LinuxCon + ContainerCon North America 2016

Intel introduced cache allocation technology (CAT) that allows partitioning of last-level cache to provide better micro-architectural isolation. Google deployed CAT using intel_rdt cgroup for mixed workloads in Borg. We experimented with various partitioning techniques for different workloads to get better performance and higher utilization.

Preventing batch tasks from affecting latency-sensitive tasks performance through static cache partitioning providing significant performance boost. Additionally, highly latency-sensitive tasks could reduce their footprint and share better when provided exclusive cache access.

In this talk, Rohit and David will cover the motivation for each use-case, describe pre-CAT strategies for interference in Borg, share CAT experiment results and performance/efficiency gains seen in production, interesting problems, and future work for isolation in Borg.

Performance monitoring and analysis using perf and BPF (Wang Nan, HUAWEI) - Jovi (jovi.zhangwei@huawei.com) worked on ktap several years ago. ktap appeared to be a good solution on performance monitoring at 2013. However, due to the dispute about the in-kernel lua virtual machine, ktap's solution is not accepted by the mainline kernel. On 2014, Wang Nan restarted Jovi's work base on perf and BPF. After 2 years development, main part of perf BPF support has been merged into mainline kernel successfully. In this presentation, Wang Nan will summarize his work on perf and BPF, give a brief tutorial about the usage of BPF scripts with perf, introduce many features of perf BPF support like BPF prologue, reading from PMU, and bpf-output perf event. He will also give some HUAWEI's real experience on performance monitoring and analysis using perf and BPF, on server and smart phone.

Workqueue is an asynchronous execution mechanism that has an ubiquitous presence in the Linux Kernel. It's used for various purposes from simple context bouncing to hosting a persistent in-kernel service thread.

In this presentation, we will first discuss several performance issues that are enhanced by the Concurrency Managed Workqueue(CMWQ). We will then look at an example execution scenario which will illustrate how CMWQ serves as a robust async mechanism without introducing any noticeable performance degradation. This will be followed by an overview of the CMWQ API and its design. We will then walk through a number of examples from various parts of the kernel that will offer a comprehensive view of workqueues and its usages.