RandomIO.implNextBytes() in NativePRNG class can become a bottleneck at high contention. This shows up as the hottest lock in SPECjvm crypto.rsa benchmark.

We think we avoid the lock by bypassing the global buffer (at least for "large enough data"). Or perhaps even using user-level library and bypassing /dev/random in some environments.

This calls for some research on our part and consulting before putting together patch.