Today I was working on a client project when a seemingly innocent refactoring made the program 2x faster.
Of course, being happy about the improvement and going on with my life would have been wrong, as random performance improvements almost always mean that something foul is at play.
I undid the refactoring step by step, until the only change remaining was that I replaced a use of the
forever function by implementing forever itself, like this:
myForever :: (Monad m) => m () -> m () myForever f = do f myForever f
How could using this function make my client's application 2x faster?
forever's code, it is basically:
myForever :: (Applicative f) => f a -> f b myForever f = f *> myForever f
Doing a couple more tests:
myForever :: (Monad m) => m a -> m a myForever f = f *> myForever f
is slow, but
myForever :: (Monad m) => m a -> m a myForever f = f >> myForever f
So the issue seems to be that
*> is 2x slower than
>>. Not great.
Of course the next question is: What Monad
m am I running on? In my program I'm running
StateT SomeStuff IO ().
I was promptly reminded by a colleague that I shouldn't be using
StateT over IO as it makes exception handling and concurrency very hard, and that I should use a
ReaderT instead, but since the program dealt neither with exceptions nor concurrency, that shouldn't be relevant here (more on it later).
Digging further into this performance difference, I found that it went away when upgrading
0.5.3.0. Looking at the
transformers changelog, the point
* Added specialized definitions of several methods for efficiency
sounded relevant. I then found the transformers issue
*> must be defined in instances to prevent space leak with forever, which reports that
forever, used with
StateT s IOor
ReaderT r IOhas a space leak
The linked GHC ticket explains that it was introduced when
base-18.104.22.168 (for GHC 8.0) switched
forever to use
Applicative. But here I sit, 15 months later with GHC 8.2, spending hours debugging this issue for my client.
This is pretty bad.
How can a performance regression like this sneak into a core library that everybody uses, triggered by a function as fundamental as
forever? And the problem also applying for
ReaderT makes it even more widely spread. This reminds me of another similar bug from 4 years ago, where I found that some operations in the
vector package (also fundamental and widely used) had become 5x slower without anyone noticing, that time because of some C-preprocessor-macros accidentally being undefined due to missing header includes.
As a community we make a big fuss about how Haskell allows you to write correct code and how refactorings are easy and safe. But we're not promoting a good image of Haskell here. In other programming languages I rarely encounter performance issues in parts that are this fundamental and this widely used. Not having major performance regressions is also a form of correctness.
Right now, as a community, we are simply bad at not introducing performance regressions. If we want Haskell to grow, we must do something about that.
So, who is to blame?
- The GHC developers for pushing the generalisation from
foreverand other functions, a change that has been desired and requested by pretty much all Haskell users for years? Looks like they did everything right. Without such changes, Haskell will stagnate and we'll have to live with historical ballast, forever.
transformersmaintainer, for immediately applying a fix after being shown the problem, in response to code that he did not change? Looks like he did everything right.
It looks like we cannot blame either of the two parties actively changing code.
I think the problem is: Lack of performance tests.
If our tools do not allow us to reason trivially about performance (rewrite rules unnoticeably not firing or being commented out, header files unnoticeably not being included, non-breaking fundamental changes accidentally introducing space leaks and huge run-time differences), then we must not rely on human inspection (people reading diffs) to immediately spot these issues.
Instead, we must write tests to automatically verify that we're not regressing the performance of core functionality. Such tests should:
- ensure that the time spent in commonly used functions stays the same,
- ensure that the space complexity doesn't accidentally change (for example,
StateTshould not space-leak),
- be ubiquitous in the core libraries we all use.
A good example of this approach is the
nofib test suite in GHC, that exercises common use cases of the compiler, and, via Continuous Integration and automatically run tests, notifies the developers if they made something slower. Unfortunately this approach is extremely rarely seen in any Haskell libraries higher-level than GHC, and tools that make it easier (such as
weigh) are recent developments.
Preaching performance tests when they are difficult to implement isn't exactly useful. To put my (or rather, FP Complete's) money where my mouth is, I have just published a new library
cpu-instruction-counter, which can measure the CPU instructions executed by a piece of Haskell code. I encourage you to use it to guard your code against accidental performance regressions, and it would also be great if we started using these approaches directly in the test suites of
transformers and so on.
By pushing more into this direction and adding performance tests to fundamental libraries, we can keep improving these fundamentals, such as moving functions from
Applicative, and be sure to get notified early if we accidentally break things, instead of wasting productive days on evil surprises one and a half years later.