Measuring single cost centers

Learn
1

Is it possible profile single SCC’s ? e.g. the result of a given function, after reduction to normal form.

Let’s say I want to analyze `primal` :

let
   tup = f input1 input2
   primal = {-# SCC primal #-} fst $!! tup

Time:

  • .prof files only contain normalized times and allocations, as percentages, e.g. :
mem-bench +RTS -p -RTS

total time  =        0.01 secs   (6 ticks @ 1000 us, 1 processor)
total alloc =   9,873,104 bytes  (excludes profiling overheads)

COST CENTRE                    MODULE                             SRC                                                         %time %alloc

unsafeLinearWrite              Data.Massiv.Array.Manifest.Unboxed src/Data/Massiv/Array/Manifest/Unboxed.hs:(156,3)-(157,74)   16.7    0.0

Is it enough to compute the proportion of total alloc (i.e. scc_time = %time * total_time ) for a given SCC?

Memory:

I imagine one would extract the eventlog with one of the RTS `-l` flags, but documentation on the various eventlog processing tools is quite sparse.

Alternatively, I’ve used `weigh` in the past but it’s incompatible with multithreading and some of my use cases use that feature instead.

Thank you for all pointers!

2

I think you may be able to answer more profiling questions by loading your profile in https://www.speedscope.app

Profile using the +RTS -pj option and then load the resulting .prof file into the speedscope website.

Some more notes about profiling: GHC Profiling a Cabal Project with an Interactive Application - #2 by romes

2 Likes
3

Thank you, that works of course, but I have a really hard time parsing flame graphs. I also notice that the SCC’s I defined do not appear anywhere in the flame graph, whereas they appear in the generated JSON file.

What I would like to show is the memory and time dynamics of those specific SCC’s in a dedicated time series plot.