This is an excellent writeup, but I'm curious about how dependent the microbenchmark is on the scripting engine doing no real work. My off-the-cuff guess is that the unshared version might still perform better with other workloads (it would have higher throughput and avoid variance caused by fluctuating degrees of cache contention) but it might be that the observable overhead from hammering a single cache line would disappear if each thread had more varied work to do. And of course if the scripting engine benefits substantially from caching internal state across invocations (things like incrementally optimized JIT code) then having 24 instances can incur substantial extra work.