On Tue, Aug 22, 2006 at 11:21:27AM -0500, David Sowder (Zothar) wrote:
> Matthew Toseland wrote:
> >
> >Sure, but we're determining whether to pre-emptively reject requests
> >based on our estimate of load. Two of the factors are ping time
> >(700-1500ms we reject some, more than that we reject all) and bwlimit
> >delay time (1000-2000ms, but I think there is a problem with the
> >limiter). The objective here is to prevent timeouts; timeouts are *bad*.
> >  
> I would agree that there is a problem with the limiter.  Perhaps we 
> could isolate that code and do some testing on it and debug until a 
> flood of test packets shows up as limited on an MRTG (i.e. the graph 
> never goes above the limit (and not by much if it does, later 
> compensating), but roughly flatlines at the limit).

Well, here's the thing: We limit the number of requests we accept using
a token bucket (80% of the output rate, and a fairly generous, perhaps
too generous, allowance for spikes [1]) and the average number of bytes
used in- and out- per request. Therefore, we should not accept more
requests than we can handle, and the bwlimitDelayTime should never be
very high. Yet it does go very high from time to time.

[1] Tokens are added at 80% of the bwlimit, the bucket size is 60
seconds' traffic or 64kB whichever is greater.
-- 
Matthew J Toseland - toad at amphibian.dyndns.org
Freenet Project Official Codemonkey - http://freenetproject.org/
ICTHUS - Nothing is impossible. Our Boss says so.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: 
<https://emu.freenetproject.org/pipermail/tech/attachments/20060822/24926da5/attachment.pgp>

Reply via email to