|
|
Posted by anthonyberet on 02/22/06 23:17
fred-bloggs wrote:
> anthonyberet <nospam@me.invalid> wrote in
> news:461k2jF8tjb2U1@individual.net:
>
>
>>fred-bloggs wrote:
>>
>>>anthonyberet <nospam@me.invalid> wrote in news:45v2e7F8mnkoU1
>>>@individual.net:
>>>
>>>
>>>
>>>>fred-bloggs wrote:
>>>>
>>>>
>>>>>That's incorrect. Primaries are more necessay than ever.
>>>>>
>>>>>
>>>>
>>>>Can I ask on what you base that assertion?
>>>>(not that I disagree necessarily).
>>>
>>>
>>>I used to have 4 secondaries, last time I looked it was 12.
>>>
>>
>>so... the density of your data store is improved,
>
>
> Increased load on the primary, resulting in fewer search results and less
> complete browses.
>
Increased load is not in itself a problem - only if the load exceeds
what the primary can handle. This is set in the network bandwidth in the
client settings. - excessive load will manifest itself in sluggish GUI
performance as well as a drop-off in the SPM showing.
Excessively light load is also damaging in terms of search results
returned to others.
>
>>the search horizon
>>of all peers within range of your primary has improved.
>
>
> I believe the search horizon is a *fixed* number of primaries, it has
> nothing to do with the number of secondaries.
>
I do not believe that the search horizon is a fixed number of primaries.
- Think about it, how does one primary handling a search know how many
other primaries are handling the same search? -unless the other
primaries are directly between it and the originator of the search.
Searches going off in opposite directions have no way to communicate to
each other how many primaries have been passed.
I think that the horizon is defined in 'hops' - research published
previously in this group suggests that the horizon could be as close as
three hops - it is desirable for that search to pass through primaries
with as many primary connections as possible, but just as importantly,
that those primaries have as rich libraries as possible. - This means
plenty of secondaries.
>
>>The number of
>>fake files experienced by those other peers might now be reduced...
>
>
> Only in that total number of search results is decreased.
>
not necessarily - although I need to think about this further.
- The fake files attack seems to work by a small number of peers
supplying fake lists of files to numerous primaries. The multiple
similar results returned are when a search passed through several
primaries with the fakers attached as secondaries.
One way to reduce the numbers could be to reduce the number of available
primaries, but this issue is complicated.
>
>>What's the problem?
>
>
> Decreased load balancing, fault-resilience, robustness, and scalability.
> Ideally there should be NO secondaries.
The whole idea of a supernodal network like the WPN is to improve
scalabilty, and with it those other things. If a supernodal network has
no ordinary nodes, then the supernodes are basically just ordinary nodes.
- This is the topography first used by GNUtella in the very early days
of distributed p2p but quickly superseded by networks like winmx.
It just isn't so that more primaries is a purely good thing - primaries
and secondaries are designed to work together, to group as many
potential search hits together in the path of searches without those
searches having to multiply so much that they clog the whole network up.
There must be some optimal ratio of primaries vs secondaries on average,
but for each primary this is governed by the bandwidth alloted for
network traffic. - a perfect primary would not upload or download at
all, but would devote all its bandwidth to search traffic and to
maintaining as large a pool of secondaries as possible, so as to supply
as many search results as possible.
- This is similar to the original Napster topography, where the central
indexing server did not supply files, but carried every connected client
as a secondary - at one point hitting 2 million simultaneous
'secondaries' *all* of whose files could be returned for *all* search
queries.
- I remember having to actually set a limit on results returned with
Napster, which I have never had to do with any other p2p system.
(having said that, my patience for watching lists grow and grow on my
screen is very large).
Navigation:
[Reply to this message]
|