Recently, a major French bank asked Tweag to help fix a performance problem they had in a large Haskell codebase. In this blog post, we show the technique we used to locate the performance bottlenecks in that codebase. Once they were located, we were able to address them and eventually obtain a speed-up of two orders of magnitude. This post also includes a repository with sample code profiled using the method described here, with all runnable commands in a Makefile—so you can easily experiment.
Preamble: cost centers and inlining
GHC can mark some regions of code as cost centers, which are the fundamental units in profiling reports. During the runtime of a profiled executable, it measures and reports the time and memory spent on evaluations of each cost center.
Since these pieces of code marked as cost centers must be clearly identifiable at runtime, they have a significant restriction—they cannot be inlined. Consequently, they cannot be optimized away.
Therefore, if we want to be sure we are profiling the same code we are running using a standard optimized binary, we must not mark functions that should be inlined for optimizations as cost centers. Since GHC is not aware before compilation of which pieces of code will be inlined or not, cost centers must be marked manually, rather than automatically. Since automatic addition of cost centers is the default, care must be taken.
Marking cost centers manually also implies that library dependencies won’t be profiled at all. That’s not a problem, since the goal of this technique is to locate performance bottlenecks, not to diagnose them. If you want to single out some function from an external library, it suffices to bind it to some local variable and mark this local variable as a cost center.
Example code
Here’s the code we are going to profile.
It consists of a small calculator with operations on integers (Op
).
module Main where
import Data.Foldable (foldl')
import qualified Data.Vector as Vector
main :: IO ()
main = do
let
pair n = [Add n, Sub n]
ops = concatMap pair [0..10000]
result = applyMany ops 0
print result
data Op = Add Int | Sub Int
deriving (Show, Eq)
addSlow :: Int -> Int -> Int
addSlow n k = Vector.foldl (+) k (Vector.replicate n 1)
{-# SCC addSlow #-}
sub :: Int -> Int -> Int
sub x y = y - x
{-# SCC sub #-}
apply :: Op -> Int -> Int
apply (Add n) = addSlow n
apply (Sub n) = sub n
{-# SCC apply #-}
applyMany :: [Op] -> Int -> Int
applyMany ops k = foldl' (\n op -> apply op n) k ops
Notice that we marked the functions apply
, sub
, and addSlow
with SCCs (Set Cost Center). The
reason for this is double:
- These are the functions we wish to observe, so they must appear in the profiling report.
- Inlining these functions would not significantly affect the performance of our software, so it is safe to mark them as SCCs, preventing inlining.
It’s not easy to give general advice for identifying functions that are safe to mark as cost centers. However, some pointers are:
- Recursive functions are never inlined, so marking them as an SCC should be safe.
- Functions that are subject to fusion, such as consumers and producers of lists and vectors, are very prone to be optimized away via rewrite rules, so much care should be taken when profiling them.
Application of the method
The naive method (and easiest) is to use the default profiling methods from
Cabal or Stack, which add cost centers automatically when compiling with
profiling activated. We want to avoid it, so at compilation time we pass the
GHC flag -fno-prof-auto
to the executable and its dependencies (see the
appendix). For Cabal one only needs to make the flag explicit:
cabal build --enable-profiling --ghc-options="-fno-prof-auto"
For Stack, one needs to add the following flags at the top-level to
stack.yaml
:
apply-ghc-options: everything
rebuild-ghc-options: true
and build with:
stack build --profile --ghc-options="-fno-prof-auto"
In all of these methods, the executable should be run with example +RTS -p
in
order to produce the .prof
file, that can be analyzed with many tools, like
Flamegraph or Profiterole.
To show this choice of flags makes a difference, these are the runtimes of the executable above in three situations:
- No profiling: 0.16s
- Profiling with default flags: 6.16s
- Profiling with overridden flags: 0.45s
The executable with profiling is always a little slower due to the runtime cost of measuring and registering, but it should never have a different complexity from the original executable: the difference should always be, at most, linear.
We also see the difference on the profiling results in the .prof
file. When
using the default flags, we see, after simplifying the file a little:
individual inherited
COST CENTRE MODULE entries %time %alloc %time %alloc
CAF:main1 Main 0 0.0 0.0 100.0 100.0
main Main 0 0.0 0.0 100.0 100.0
main.result Main 1 0.0 0.0 100.0 100.0
applyMany Main 1 0.0 0.0 100.0 100.0
applyMany.\ Main 20002 0.0 0.0 100.0 100.0
apply Main 20002 0.0 0.0 100.0 100.0
Main.apply Main 20002 0.0 0.0 100.0 100.0
addSlow Main 10001 0.0 0.0 99.9 100.0
Main.addSlow Main 10001 0.1 0.0 99.9 100.0
>>= Data.Vector.Fusion.Util 100020001 84.9 83.0 98.6 94.0
basicUnsafeIndexM Data.Vector 49995000 13.6 11.1 13.6 11.1
...
which results in the following flamegraph:
Here we see that the function addSlow
is slow as expected, however,
most of the cost is in the vector
library, in (>>=)
. Why?
Because replicate
failed to fuse with foldl
in the code above, making
the slow code even more anomalous.
If we use the correct GHC flag instead, we see:
individual inherited
COST CENTRE MODULE entries %time %alloc %time %alloc
CAF:main1 Main 0 0.0 0.0 100.0 100.0
Main.apply Main 20002 0.0 0.0 100.0 100.0
Main.addSlow Main 10001 100.0 100.0 100.0 100.0
Main.sub Main 10001 0.0 0.0 0.0 0.0
which results in the following flamegraph:
This is much more straightforward, showing that addSlow
is the culprit
that should be corrected, without introducing spurious information to mislead
our profiling.
Conclusion
This profiling technique is simple and should produce fairly predictable results. This predictability is by design—code that can be inlined and be optimized away should never be marked as a cost center, making the profiled executable performance very similar to the regular executable one. This makes sure that the profiling is faithful. This was essential for the codebase of this French bank, since the default profiling report always pointed to the wrong place, wasting time and resources.
This method works well with large codebases—one can start by profiling key functions at the top level, and little by little add cost centers to locations that are deeper in the file hierarchy, up to the point where it finds the actual bottleneck.
Since this method requires very precise flags to work, one must be sure that these are being passed correctly to the repository at hand. If even one dependency gets cost centers automatically attributed, it can compromise the whole profiling result. However, this should be easy to diagnose, since one would see the dependency code in the profiling report.
Remember to check out the repository we included—so you can easily experiment with the code and commands above!
About the author
If you enjoyed this article, you might be interested in joining the Tweag team.