RFR: 8317453: NMT: Performance benchmarks are needed to measure speed and memory [v4]

[Thomas Stuefe]

On Mon, 20 Jan 2025 11:45:14 GMT, Johan Sjölen <jsjolen at openjdk.org> wrote:

> I haven't read the code yet, but this is something that we have used internally for testing different optimizations in NMT.

I had a quick read through the code to get the gist of it. Some sort of design document would be rather helpful here.

I am not really sure what that benchmark does. We can "replay" operations, which for malloc means we re-execute the mallocs, for maps it means we just execute the NMT bookkeeping operation without the associated sys calls and without the locking needed to bracket the latter with the former. What do we measure here?

I have the feeling the purpose is to somehow extract the time cost for NMT bookkeeping? To separate it from the cost associated with synchronization and associated with OS-side allocation calls (but then, why do we execute mallocs on replay?) ?

-------------

PR Comment: https://git.openjdk.org/jdk/pull/23115#issuecomment-2602450326