[geeks] Game GPU clusters for supercomputering

[Joshua Boyd]

On May 24, 2008, at 8:52 PM, Sridhar Ayengar wrote:

> Joshua Boyd wrote:
>> As you can imagine, between the 128x128 register file, and the  
>> SRAM, a
>> context switch on a SPE is horrendous.  This means that in a
>> multitasking environment SPEs can't really be shared between
>> applications, so if you have a video player that uses 2 SPEs, it  
>> grabs
>> those SPEs and doesn't release them until you stop it.
>
> Is there anything stopping one from just requesting SPEs and  
> releasing them as if you were starting up and shutting down on a  
> process level? As in, is it possible to dynamically allocate SPEs  
> with really, really coarse granularity?
>
> Say, if said video app is playing, happily running along on two  
> SPEs, but then you hit stop.  Could it let go of those two SPEs  
> until you hit play again?  I understand that there would be a big  
> pile of overhead involved, but would the cost of the overhead  
> outweigh the benefit from relinquishing the SPEs for a few seconds?

Technically, I think you could make what you suggest work.  According  
to the IBM articles on Cell Linux, each SPE is accessed via it's own / 
dev entry, so I would assume that would mean that different  
applications could grab SPEs and release them as they pleased.