Indeed. I filed a bug earlier today to make use of this. On Thu, Jul 31, 2014 at 6:15 PM, Edward Kmett wrote:

Interesting.

I suppose ekg could also (ab)use this.

Johan?

-Edward

On Thu, Jul 31, 2014 at 5:51 AM, Simon Marlow wrote:

...

Hey Bryan,

Sorry for the delay.

On 15/07/14 01:57, Bryan O'Sullivan wrote:

...

I spent a bit of time over the weekend trying to figure out how to force the RTS to collect GC statistics, but was unable to do so.

I'm currently working on enriching criterion's ability to gather data, among which I'd like to see GC statistics. If I try to obtain GC stats using criterion when I'm not running the benchmark app with +RTS -T, I get an exception.

Is there a way to allow criterion to forcibly enable stats collection? My efforts to do so have gotten me nowhere. It would be unfortunate if I had to tell users of criterion that they should always run with +RTS -T or add a -rtsopts clause, as they'll simply forget.

And while I'm asking, why does GHC not simply collect GC stats by default? Collecting them seems to have zero cost, from what I can see?

So you can do this in the same way as GHC. See

https://phabricator.haskell.org/diffusion/GHC/browse/master/ghc/hschooks.c;6...

Which is imported into Haskell like this:

https://phabricator.haskell.org/diffusion/GHC/browse/master/ghc/Main.hs;6fa6...

I'm not sure why it's marked "safe", but it doesn't hurt.

This API is kind-of public, in the sense that we deliberately expose it via the Rts.h header, and I'll try not to break it gratuitously.

Cheers, Simon

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

We already have GHC.Stats. It would be nice if it was put there. On Thu, Jul 31, 2014 at 6:55 PM, Austin Seipp wrote:

The 'safe' mark was just an oversight from me - I don't really think it matters, considering it's only called once at startup anyway. Similarly I imagine the oversight is negligible in the Criterion or ekg case.

And also, take note the conditional is very specific; you want COLLECT_GC_STATS only when NO_GC_STATS is the current setting - if you unconditionally force COLLECT_GC_STATS, then things like `+RTS -sstderr -RTS` will no longer work. I mention this because I messed it up originally. :)

Given we already use this in GHC, and other users want it, it's probably reasonable to put such an interface in `base` somewhere, e.g.

data GCStatisticsOption = NoStatistics | CollectStatistics | VerboseStatistics ...

getGcStatistics :: IO GcStatisticsOption setGcStatistics :: GcStatisticsOption -> IO () modifyGcStatistics :: (GcStatisticsOption -> GcStatisticsOption) -> IO ()

or something like that.

The special case inside GHC could then be handled with such an API I suppose.

On Thu, Jul 31, 2014 at 11:41 AM, Johan Tibell wrote:

...

Indeed. I filed a bug earlier today to make use of this.

On Thu, Jul 31, 2014 at 6:15 PM, Edward Kmett wrote:

...

Interesting.

I suppose ekg could also (ab)use this.

Johan?

-Edward

On Thu, Jul 31, 2014 at 5:51 AM, Simon Marlow wrote:

...

Hey Bryan,

Sorry for the delay.

On 15/07/14 01:57, Bryan O'Sullivan wrote:

...

I spent a bit of time over the weekend trying to figure out how to force the RTS to collect GC statistics, but was unable to do so.

I'm currently working on enriching criterion's ability to gather data, among which I'd like to see GC statistics. If I try to obtain GC stats using criterion when I'm not running the benchmark app with +RTS -T, I get an exception.

Is there a way to allow criterion to forcibly enable stats collection? My efforts to do so have gotten me nowhere. It would be unfortunate if I had to tell users of criterion that they should always run with +RTS -T or add a -rtsopts clause, as they'll simply forget.

And while I'm asking, why does GHC not simply collect GC stats by default? Collecting them seems to have zero cost, from what I can see?

So you can do this in the same way as GHC. See

https://phabricator.haskell.org/diffusion/GHC/browse/master/ghc/hschooks.c;6...

Which is imported into Haskell like this:

https://phabricator.haskell.org/diffusion/GHC/browse/master/ghc/Main.hs;6fa6...

I'm not sure why it's marked "safe", but it doesn't hurt.

This API is kind-of public, in the sense that we deliberately expose it via the Rts.h header, and I'll try not to break it gratuitously.

Cheers, Simon

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

-- Regards,

Austin Seipp, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/

Right, so if you look in ./includes/rts/Flags.h, you'll see a set of different options for the GC stats: #define NO_GC_STATS 0 #define COLLECT_GC_STATS 1 #define ONELINE_GC_STATS 2 #define SUMMARY_GC_STATS 3 #define VERBOSE_GC_STATS 4 By default the RTS sets the stats flag to NO_GC_STATS, which does nothing. Note that *every* flag here (except NO_GC_STATS) implies the collection of GC statistics; it's just a matter of what's displayed to the user (if anything). There are several RTS options, each of which correspond to one of the above (see RtsFlags.c): `+RTS -T` corresponds to COLLECT_GC_STATS `+RTS -S` corresponds to VERBOSE_GC_STATS `+RTS -s` corresponds to SUMMARY_GC_STATS `+RTS -t` corresponds to ONELINE_GC_STATS These flags are parsed at startup before any Haskell code is executed obviously. So here's the problem: let's say you unconditionally force the statistics flag to be COLLECT_GC_STATS from Haskell-land - if I started my application with `+RTS -s`, then you've changed the statistics reporting from SUMMARY_GC_STATS, which means that when the program terminates, no statistics about the program will be displayed! Which will break end-user expectations. So the rule of thumb is: all of the above flags enforce the collection of statistics, and you *only* want to set the flag to COLLECT_GC_STATS *iff* it is currently set to NO_GC_STATS. In all other cases, GC statistics will already be enabled for the runtime, so you don't need to do anything. Hope that explains it well enough. On Thu, Jul 31, 2014 at 12:17 PM, Johan Tibell wrote:

On Thu, Jul 31, 2014 at 6:55 PM, Austin Seipp wrote:

...

And also, take note the conditional is very specific; you want COLLECT_GC_STATS only when NO_GC_STATS is the current setting - if you unconditionally force COLLECT_GC_STATS, then things like `+RTS -sstderr -RTS` will no longer work. I mention this because I messed it up originally. :)

Could you expand a bit on this. I want to programtically set whatever flags/settings would be set if the user passed +RTS -T on the command line.

-- Regards, Austin Seipp, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/