Le 12-mars-08 à 04:36, Brian E Carpenter a écrit :
On 2008-03-07 23:32, Luigi Iannone wrote:
...
Finally, I cannot find why EID-to-RLOC cache is significantly smaller
(preferably, in terms of big-O) than the EID-to-RLOC database. Is
there any study of that issue?
You can get a look to:
There you can find the size of a cache.
That's a nice paper, but of course the size of cache is arbitrary - you
chose to limit it by setting the timeout, but if the cache is limited
by memory size, it could be much smaller (or much bigger). As with
any cache, the question is what is the working set needed to meet
a given performance goal.
You are right. And actually our measurements hold in the context of a stub network having more or less the same size and traffic pattern as ours.
You cache the size of the nework and its traffic pattern and you can have different results.
The size of the database, assuming granularity of prefixes annonced by
BGP, is just the size of a BGP table.
Not if you want map granularity to be arbitarily fine. We've talked
about map sizes of at least 10^8.
As i said, we assume a granularity equivalent to BGP prefixes.
Your values for the working set
for a given site may be valid, however - that probably needs data
collection from a large sample of sites.
See my first point.
Luigi
Brian