On Fri, Aug 28, 2015 at 1:25 AM, Ben Laurie <[email protected]> wrote: > > On Thu, 20 Aug 2015 at 22:33 Trevor Perrin <[email protected]> wrote: >> >> I think to be practical here computational PIR would need to handle >> something like: >> >> 1B users (e.g. phone numbers or email addresses) in system >> 1K users = average address book size >> 1% of users join or leave system each day >> 1% of address book entries change per day >> 1B queries per day (one per user) >> 100 queries per core per second (on server) > > > The rest of your numbers I can mostly agree with, but where does this one > come from? > > Clearly, the system has to handle ~10,000 qps as a whole. Using Ian's > numbers, that would be 1.2M cores. Ouch! And actually, it would presumably > be substantially worse because your numbers are a lot higher.
Yeah, I was proposing 2^30 registered users instead of Ian's 2^24. I think PIR cost on the server scales at least linearly with # of database entries, because the server has to consider all of them per query. So I think the naive estimate would be 64x higher, so ~80M cores? > Are you saying the system has to be sized to 100 cores? If so, why? It's a crude cost estimate - the more expensive this gets, the less likely it gets deployed. But precision isn't super important here as long as PIR techniques are several orders of magnitude away. >> 1 second query response for mobile client (including communication) > > > Another number I would argue with - for a daily address book refresh, this > could take substantially longer. Even if I were wanting the lookup live, it > can take as long as it takes me to construct my message, at least. Recipient selection is often done *before* composing a message. For example: * User starts an app * User taps "Send new message" * Contact discovery happens while user waits * App displays the intersection of address book with registered users * User selects recipient(s) from intersection list * User composes message Contact discovery delay here is very noticeable. Trevor _______________________________________________ Messaging mailing list [email protected] https://moderncrypto.org/mailman/listinfo/messaging
