I suggest you try this benchmark by at least mimicking a real application, by e.g. adding a use Moose to your program. The problem with Perl 5 ithreads is not in tests like this. The problem is in real applications that set up a lot of data structures that need to be copied to all 3000 threads in your example. And whose memory will not be shared. I bet it will run a tad over 2 GB by just adding use Moose.

My suggestion was to add a use Moose (and nothing else) to your benchmark program, so that it would be closer to a real world benchmark.

That's like responding to being challenged to a race by a guy in a Ferrari, by accepting on the basis that he tow a U-Haul trailer full of kitchen sinks.

I'm glad for you that you are part of the jet-set that only needs to deal with Ferraris. I think most Perl 5 programmers have to deal with U-Haul trailers full of kitchen sinks every single day of their working life. And thus I felt that adding a use Moose to your benchmark, would add a touch of reality for them.

By the way, the use of that Ferrari is officially discouraged as it may be taken off the market:

From time to time, we may mark language constructs and features which we consider to have been mistakes as discouraged. Discouraged features aren't currently candidates for removal, but we may later deprecate them if they're found to stand in the way of a significant improvement to the Perl core.

Looks to me you didn't examine all that I said. Or simple preferred not to react to this point. FWIW, telling people to use features of Perl 5 that you know may be removed in the future, does not feel right to me. Regardless of whether they're shiny red or not.