On Sat, May 26, 2001 at 10:49:19AM -0700, Ian Clarke wrote:
> On Sat, May 26, 2001 at 12:27:55PM -0500, Brandon wrote:
> >
> > I've been thinking on the caching problem recently and it seems to me that
> > it is is indeed disadvantageous to cache aggresively because files cached
> > on the edge of the network have a decreased probability of being found by
> > an arbitrary node. The closer a file is to the epicenter, the greater its
> > global probability of discovery.
> >
> > So rather than caching all the way back the search path, it would be most
> > advantagous to cache in an expanding ring around the epicenter. This can
> > be done by having a rapidly decaying probability of caching.
>
> I agree entirely, in fact here was part of my reply to that guy's post:
>
> ----
> Many thanks for your insights, your document is rather interesting. I
> tend to agree with your assessment, I think that over-zealous caching is
> having a negative effect. One experiment I hope to try is caching
> probabilistically (as you suggest), but with a decreasing probability
> the further the DataReply gets from the origin of the data. This means
> that the node where the request originated, which is extremely unlikely
> to be a "specialist" in the data being requested, should therefore be
> unlikely to cache the data.
> ---
As has been discussed previously, this is one of the things I want to look
at testing as soon as I start simulating.
I'm thinking something along the line of
P[cache] = k / (k + htl^2)
with k = (average HTL of messages that reach node)^2 .
Of course, this is not the only model. Another model might be to cache
everything, but to place it later in the list when the HTL is greater.
> Of course, we will need to simulate this behavior before implementing
> it, time to dust off Serapis...
Book three: Return of the Coder.
--
'DeCSS would be fine. Where is it?'
'Here,' Montag touched his head.
'Ah,' Granger smiled and nodded.
Oskar Sandberg
[EMAIL PROTECTED]
_______________________________________________
Devl mailing list
[EMAIL PROTECTED]
http://lists.freenetproject.org/mailman/listinfo/devl
