I can't show a simple script to defend my earlier claim, since the times were taken using Time::HiRes qw(gettimeofday).

The sub-system used is 20,000+ lines of code running persistently with pre-forked servers that dynamically expand/contract on activity. The servers respond to tcp calls from a web server with a very small mod-perl cgi script preloaded.

The design goal was to support 3 transactions per second per core with enough memory to prevent paging.

One bottleneck was to reload all hashes required for each transaction. Testing found that pre-allocting the hash key number had a significant improvement in production.

The last stress test, we exceeded 3.52 tranactions per second per core. However, the difference between hardware, operating systems, web-servers, mod-perl, and versions of perl is significant, so your mileage may vary!

Thank you