[Tinyos Core WG] Meeting: 7/11/07

[John Regehr]

> PL: Large atomic sections perhaps?

I wrote a bit of code for Avrora (UCLA simulator) that makes a histogram of 
cycles spent in each atomic section.  Anyone wants this, let me know.  Scanning 
these results may be a reasonable part of regression testing.  For example this 
is from RadioCountToLeds from current CVS with the histogram aggregated over all 
atomics over a 30 seconds run:

   bucket 00 (0000-0099) : 6132
   bucket 01 (0100-0199) : 135363
   bucket 02 (0200-0299) : 1177
   bucket 03 (0300-0399) : 2359
   bucket 04 (0400-0499) : 158
   bucket 05 (0500-0599) : 4
   bucket 06 (0600-0699) : 37
   bucket 07 (0700-0799) : 0
   bucket 08 (0800-0899) : 0
   bucket 09 (0900-0999) : 0
   bucket 10 (1000-1099) : 0
   bucket 11 (1100-1199) : 0
   bucket 12 (1200-1299) : 0
   bucket 13 (1300-1399) : 118
   bucket 14 (1400-1499) : 0
   bucket 15 (1500-1599) : 0
   bucket 16 (1600-1699) : 0
   bucket 17 (1700-1799) : 1
   bucket 18 (1800-1899) : 118
   bucket 19 (1900-1999) : 0
   bucket 20 (2000-2099) : 0
   bucket 21 (2100-2199) : 0
   bucket 22 (2200-2299) : 0
   bucket 23 (2300-2399) : 0
   bucket 24 (2400-2499) : 0
   bucket 25 (2500-2599) : 0
   bucket 26 (2600-2699) : 0
   bucket 27 (2700-2799) : 0
   bucket 28 (2800-2899) : 0
   bucket 29 (2900-2999) : 118

(Nothing in any larger bucket.)

A problem is that Avrora only gives results for Mica2.  I will just repeat my 
plea for a cycle accurate simulator for all major platforms.

The question "How long is too long for an atomic?" is harder in general, though 
easy for the case of a serial port that interrupts on every byte.  For example 
at 115 kbps we might expect problems if there are too many atomics longer than 
~500 cycles.

> DG: Tasks assume they are never interrupted -- they don't protect busy
> flags with
> atomic for example.

Conservative detection of variables shared across tasks is something we can do 
very easily.  This would seem to be useful if anyone wants to threadify.

John Regehr