Status of GPU garbage collection efforts?

[Frost, Gary]

Philip,

Actually, so far we have not dived too deep into discussions centered on garbage collection. 

Myself and colleagues from AMD have been focused on HSA ( see  http://hsafoundation.com) enablement (via Graal). As you may know, HSA offers a number of compelling advantages for GPU execution. Including, fully coherent memory access from GPU and CPU side. We believe that  (for HSA compatible devices) many of the memory/allocation/GC issues traditionally associated with CPU + discrete GPU devices can be sidestepped.  

I do think it would be great to start a discussion regarding GC and allocation strategies for GPU execution. I think your previous work with Jikes will be very valuable here.

Gary


-----Original Message-----
From: sumatra-dev-bounces at openjdk.java.net [mailto:sumatra-dev-bounces at openjdk.java.net] On Behalf Of Philip Reames
Sent: Sunday, July 07, 2013 4:02 PM
To: sumatra-dev at openjdk.java.net
Subject: Status of GPU garbage collection efforts?

Good afternoon,

I am curious what the current status of the GPU garbage collection effort is.  I can't find much online or in the mailing list archive that describes your efforts.  I am curious to see what approach you've taken and what bottlenecks you've run into.  In particular, I'd be really interested in knowing what object layouts you've chosen, whether you're doing stop-the-world or concurrent GC, and how you're handling load balancing across compute units on the GPU.

I am one of the authors of a previous GPU GC implementation using the Jikes RVM.  For those curious, our implementation is described in a paper (1) that appeared at ISMM 2012 and our source code is available on github (2).

(1) http://www.philipreames.com/publications/ismm2012-gpugc.html
(2) https://github.com/preames/gpu-garbage-collection

We've been weighing the possibility of extending our previous work to a truly concurrent collector.  Before we move in that direction, I'm very much interested in hearing about your efforts.

Yours,
Philip Reames