Скачать или смотреть Methodology for performance benchmarking of containerized microservices using CMS memory pooling

Presented by Vikrant Soman (Uber)

Uber adopted a microservice oriented architecture to support high feature velocity, at present, we have 1000s of containerized microservices. Instead of attempting to create benchmarks using real services / mocked up services, we took a completely different approach to measuring microservice “performance” on our hardware. We created a service agnostic performance abstraction, by profiling hundreds of hosts in our fleet using Intels top-down telemetry package. Analyzing this data provided us a clear top-down pattern of which resources within the core pipeline and memory subsystem were heavily used when running our microservices in production. From this data, we then selected a few SPECCPU2017 subcomponent(s) whose top-down profile(s) closely matched the “top down profile” seen in our production fleet. We propose a similar approach can be applied to study the performance of container workloads using CMS memory pooled solution