Measuring, Modeling, and Optimizing Counterintuitive Performance Phenomena in Power-Scalable, Parallel Systems.

摘要

The demands of exascale computing systems and applications have pushed for a rapid, continual design paradigm coupled with increasing design complexities from the interaction between the application, the middleware, and the underlying system hardware, which forms a breeding ground for inefficiency. This work seeks to improve system efficiency by exposing the root causes of unexpected performance slowdowns (e.g., lower performance at higher processor speeds) that occur more frequently in power-scalable systems where raw processor speed varies. More precisely, we perform an exhaustive empirical study that conclusively shows that increasing processor speed often reduces performance and wastes energy. Our experimental work shows that the frequency of occurrence and magnitude of slowdowns grow with clock frequency and parallelism, indicating that such slowdowns will increasingly be observed with trends in processor and system design.;Performance speedups at lower frequencies (or slowdowns at higher frequencies) have been anecdotally observed in the prevailing literature since 2004, but no research has explained nor exploited this phenomenon. This work conclusively demonstrates that performance slowdowns during processor speedup phases can exceed 47% in common I/O workloads. Our hypothesis challenges (and ultimately debunks) a fundamental assumption in computer systems: faster processor speeds result in the same or better performance.;In this work, with the use of code and kernel instrumentation, exhaustive experiments, and deep insight into the inner workings of the Linux I/O subsystem, I overcome the aforementioned challenges of variance, complexity, and nondeterminism and identify the I/O resource contention as the root cause of the slowdowns during processor speedup. Specifically, such contention comes from the Linux kernel when the journaling block device (JBD) interacts with the ext3/4 file system that introduces file write delays and file synchronization delays. To fully explain how such I/O contention causes performance anomaly, I propose analytical models of resource contention among I/O threads to describe the root cause of the observed I/O slowdowns when processors speed up. To this end, I introduce LUC, a runtime system to limit the unintended consequences of power scaling and demonstrate the effectiveness of the LUC system for two critical parallel transaction-oriented workloads, including a mail server (varMail) and online transaction processing (oltp).