Lenz Weber writes:
Hello, I'm currently investigating some serious performance problems when using courier-0.66.1ubuntu4 in combination with gnutls-3.0.11+really2.12.14-5ubuntu3.5, which would be the usual combination if one were to compile courier using the --with-gnutls config flag in the lastest Ubuntu LTS version 12.04.
The key part in the strace, here, http://paste.ubuntu.com/6724497/ starting around line 500. What should be happening there is nothing more than a read()ing a nicely-sized chunk from a pipe, file descriptor 6, and then tossing the whole thing to gnutls_record_send() which writes it to a socket on file descriptor 0.
Between each call to read(), and a subsequent writev() of the encrypted chunk, there are two sets of calls gettimeofday(), getrusage(), and times(). I can't see why that alone would be enough to grind things to a halt; however it's hard to see the actual timings here.
However, I happened to have sources of 3.2.7 handy, and from poking around, those syscalls appear to be coming out of wrap_nettle_rnd_refresh() in lib/nettle/rnd.c, which appears to have something to do with reseeding an internal random number pool. It looks like there's a lot of work happening here; and I don't quite follow why that needs to be done for each call to gnutls_record_send(); but that's as far as I could go, with what I know.
pgpubYCd6F3im.pgp
Description: PGP signature
_______________________________________________ Gnutls-help mailing list [email protected] http://lists.gnupg.org/mailman/listinfo/gnutls-help
