gnutls_rnd() is more heavy-weight than I originally anticipated,
and in simple profiling it would take roughly 1% of CPU.
With this simple buffering the usage dropped well under 0.1%.