Hello,

this is not strictly related to D, but as D's runtime have a GC, I thought it can be useful to show an interesting research on GCs.

Traditional GCs perform poorly when memory gets tight: some of the page managed by the GC gets evicted to the disk by the Virtual memory Manager (VM) of the OS, but when the GC does a garbage collection run, it will swap in the evicted page which will cause the VM to evict another page, etc.

The Bookmarking Collector (BC) is a GC which communicates with the VM to select the page to evict to avoid this memory trashing, it necessitates a (small) patch to the OS so that the VM and the GC can communicate, but the results are impressive compared to traditional GCs when there is memory pressure: improved runtime and reduced pause (even if BC is not a realtime GC).

Here's the link to the video presentation (too bad, the questions at the end are difficult to hear):
http://content.digitalwell.washington.edu/msr/external_release_talks_12_05_2005/12631/lecture.htm

And the paper:
http://www.cs.umass.edu/~emery/pubs/f034-hertz.pdf


BC is a moving memory GC implemented in a JVM, so the results are not directly applicable to D, there's a non-moving variant of BC studied in the paper it gets nearly as good runtime as BC, but the pause are not as good.

I hope some of you will find this interesting.

Regards,
renoX

Yup, it's a good technique.  About the only drawback is that it requires VMM integration to work, so it isn't terribly portable.  However, such a GC could fall back on normal mark/sweep if run on a platform it doesn't support.


Sean

Sean Kelly wrote:

> Yup, it's a good technique.  About the only drawback is that it requires VMM integration to work, so it isn't terribly portable.  However, such a GC could fall back on normal mark/sweep if run on a platform it doesn't support.

Even better: D defines that a GC must EXIST, not what kind of GC you have to implement and how it must work :-)

And your nice Tango library makes it even run time switchable AFAICS.

So, people can really play here and you guys did a damn good design here!


Best Regards

Ingo Oeser

Ingo Oeser wrote:
> Sean Kelly wrote:
> 
>> Yup, it's a good technique.  About the only drawback is that it requires
>> VMM integration to work, so it isn't terribly portable.  However, such a
>> GC could fall back on normal mark/sweep if run on a platform it doesn't
>> support.
> 
> Even better: D defines that a GC must EXIST, not what kind of GC you have to implement and how it must work :-)
> 
> And your nice Tango library makes it even run time switchable AFAICS.

Link-time switchable, actually.  I'd originally considered run-time switchable, but it introduced too many weird problems.



Sean

Sean Kelly wrote:

> Ingo Oeser wrote:
[GC]
>> And your nice Tango library makes it even run time switchable AFAICS.
> 
> Link-time switchable, actually.  I'd originally considered run-time switchable, but it introduced too many weird problems.

That is amazing enough for me. So better keep that stuff clean and fast.


Best Regards

Ingo Oeser