larger, it now took longer to load than compressed data.
At some specific point in time, the old wisdom that compressed data is slower became a lie. Nobody told us. We weren’t heros, we just lived on false premises.
Our customer was pleased and continued to be pleased for years. Every few years, the CPUs of the target machines would become even faster, but the harddisk performance would hardly improve. The new wisdom was truer than ever: Less data is faster, regardless of what the CPU has to do with it. This was a performance optimization that could be used everywhere. You want to increase I/O speed? Compress the data.
The tables begin to turn
Fast forward to modern days. A new technology promised to improve harddisk performance by orders of magnitude: The solid state disk (SSD) delivers impressive amounts of data per second, but more important, gets rid of the initial seek time that magnetic spindle disks had (those few milliseconds to locate the data on the disk that felt like ages for modern CPUs). We started our migration to SSD in 2010 and were SSD-exclusive at the end of 2013. Our customer was a bit more hesistant, but the latest generation of target machines run on SSDs, too. So how did this affect our performance optimization?
As you can probably deduce by now, the decompression work of the CPU re-emerged in the archive access time. It’s by no means as bad as in ancient times, but the days of “no tradeoff” are gone, again. The performance optimization isn’t an optimization anymore. We still have it in the system, because it was never meant to improve performance, but storage efficiency, and it still does that perfectly. And needs to, because SSD are still expensive and don’t offer storage space in abundance. But the old wisdom is back (partially): compressed data is slower, if not by much.
What do we learn from this?
Over the course of a few years, a specific feature (transparent storage compression) in our system was a performance burden, then a performance booster and a burden again. We didn’t change the code, we just changed the hardware circumstances. The best lesson to be learnt is that no performance optimization lasts forever. Premises will change. Effects will be negated. Bottlenecks will be shifted. It’s best to know when that happens. And then it’s best to be able to revoke the optimization. Or, if all this sounds way too much trouble for such a tiny performance gain, remember the first rule of performance optimization (Don’t) and just leave it alone. You can always tell yourself that you’ve just optimized your code for the machine it will run on in ten years.