Abstract
Modern On-Line Data Intensive (OLDI) applications have evolved from monolithic systems to instead comprise numerous, distributed microservices interacting via Remote Procedure Calls (RPCs). Microservices face single-digit millisecond RPC latency goals (implying sub-ms medians)---much tighter than their monolithic ancestors that must meet ≥ 100 ms latency targets. Sub-ms-scale OS/network overheads that were once insignificant for such monoliths can now come to dominate in the sub-msscale microservice regime. It is therefore vital to characterize the influence of OS- and network-based effects on microservices. Unfortunately, widely-used academic data center benchmark suites are unsuitable to aid this characterization as they (1) use monolithic rather than microservice architectures, and (2) largely have request service times ≥ 100 ms. In this paper, we investigate how OS and network overheads impact microservice median and tail latency by developing a complete suite of microservices called µSuite that we use to facilitate our study. µSuite comprises four OLDI services composed of microservices: image similarity search, protocol routing for key-value stores, set algebra on posting lists for document search, and recommender systems. Our characterization reveals that the relationship between optimal OS/network parameters and service load is complex. Our primary finding is that non-optimal OS scheduler decisions can degrade microservice tail latency by up to ∼87%.
Akshitha Sriraman
Assistant Professor
I am an Assistant Professor in the Department of Electrical and Computer Engineering at Carnegie Mellon University. My research bridges computer architecture and software systems, with a focus on making hyperscale data center systems more efficient (via solutions that span the systems stack). The central theme of my work is to (1) design software that is aware of new hardware constraints and (2) architect hardware that efficiently supports new hyperscale software requirements.