Workshop: 8th Workshop on Programming and Performance Visualization Tools (ProTools)
Authors: Viraj Thakkar (Arizona State University, HPE) and Timo Soekler, John Byrne, Myungjun Son, Sekwon Lee, Harumi Kuno, and Aditya Dhakal (HPE)
Abstract: We demonstrate that although eBPF and bpftrace are widely used for low-overhead tracing, their overheads are large when considering microsecond-scale events. We decompose these overheads into two parts that were not considered separately by prior work: a fixed floor paid for every probe that fires, and per-operation work layered on top. We measure these overheads on a controlled single-syscall harness, then characterize how overheads scale as probes accumulate: overhead is driven by the number of distinct events that fire (~42 ns each), not by the number of probe definitions (~3 ns per additional probe on an already-traced event). For example, a bare probe costs ~50 ns while a single map update or timestamp adds 100–300 ns. From these measurements we derive an observability budget that predicts, before any probe is attached, how much a proposed trace will distort its target and therefore how faithfully the event can be observed.
Back to 8th Workshop on Programming and Performance Visualization Tools (ProTools) Archive Listing
Back to Full Workshop Archive Listing