System time is increasing on enterprise workloads as multi-core and non-uniform memory architecture (NUMA) systems become mainstream. A defining characteristic of the increase in system time is an increase in reference costs due to contention of shared resources - instances of poor memory locality also play a role. An exploration of these issues reveals further opportunities to improve kernel scalability.
This paper examines kernel lock and scalability issues encountered on enterprise workloads. These issues are examined at a software level through structure definition and organization, and at a hardware level through cache line contention characteristics and system performance metrics. Issues and opportunities are illustrated in the process scheduler, I/O paths, timers, slab, and IPC.