Workshop: PDSW’26: The 11th International Parallel Data Systems Workshop
Authors: Alexander Klassen and Osamu Tatebe (University of Tsukuba)
Abstract: STARCH is a low-perturbation LTTng-based tracing system targeting RPC-Lifecycles, designed to instrument Margo-based applications without any code modification on the target. Per-node collectors aggregate events from clients and servers, while causal association of a client request to a server response is handled outside of the HPC job. The resulting trace database can be used to generate statistics and test many hypotheses without needing to run multiple manually instrumented experiments. In this paper we use STARCH to analyze a known CHFS performance degradation using IOR, and present the potential general applicability of STARCH on GekkoFS. STARCH typically incurs under 1.7\% overhead on our IOR+CHFS benchmark, and reliably collects and reconstructs tracing data in a manageable per-node data volume. STARCH was able to surface a nontrivial performance finding about CHFS, reducing the problem scope to a specific area of RPC handling.
Back to PDSW’26: The 11th International Parallel Data Systems Workshop Archive Listing
Back to Full Workshop Archive Listing