Not a huge amount of PySNMP speedup for the day, wound up doing quite a lot of whole-application profiling, but hours of that only represents a few runs in total (it takes a long time to load 100MB of profile data in wxProfile). Discovered that, in the demo code, almost 33% of running time is spent in PySNMP's "is a prefix" method for OIDs, which is heavily exercised by the table-retrieving code.
I'm still working on the fix for that, but while I was doing so, I wanted to generate a patch to apply to the demo server to improve performance now. Unfortunately, with SourceForge's anonymous CVS down, I can't work with PySNMP (I'm not an official PySNMP developer).
Tomorrow I'm going up to ACI to show the system to the technicians, suppose I should at least get the newest code release running tonight, even if I can't get the optimisations online.