Open zl-wang opened 4 years ago
@babsingh has been working on retuning our locks to use an MCS-based lock to address high contention scenarios. I'd like to see that work complete before we patch an additional test+branch into our locks
@babsingh can you link this to your overall MCS lock design?
Related: https://github.com/eclipse/openj9/issues/5681 (linking to the overall MCS lock design).
During previous customer PoC, it was found that there was big performance benefit to be gained tuning the spin count and getting rid of the yieldCPU() call. The workload is not that friendly on a big machine. It is also understood that getting rid of yieldCPU() may hurt other workloads. Such that, this asks for an option to meet both worlds on the best possible terms.