[mnet-devel] Re: [mnet-cvs]calculate timeouts according to size of sent message

[Zooko O'Whielacronx]

> if you have another description of network behaviour, feel free to 
> fine-tune or change my patch...

Well, I really think that mu + 6*sigma might match network behavior better than 
mu*scalingfactor + sigma, for two reasons:

1.  There's that darned scaling factor.  If it's too low, you consistently incur 
unnecessary timeouts.  If it's too high, you consistently wait longer than you 
have to before launching a replacement request.  mu + 6*sigma has no magic 
constant, and it can *never* be consistently wrong in either direction over the 
long run.  (Although it can be inconsistently wrong -- too far one way half the 
time and too far the other way the other half of the time.)

2.  The 6*sigma is more accepting of high variance than the sigma.  I sort of 
think of these transaction timings as "the left half of a tight normal curve 
followed by a Poisson distribution".  I have no idea if there's a word for that 
pattern, or if mu + 6*sigma is a good predictor for it, but what the heck: 
mu + 2*sigma is a 90-something% predictor for normal distribution and 6*sigma is 
I think a 90-something% predictor for exponential distribution, which I think is 
the same as a Poisson distribution.

If I *really* wanted timeouts scaled to the size of the request message, which 
I'm not convinced that I do, I might consider actually calculating a linear 
regression (see pyutil.memutil.measure_mem_leakage() for an example of how to 
simply calculate a linear regression estimate) and add some fuzz factor.  Or 
maybe be super clever and do some kind of statistical match against a Poisson 
distribution or something.

Regards,

Zooko



-------------------------------------------------------
SF.Net is sponsored by: Speed Start Your Linux Apps Now.
Build and deploy apps & Web services for Linux with
a free DVD software kit from IBM. Click Now!
http://ads.osdn.com/?ad_id=1356&alloc_id=3438&op=click
_______________________________________________
mnet-devel mailing list
mnet-devel at lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mnet-devel