In JDK-8314438 we have a very nice synthetic benchmark that can be used to measure speed performance of unsafe malloc.

However, we need a benchmark that measures the memory used by NMT as well, and for that we need to drop into native code on the VM level.

I propose a mechanism to record all the NMT requests to allocate the memory, which can be later used to "replay" it to measure performance that is as close to the real thing as possible - see the "NMTBenchmark design document.pages.pdf" for more details.