Rerunning e.g. octane.richards, the same number of runs on the same machine with the same binary gives me everything between 53000 ops/s and 123000 ops/s on my machine at steady state. No more code is generated at this steady state, yet we have a huge diversion between peak scores between runs.

I understand Roland already has a partial fix for this. Filing a bug to get it tracked.

Roland- can you elaborate a little bit what the reason for this variation is in this CR, please?

We'd be willing to take your partial fix that still has some sway left into JDK9 and file another CR for any additional work that might come later.