Okay, first things first. I have put a change in which I believe will
significantly increase insert performance. It goes as follows:
- Inserts visit many more nodes than requests.
- Therefore they take longer - the round-trip-time part of the throttle
  is higher.
- They also have a much higher chance of running into an overloaded node
  at some point during the request, so the window will also be much
  greater.

IMHO they are being excessively and unduly penalized, so I have changed
the load limiting system to have a single window tracker for all types of
requests, and per-type round-trip-time's. Is this reasonable?

If you accept the above, then we have another problem: If the user
starts the node and does an insert, the above change is subverted; there
are inserts happening but no requests, so even if we remember what our
window size was before, it will rapidly get smaller due to only doing
inserts. The problem is lack of information. If we allow all requests to
contribute to the window size, then we have more information, more
accurate limiting, and faster inserts.

BTW congrats on getting PGP working.

On Tue, Apr 11, 2006 at 11:42:16AM -0700, Ian Clarke wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> 
> On 11 Apr 2006, at 09:15, Matthew Toseland wrote:
> 
> >Which requests should count for load limiting?
> >
> >Load limiting is the process whereby if we get a RejectedOverload or a
> >timeout we reduce the rate at which we send (locally originated)
> >requests, and if we don't, we increase it.
> >
> >The original intention I think, based loosely on the TCP-over-Ethernet
> >metaphor, was to only count locally originated requests. So if we  
> >get a
> >timeout on any other request, this doesn't affect the rate at which we
> >send requests.
> >
> >Is this the best option? It is perhaps closest to "propagate the load
> >back to the originator"? But maybe more information - counting other
> >requests - would be better?
> 
> We should not change this, we need to keep it simple - only  
> introducing further complexity if there is a *clear* justification.   
> I don't see a clear justification here.
> 
> Ian.
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.2.2 (Darwin)
> 
> iD8DBQFEO/iLQtgxRWSmsqwRApq3AJ0brx1vu7aK1WtZhgyxl95wmxhl2wCeMaVm
> i9RW9QqbzcE2vpalivRTryY=
> =SM3m
> -----END PGP SIGNATURE-----
> 

-- 
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/devl/attachments/20060411/0cc80259/attachment.pgp>

Reply via email to