Most of the paper’s authors are connected to Google, and their work centers on Google workload performance improvement. They’ve developed runtime controller software, Heracles, that uses real-time feedback and static modeling rules to adjust resource allocation within servers in order to meet service-level objectives (SLOs). The paper’s closing section demonstrates that Heracles improved performance in test systems. It’s great that they’ve made Google faster, but what use is this Google performance tool to the rest of us?

Unless and until they make Heracles freely downloadable (and we have server applications that can make use of it), the tool itself is of no general interest. The value of this paper, though, lies not in the end product, but in the journey that led to the end product. The first nine or ten pages describe the authors’ analysis of resource contention, specifically the interplay between latency critical (LC) tasks and noncritical, best-effort (BE) tasks. All applications, from Internet of Things (IoT) to the cloud, on smartphones and in data center servers, face the problem of guaranteeing quick response from critical services despite the unpredictable activity of background tasks. At present, we solve the problem by over allocating resources. We throw money at the problem, paying for peak usage scenarios, while day in and day out we tolerate idle central processing units (CPUs) and underutilized storage. The analysis that led to Heracles, summarized in this paper, brings us a step closer to building efficient systems.

The authors have created a template we can all use to analyze resource contention. They also identify specific tools and techniques used to address contention issues, quantify performance improvements achieved by using those tools, and survey numerous research contributors for further investigation. The rest of us will probably never use Heracles, but we can all use this advice to improve our own little corner of the universe.