Flame graphs for GHC time profiles
GHC comes with a number of nice profiling facilities. Among other things, GHC can generate time profiles, a useful facility for answering the following question: "where in the source code is my program spending all its CPU time?". With the right flags turned on, GHC's RTS dumps a time profile in a
.prof file when your program exits, providing textual summary and detailed views of the program's runtime, broken down by cost centre.
However, in large programs these
.prof files can become quite hard to make sense of. Visualizing profiling data is a common problem, and one neat solution is to use flame graphs to get a high-level view of where time is spent, and why it is spent there. That's why we wrote
ghc-prof-flamegraph, a new utility useful for turning textual
.prof reports into a pretty picture (click on the image to get to the interactive SVG):
In the figure above we have the flame graph for a run of a small Haskell program, which we will describe later. Paraphrasing the description of flame graphs from the website:
The x-axis shows the stack profile, sorted alphabetically (it is not the passage of time), and the y-axis shows stack depth. Each rectangle represents a cost center. The wider the rectangle is is, the more time is spent in that cost centre or its descendants. Cost centers often represent function calls, in which case each rectangle can be thought of as a stack frame in the call stack. The top edge shows what is on-CPU, and beneath it is its ancestry. The colors are usually not significant, picked randomly to differentiate frames.
Notice how the generated SVG image is interactive. Hovering over a stack frame gives us more information about it, and double clicking on it we can drill down that particular code path.
Installation is easy:
$ cabal install ghc-prof-flamegraph
You'll also need the FlameGraph scripts to produce SVG files. I will assume that the
flamegraph.pl script is in the
$PATH, but it can also be called from some other location.
(Example taken from from http://jaspervdj.be/posts/2014-02-25-profiteur-ghc-prof-visualiser.html)
$ ghc --make -auto-all -prof -rtsopts binary-trees.hs [1 of 1] Compiling Main ( binary-trees.hs, binary-trees.o ) Linking binary-trees ...
Then we run it enabling time profiling:
$ ./binary-trees 15 +RTS -p -RTS stretch tree of depth 16 check: -1 65536 trees of depth 4 check: -65536 16384 trees of depth 6 check: -16384 4096 trees of depth 8 check: -4096 1024 trees of depth 10 check: -1024 256 trees of depth 12 check: -256 64 trees of depth 14 check: -64 long lived tree of depth 15 check: -1
Which will generate
binary-trees.prof. Now we can use
ghc-prof-flamegraph to convert it into a format understandable by
$ cat binary-trees.prof | ghc-prof-flamegraph > binary-trees.folded
and finaly use
flamegraph.pl to convert it to an interactive SVG image:
$ cat binary-trees.folded | flamegraph.pl > binary-trees.svg
The result is shown at the beginning of the post. Note that
flamegraph.pl assumes the data is derived from sampling the execution of the program, and thus
ghc-prof-flamegraph uses a fictitious numbers for the number of entries of each stack frame, derived from the individual time as reported in the
A larger example
Let's scale this up to a larger application: consider this
.prof file, resulting from running
hoogle generate, and the resulting flame graph:
Looking at the flame graph we are immediately able to understand the two code paths that take the vast majority of the time:
General.Store.storeWriteFile. We are then able to drill down on each path by double clicking on it to explore where time is spent in detail. On the other hand, if we want to examine the
.prof file directly, we can quickly identify the hotspots:
myParseDecl Input.Type 29.3 21.8 writeItems.\.\.bs Output.Items 22.9 21.9 pretty General.Util 13.5 15.4
but we need to manually chase down their occurrences in the
.prof file to understand where these functions are being call:
myParseDecl occurs twice,
writeItems.\.\.bs only once, and
pretty 7 times. It is often the case that the hotspots are even more fragmented, making them even harder to interpret.