The Metric of Scale: Real-World Experiences

via Infrastructural Software

Everyone involved in Grand Challenge HPC is concerned with the metric of scale. Application developers covet linear speedup as a function of processor count. System vendors are driven by scale in the development of next-generation architectures. Interconnect vendors aim to maximize bandwidth while minimizing latency irrespective of scale. Programming models, plus the compilers and tools that support them, are also designed with scale in mind. The case is no different for infrastructural software: Scale is a gating factor. Drawing on real-world experience in managing shared and distributed-memory applications, illustrations of the incremental and monumental impact of scale is demonstrated. Although these capability-oriented workloads are very effective in illustrating the challenge of scale for infrastructural software, there is another class of workloads that manifest scale quite differently. With the focus on parallel processing, application throughput exposes the metric of scale for capacity-oriented workloads. Again drawing on real-world experience, incremental and monumental ramifications of capacity-driven scale are presented. With respect to infrastructural software, scale has been a shrewd motivator in the past, and will continue to create challenges and opportunities in the future.