[Cbe-oss-dev] Costs of the SPE Context Switch

[Jeremy Kerr]

Hi Glenn,

> I have recently been familiarizing myself with the Cell architecture
> and I am interested the costs of SPE context switches on Linux.  The
> material I have found states that the cost of the context switch
> operation is relatively high (compared to a standard CPU) and is on
> the order of 20 microseconds.

I did some measurements last year and found that it was in the order of 
200us. However, that was a context switch for a basically idle SPE, so 
we didn't have to do some of the optional operations, like purge and 
save the DMA queue.

> The Cell BE Programming Handbook has a chapter (12) on the steps
> needed to carry out an SPE context switch.  I'd like to find
> benchmarks as fine-grained as possible (perhaps paralleling the
> Handbook's steps), but I would be happy with numbers characterizing
> the average costs of managing the SPE Local Store, Registers, and DMA
> queue.
>
> Does anyone know if such information is available?

I haven't heard of any numbers reported so far.

I wouldn't think this is too difficult; I'd suggest either:

 * running a context-switch-only workload and oprofiling the results; or
 * adding some trace points to switch.c and recording the hits

The best approach probably depends on what you're looking to do with the 
timing data.

Cheers,


Jeremy