Atomic maximum/minimum

Paper P0493 for WG21

authors: Al Grant ([email protected]), Bronek Kozicki ([email protected]), Tim Northover ([email protected]) audience: SG1, LEWG

Benchmarks

• bench1 populates fixed size queue with empty data
• bench2 finds maximum value from a series generated by a PRNG

Benchmark results from bench1 have typically higher standard deviation and are more susceptible to outside noise, due to contention in the Gong algorithm in CAS loop searching for the next free entry.

Note, the default architecture for benchmarks is not Intel x86_64, as our knowledge this architecture does not implement fetch_max in hardware; instead it is ARMv8.1. The benchmarks were run on AWS Graviton2 EC2 instances (type C6g).