Geoff, I thought the reason for using (and mandating) rsync was precisely
to avoid re-load of the whole data space on each synchronization.
If so, what estimates would you use of how much of the space would be new
(what Matt calls "fresh") data at each synchronization time point?
--Sandy
(This is a clarification query, which might be viewed as co-chair hat on
or member hat on, or both, your choice.)
On Tue, 27 Oct 2009, Geoff Huston wrote:
WG Co-Chair Hat OFF
Hi Matt,
entities who are actually using RPKI data for routing SHOULD be fetching
fresh data from the repositories at least once every three hours.
3 hours?
At a first pass that seems very frequent.
From a server's perspective if there are 30,000 AS's out there and each is
running a local cache and each is a distinct relying party of the RPKI
system, then the local hit rate at the server would be 3 per second, assuming
that all the relying parties evenly spread their load (which is a pretty wild
assumption - the worst case is that all 30,000 attempt to resync at the 3
hour clock chime point) Assuming that a repository sweep with no updates
takes 30 seconds to complete then the server would have an average load of
some 90 concurrent sync sessions. If there is a local rekey then the refresh
would also imply a reload of all the signed products at this repository
publication point. Assuming that this would then take 3 minutes to download,
then the rekey load per server would be of the order of 540 concurrent rsync
sessions as an average load. These load numbers appear to me to be somewhat
large.
From the relying party's perspective if there are 30,000 distinct RPKI
repository publication points, and a serial form of local synchronisation
using a top-down tree walk then the same set of assumptions imply that the
relying party's perspective then it needs to process the synchronisation with
the remote cache (including minimally the manifest crypto calculation at a
rate of 3 per second. Assuming that there are 200,000 distinct ROAs out there
that are re-validated at each fetch then once more the numbers imply that a 3
hour refresh would infer that the relying party would need to validate
200,000 ROAS in 10,800 seconds. That probably needs some pretty quick
hardware.
These numbers are pretty much a toss at a dart board, and the draft's
authors' may well be using a different scale model to justify this
recommended time cycle. What numbers did you have in mind Matt that would
make this "SHOULD" 3 hour refresh cycle feasible in a big-I Internet scenario
of universal use?
Geoff
WG Co-Chair hat off
_______________________________________________
sidr mailing list
sidr@ietf.org
https://www.ietf.org/mailman/listinfo/sidr
_______________________________________________
sidr mailing list
sidr@ietf.org
https://www.ietf.org/mailman/listinfo/sidr