[net2-wg] 600 pps, high throughput

[Philip Levis]

On Mon, 2006-03-13 at 23:54 +0100, henri dubois-ferriere wrote:
> On 13/03/06, Philip Levis <pal at cs.stanford.edu> wrote:
> 
> > If the idle time per packet is I, then
> >
> > 616 * (I + 160/256k) = 1
> > 380 * (I + 368/256k) = 1
> >
> > so the idle time is ~1ms per packet, but it looks like it's not
> > independent of the packet size (you do have to read it out, after all).
> >
> 
> On that topic, I believe the SpiByte and SpiPacket msp430
> implementations in 2.x do slow things down a bit in comparison with
> 1.x.
> Something like 40us per byte (*), so you can quickly run up to 1ms
> transfer time with large packets. For SpiPacket, this could be
> improved by growing SPI_ATOMIC_SIZE in Msp430SpiP.nc
> 
> Henri
> 
> (*) this is from memory, not a firm number -- but i don't have an
> oscilloscope at home to verify.

As Joe's pointed out in the past, the best solution is to use DMA
(assuming you have a free DMA channel). If you have CSMA enabled, none
of this makes a big difference, though, as the CSMA latency is greater
than the loading time. I think Jonathan was planning on checking in a
DMA version at some point, based on the re-architecture of the MSP430
USART system Vlado, he, and I came up with after the TTX. He's been
doing some pretty serious performance tradeoff comparisons, so probably
has some good thoughts on the matter.

Jonathan?

Phil