On Sun, 24 Apr 2016 22:02:23 -0700 David Fifield <david@bamsoftware.com> wrote:
I saw you say on IRC that you had an idea for improving the efficiency of meek-server. What's your idea? The server hosting meek-azure is passing 90% CPU at times.
Hmm, you're building meek-server with Go 1.6.x right? CloudFlare of all companies contributed some of the optimizations I wanted to do courtesy of V. Krasnov, formerly of Intel...[0] Alternatively using nginx or apache, since meek-server implemented as a module would be fairly trivial, and likely would perform better. [snip]
I've also done some preliminary work on allowing the server to send data in larger chunks, which should reduce polling: https://trac.torproject.org/projects/tor/ticket/12857#comment:4 All that's really needed here is a decision on how to let the client indicate that it supports such a server, as the change is backward incompatible.
Neat. I think the header value sounds reasonable (let me know when this is finalized so I can update `meek_lite`). Regards, -- Yawning Angel [0]: Specifically a non-terrible GCM-AES. The old code didn't use PCLMULQDQ when calculating GHASH and wasn't stitched so performance was total garbage even with AES-NI (NB: CBC and CTR mode are currently still unoptimized, but that's orthogonal to this).