Interestingly it turns out that using the native base64 routines does
not improve performance. Likely because the actual time is in
marshalling/unmarshalling between strings and arrays (and associated
garbage collection overhead) which has to be done either way.