[m-dev.] enabling stseg component by default

Julien Fischer jfischer at opturion.com
Wed May 22 14:46:40 AEST 2013


On Wed, 22 May 2013, Paul Bone wrote:

> On Wed, May 22, 2013 at 02:34:22PM +1000, Julien Fischer wrote:
>>
>>
>> On Wed, 22 May 2013, Paul Bone wrote:
>>
>>> On Tue, May 21, 2013 at 05:14:22PM +1000, Julien Fischer wrote:
>>>>
>>>> Hi all,
>>>>
>>>> I propose that we enable the stseg (stack segments) grade component
>>>> by default for all low-level C grades in the post 13.05 versions
>>>> of the compiler.
>>>>
>>>> Any objections?
>>>
>>> I don't mind.  Although the memory zones code probably needs further tuning,
>>
>> In which respect?
>
> There are three things I"m aware of:
>
>    1) You requested a way to get stats about memory zone usage, I never
>       implemented this.
>    2) The caching and freeing of zones (how many do we cache, what policy
>       is used to prevent the cache from growing too large) has never been
>       benchmarked/tuned.

One thing we could do is add something to the stdlib's gc module that
clears the zone cache, or possibly just extend gc.garbage_collect/[02]
to do that.

So far as tuning the other parameters you mention goes, we probably need
to start using the stack segement grades more to do that ...

>    3) I planned, but never implemented, per-engine caching of stack zones
>       in parallel grades (in addition to a global cache).  My current
>       thinking is that an engine should hang onto a small number of zones,
>       1 or perhaps 2 so that if a process keeps crossing a zone threshold
>       it doesn't need to bother the global cache often.  The local cache's
>       data structure should also be very simple to make this operation as
>       quick as possible.

That's not really an important issue for the non-parallel grades! ;-)

What I am aiming to do here is to make it less likely that users will
bump into stack exhaustion problems due to the stacks in the non-stseg
grades having a fixed size.

Cheers,
Julien.




More information about the developers mailing list