Schneider, D. (2024):

A Symphony of Metrics: Assessing the Advantages of eBPF over conventional Benchmarking Tools

Supercomputers are configured to handle demanding calculations, and their architecture varies based on the workloads they run. Some workloads involve extensive data communication between instances, requiring a high grade of network traffic throughput capabilities of the cluster and an optimized architecture for intercommunication of these workloads. Other workloads may require more hardware-based resources, such as high amounts of system memory, compute performance, and disk speeds, calling for cluster architectures optimized for these resource types. In general, It can be said, that various cluster architectures have their own benefits and non-benefits, on which informed decisions have to be made. Choosing the ideal cluster architecture for a particular workload has been a topic of extensive research. The decision-making process begins with gathering behavioral data on the workload; acting on this data is a separate field of study with various approaches. Most systems utilize a user-space program that can benchmark the workload to make its behavior more measurable.

Recent work on the extended Berkeley Packet Filter (eBPF) has made it possible to bench- mark at a lower level within the kernel space. eBPF provides more control over resources and better access to operating system functions and information that are not easily accessible through user-space programs. This thesis aims to assess eBPF’s advantages in Linux systems compared to conven tional benchmarking solutions. We will develop an eBPF-based benchmarking reference implementation to collect per-process usage statistics, including CPU, Memory, and Disk utilization. Our assessment will compare this system against traditional solutions, focusing on correctness, modularity, simplicity, and scalability.

• PDF Version (974.9 KB)