On Tue, 4 Aug 2015 17:55:51 -0400 (EDT) "Steve Snyder" swsnyder@snydernet.net wrote:
Given a contemporary release of Tor with a contemporary version of OpenSSL, under what circumstances is the intrinsic curve25519_donna() preferred over the libsodium/NaCl crypto_scalarmult_curve25519(), or vice versa?
A quick peek at the libsodium code reveals that it uses curve25519-donna (64 bit targets), and ref10 (32 bit targets?).
Standard Tor uses curve25519-donna for both. I *think* donna's 32 bit code should out perform ref10, so libsodium is pointless.
djb's NaCl might have assembly versions that are faster, but the difference doesn't look massive.
http://bench.cr.yp.to/impl-scalarmult/curve25519.html
Does it come down to 32-bit vs. 64-bit? Or CPU instruction sets detected at build time?
Or has libsodium/NaCl been deprecated altogether and just not removed from the configure script?
I think mostly this, at this point. Deprecating it might make sense, since libsodium is basically never going to outperform donna, and it's not practical to use libsodium/NaCl for Ed25519 due to Tor specific changes that would be required.
I'm trying to determine if installation of libsodium in a Tor build environment will yield a "better" Tor binary.
If libsodium, then no. If djb's NaCl then *maybe*, but the difference in practice will probably be fairly negligible (NB: discussing 64 bit platforms. 32 bit platform performance is somewhat non-interesting to me, so I haven't measured things in depth).
I'm planning on revisiting this issue at some point, but last I looked into it, using an assembly optimized Curve25519 implementation was potentially a 7-10% gain (but is neither from libsodium nor djb NaCl).
https://trac.torproject.org/projects/tor/ticket/8897