Evaluate mold on high-core-count server processors

Investigate the performance of the mold massively parallel ELF linker on high-core-count server processors with additional memory channels, particularly whether the additional memory bandwidth raises the random-access performance ceiling observed in the scalability experiments.

Background

The paper reports that mold’s scalability plateaus beyond approximately 32 threads because its parallel passes become increasingly constrained by metadata working-set growth, demand loads served from DRAM, and random-access memory latency. Adding threads beyond 32 increases CPU time without improving wall-clock time on the evaluated 64-core Threadripper system.

The authors identify evaluation on high-core-count server processors with more memory channels as future work. Such systems may provide greater memory bandwidth and could reveal whether mold can scale beyond the observed 32-thread plateau.