The performance of gnutls is dramatically better. For a small file
(5k) and DHE-RSA ciphersuites the performance is equivalent. For the
plain RSA ciphersuite the performance is still low (about 40% of the
openssl performance).
For a larger (300k) file the performance for both ciphersuites is
exactly the same.
So it seems libgcrypt is quite optimized in amd64... However there
seems to be some overhead in the plain RSA ciphersuites that affects
performance when the number of transactions increases (the first case
with the small file). Possibly the RSA blinding...
Yeah, or the TCP stack becomes the bottleneck since gnutls sends more
packets than mod_ssl. Although this needs more investigation, my guess
is that the TCP overhead for another packet is pretty small. Especially
when run on localhost.