The GHC profiler is today one of the most advanced profiling tools for analyzing how Haskell programs use CPU time and memory. It produces information on what parts of a program could use optimizations to speed it up.
Unfortunately, the very definition of the GHC profiler makes it of limited use when estimating time on two classes of computations: firstly, those that need to do blocking IO; and secondly, some computations that invoke functions written in other programming languages. In these cases, one could just turn to a tool like ghc-events-analyze, however, it is not available to plugins loaded in GHC 8, such as Liquid Haskell.
In this post I’m discussing a bit the limitations of the GHC profiler, and I’m announcing timestats, a simple profiling library that you might want to grab when everything else fails.
The limitations of the GHC profiler
Consider the following program from a GHC issue:
main = g
g = getContents >>= print . length
If a person interacts with this program in a terminal, the program
is going to spend most of the time blocking while waiting for input.
The GHC profiler, however, won’t account any time to function g
.
This is because the profiler doesn’t estimate wall-clock time but
CPU time and, so far, only for the Haskell portion of a program. As
function g
uses little CPU, g
gets little time attributed to it
in the reports of the profiler.
The example program is contrived but the implications are far reaching. It is not uncommon for programs to block when doing IO to communicate with a database, a remote server, a file system, or another local process. The larger the program is, the more likely many of these blocking IO calls are occurring during its lifetime.
The other class of problematic computations can be exemplified with:
{-# LANGUAGE CApiFFI #-}
foreign import capi unsafe "stdio.h getchar" getchar :: IO Int
main = g
g = getchar >>= print
Much like in the previous case, the program will spend most of the
time waiting for input. The important difference, though, is that the
program is calling into C first via the
Foreign Function Interface, and only then blocking for input.
According to our assumptions this far, this shouldn’t make a
difference in the results. The program uses little CPU time, so g
would not show any time attributed to it. This is not what happens
though!
Haskell programs can call functions written in other languages using the so called foreign functions. Foreign functions come in two flavors: safe and unsafe. The meaning of the flavors doesn’t affect the discussion here, but it is necessary to note that the foreign function in our example is unsafe.
It turns out that the GHC profiler estimates CPU time for Haskell computations. But if a computation calls an unsafe foreign function, it will switch to estimating the wall-clock time of the foreign call (!) and attribute it to the calling computation1. The main consequence of this behavior is that interpreting the results of profiling requires knowledge of the call graph to learn which of CPU or clock time is being measured at different places. The bigger the application, the harder it becomes to interpret the results.
timestats
timestats
is a library that measures the time it takes to execute
selected fragments of a program. It requires the program to be instrumented
with calls that identify these fragments, and then relies on the
function getMonotonicTimeNSec
to measure the execution time.
It features a measureM
function:
measureM :: MonadIO m => String -> m a -> m a
which associates a label with a monadic computation, measures
the time it takes to perform the computation, and records it in a hidden
and globally available location. If multiple measureM
calls use the
same label, their time measures are added together.
import qualified Debug.TimeStats as TimeStats
a_task :: IO ()
a_task = TimeStats.measureM "a_task" $ ...
another_task :: IO ()
another_task = TimeStats.measureM "another_task" $ ...
main :: IO ()
main = do
a_task
a_task
another_task
TimeStats.printTimeStats
In the above example we have instrumented functions a_task
and
another_task
. Whenever they are invoked, we now collect timing measures.
At the end of the program, we added a call to print the measures that have
been collected so far. It produces an output like the following when
measuring is enabled.
a_task: 2.055s count: 2
another_task: 3.071s count: 1
By default, no measures are collected, though. None of the calls coming
from timestats
are meant to change the observable behavior of the
program unless the environment variable DEBUG_TIMESTATS_ENABLE
is set
to any value.
timestats in a larger program
I developed timestats
when profiling Liquid Haskell,
a GHC plugin that interacts with SMT solvers
that run in external processes. I needed to understand how much time
was spent in different stages of Liquid Haskell.
Besides the limitations with blocking IO, the GHC profiler posed the challenge of making the profiling runtime available to Liquid Haskell when loaded in GHC. GHC is not linked by default with this runtime.
In addition, Liquid Haskell support for GHC 9 is partial still, and the
profiling work was to be done on GHC 8. Now, GHC 8 is not delivered with
the eventlog variant of the runtime system, which is necessary to use
ghc-events-analyze
.
In contrast, with timestats
I could arrive at the following summary
by inserting a handful of calls. These numbers are the aggregates of
processing multiple modules with Liquid Haskell.
filterValid_: 6.611s count: 1532
ple: 2.135s count: 5
refine: 7.713s count: 15
solveCs: 20.059s count: 15
solveNative': 16.426s count: 15
typecheckHook: 24.995s count: 16
typecheckHook
is the top-level call of the Liquid Haskell plugin.
filterValid_
captures much of the time that is spent communicating with
the SMT solver. ple
is an algorithm that has undergone
much optimization lately, to the point that it has
become rather negligible in this example.
refine
captures the time spent on another verification
algorithm . solveCs
and solveNative'
stand for calls that deserve
further exploration.
For the insight that these numbers provide, it is quite a return in exchange for inserting six instrumentation calls!
Closing remarks
timestats
is as basic to time profiling as print statements are to
debugging. It is a last resort when more sophisticated tools are out
of reach or otherwise inconvenient to use.
You are welcome to use timestats
in your projects, hack it to your
needs, and share your experience.
In addition, I haven’t found other resources yet that discuss the limitations of the GHC profiler1. I hope this post helps to raise awareness so people can veer early for the alternatives whenever appropriate. Happy profiling!
- Since the time of this writing the profiler documentation
has been expanded to explain the interactions of the profiler
with foreign calls. Because an unsafe call isn’t expected to block, the
profiler continues to track the time that the Haskell runtime
waits for the unsafe call to complete. In this light, the profiler is not
mixing CPU and wall-clock time, but estimates CPU time mostly spent
in the Haskell portion of a program, and
timestats
stays as an alternative to measure wall-clock time instead. Special thanks to Matthew Pickering and Duncan Coutts for improving the documentation.↩
About the author
Facundo is a software engineer supporting development and research projects at Tweag. Prior to joining Tweag, he worked in academia and in industry, on a varied assortment of domains, with an overarching interest in programming languages.
If you enjoyed this article, you might be interested in joining the Tweag team.