Developing software that scales on multicore processors is an inexact science dominated by guesswork, measurement, and expensive cycles of redesign and reimplementation. Current approaches are workload-driven and, hence, can reveal scalability bottlenecks only for known workloads and available software and hardware. This paper introduces an interface-driven approach to building scalable software. This approach is based on the scalable commutativity rule, which, informally stated, says that whenever interface operations commute, they can be implemented in a way that scales. We formalize this rule and prove it correct for any machine on which conflict-free operations scale, such as current cache-coherent multicore machines. The rule also enables a better design process for scalable software: programmers can now reason about scalability from the earliest stages of interface definition through software design, implementation, and evaluation.
Until the mid-2000s, continuously rising CPU clock speeds made sequential software perform faster with each new hardware generation. But higher clock speeds require more power and generate more heat, and around 2005 clock speeds reached the thermal dissipation limits of a few square centimeters of silicon. CPU architects have not significantly increased clock speeds since, but the number of transistors that can be placed on a chip has continued to rise. Architects now increase parallelism by putting more CPU cores on each chip. Total cycles per second continues to grow exponentially, but software must scale—must take advantage of parallel CPU resources—to benefit from this growth.
Unfortunately, scaling is still an untamed problem. Even with careful engineering, software rarely achieves the holy grail of linear scalability, where doubling hardware parallelism doubles software performance.
No entries found