Isolating the GC code from the jvm to run on external accelerator

Ofir Gordon ofirg6 at gmail.com
Tue May 5 07:25:41 UTC 2020


Another question after reading the mark-compact based on the Epsilon GC
tutorial - I still don't understand how it gets from oop_iterate to
preorming the closure's do_oop_work method which actually does the marking..
I understand that if I want to implement a new GC i don't need to mind
that, but I'm trying to understand all the dependencies in the marking dlow
for the existing GCs, so I need to figure out how is the call to
do_oop_work is being made and where? what is the complete flow?

Thanks again,
Ofir

‫בתאריך יום ג׳, 5 במאי 2020 ב-9:39 מאת ‪Ofir Gordon‬‏ <‪ofirg6 at gmail.com
‬‏>:‬

> Thanks a lot!
>
> I will read the documents you attached and look for the Epsilon GC.
> Regarding my other question, is it possible to do something like this? to
> take the gc into an external program that compiles and runs independently
> from the jvm?
>
> ‫בתאריך יום ב׳, 4 במאי 2020 ב-15:48 מאת ‪Aleksey Shipilev‬‏ <‪
> shade at redhat.com‬‏>:‬
>
>> On 5/4/20 2:37 PM, Ofir Gordon wrote:
>> > If there any other advice how to use run the gc on an external hardware
>> > (than the one running the jvm) or if someone had expirience with such a
>> > task, I'll be glad to hear about it.
>>
>> From the problem description, I would suggest to look at Epsilon to
>> understand where it is hooked
>> into runtime interface in allocation paths, and then this post [1] that
>> builds the trivial
>> mark-compact over it. I think it would then be evident how marking is
>> done, and how objects are
>> iterated with shared code [2]. After that, understanding other collectors
>> would hopefully be simpler.
>>
>> --
>> Thanks,
>> -Aleksey
>>
>> [1] https://shipilev.net/jvm/diy-gc/
>> [2] https://shipilev.net/jvm/diy-gc/#_marking
>>
>>


More information about the hotspot-gc-dev mailing list