[GHC] #13288: Resident set size exceeds +RTS -M limit with large nurseries

[GHC]

#13288: Resident set size exceeds +RTS -M limit with large nurseries
-------------------------------------+-------------------------------------
        Reporter:  j6carey           |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:
       Component:  Runtime System    |              Version:  8.0.2
      Resolution:                    |             Keywords:
Operating System:  Linux             |         Architecture:  x86_64
 Type of failure:  Runtime           |  (amd64)
  performance bug                    |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------
Description changed by j6carey:

Old description:

> We observed high resident set size well in excess of the `+RTS -M` limit
> in a long-running, high data-volume Haskell application that Awake
> Networks is deploying on a network appliance.
>
> We think that the current `GC.c` code has two bugs that, at least in
> combination with each other, become significant when high `+RTS -N` and
> very high `+RTS -A` values are used.
>
> 1. As we approach the `-M` limit, the computation of the new size for
> generation 1 appears to be based on an incorrect figure for the total
> size of the nursery.  The `-A` value is used instead of the product of
> that value with `-N`.  This problem could lead to the total heap size
> exceeding the `-M` limit.
>
> 2. Memory allocated from the operating system is freed only if the RTS
> thinks that it would not be reallocated soon.  The estimate for what will
> be needed soon is based on fewer inputs than the actual resizing logic,
> and in particular it is not affected by `-M`.  Thus it might keep free
> mblocks in excess of the `-M` limit, based on an expected heap growth
> that would be forbidden by `-M`.
>
> We prepared to address these issues; it points out the particular lines
> of code: https://phabricator.haskell.org/D3143

New description:

 We observed high resident set size well in excess of the `+RTS -M` limit
 in a long-running, high data-volume Haskell application that Awake
 Networks is deploying on a network appliance.

 We think that the current `GC.c` code has two bugs that, at least in
 combination with each other, become significant when high `+RTS -N` and
 very high `+RTS -A` values are used.

 1. As we approach the `-M` limit, the computation of the new size for
 generation 1 appears to be based on an incorrect figure for the total size
 of the nursery.  The `-A` value is used instead of the product of that
 value with `-N`.  This problem could lead to the total heap size exceeding
 the `-M` limit.

 2. Memory allocated from the operating system is freed only if the RTS
 thinks that it would not be reallocated soon.  The estimate for what will
 be needed soon is based on fewer inputs than the actual resizing logic,
 and in particular it is not affected by `-M`.  Thus it might keep free
 mblocks in excess of the `-M` limit, based on an expected heap growth that
 would be forbidden by `-M`.

 We prepared a fix to address these issues; it points out the particular
 lines of code: https://phabricator.haskell.org/D3143

--

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13288#comment:2>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler