You are right

As of today as SGen works is like this:
Gen0 (or nursery) and Old Generation split in two: the free list for large objects and the old generation.

Your task will consist in following: create Gen1 and you will have the following design:
Gen0: short lived objects
Gen1: medium lived objects (using bump-pointer)
Gen2: long lived objects (using bump-pointer)
GenFreeList: large objects (regardless of generation)

I really hope that your project will evolve nicely and I would love to see this addition to Mono runtime

On Sat, Nov 6, 2010 at 1:56 AM, Jérôme De Cuyper <jerome.decuyper <at> gmail.com> wrote:

This is my first message on the mono-gc mailing list and I hope I will not break any implicit rule it may have. Otherwise please let me know!

I decided to use my CS thesis as an opportunity to contribute some code to the mono project. Lupus kindly offered the following task:

"we currently have 2 generations: introduce a third one that sits between the nursery and the old generation. It should be a bump-pointer

style generation for fast gen0 collections."

Before starting to dig into the code, I need to justify the change that is proposed. I suppose the purpose of a new generation is to create a  group of objects from

a different age in order to recollect as much of them as possible before sending them to the two oldest generations. Is that correct?

Please correct me but I'm guessing the new flow of collection would be as follow: 

gen0 - newest objects - bump-pointer based

gen1 - middle aged objects  - bump-pointer based (the new generation)

gen2 - oldest generation - size-segregated freelists

gen3 - oldest generation - bump-pointer based

Thank you!

Jérôme De Cuyper
http://www.jdecuyper.com

_______________________________________________
Mono-gc-list maillist  -  Mono-gc-list <at> lists.ximian.com
http://lists.ximian.com/mailman/listinfo/mono-gc-list

_______________________________________________
Mono-gc-list maillist  -  Mono-gc-list <at> lists.ximian.com
http://lists.ximian.com/mailman/listinfo/mono-gc-list

Hi Jérôme,

First, a minor correction on the current setup of sgen. Right now mono use a 2 generations / 4 spaces setup.

There are 2 generations: young and old.

There are 4 spaces: nursery, major, los and los-huge.

The distinction before generation and spaces is important when discussing things like write barriers and collections.

We have los and los-huge to better handle allocation of large objects[1] against huge objects.

Said that, adding the extra generation will certainly help making gen0 collections faster. Some aspects of it

are kind of open questions that will require experimentation as their answer are not always obvious.

Things like how handle write barrier efficiently and figuring out a good size relation between gen0 and gen1.

I would love to see someone interested in cracking this issue. 

Cheers,

Rodrigo

[1]Right now between 8k and 1020k

On Fri, Nov 5, 2010 at 10:56 PM, Jérôme De Cuyper <jerome.decuyper <at> gmail.com> wrote:

This is my first message on the mono-gc mailing list and I hope I will not break any implicit rule it may have. Otherwise please let me know!

I decided to use my CS thesis as an opportunity to contribute some code to the mono project. Lupus kindly offered the following task:

"we currently have 2 generations: introduce a third one that sits between the nursery and the old generation. It should be a bump-pointer

style generation for fast gen0 collections."

Before starting to dig into the code, I need to justify the change that is proposed. I suppose the purpose of a new generation is to create a  group of objects from

a different age in order to recollect as much of them as possible before sending them to the two oldest generations. Is that correct?

Please correct me but I'm guessing the new flow of collection would be as follow: 

gen0 - newest objects - bump-pointer based

gen1 - middle aged objects  - bump-pointer based (the new generation)

gen2 - oldest generation - size-segregated freelists

gen3 - oldest generation - bump-pointer based

Thank you!

Jérôme De Cuyper
http://www.jdecuyper.com

_______________________________________________
Mono-gc-list maillist  -  Mono-gc-list <at> lists.ximian.com
http://lists.ximian.com/mailman/listinfo/mono-gc-list