Trace-Based Workload Generation and Execution

• Published in: Euro-Par 2021: Parallel Processing pp. 37 - 54
• Main Authors: Sfakianakis, Yannis, Kanellou, Eleni, Marazakis, Manolis, Bilas, Angelos
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Benchmarking; Cloud computing; Simulation; Trace; Workload
• ISBN: 9783030856649; 303085664X
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-85665-6_3

Although major cloud providers have captured and published workload executions in the form of traces, it is not clear how to use them for workload generation on a wide range of existing platforms. A methodological challenge that remains is to generate and execute realistic datacenter workloads on any infrastructure, using information from available traces. In this paper, we propose Tracie, a methodology addressing this challenge, and introduce the tool supporting its implementation. We present all the necessary steps starting from a trace up to workload execution: analysis of datacenter traces, extraction of parameters, application selection, and scaling of a workload to match the capabilities of the underlying infrastructure. Our evaluation validates that Tracie can generate executable workloads that closely resemble their trace-based counterparts. For validation, we correlate the recorded system metrics of a trace against the actual execution. We find that the average system metrics of synthetic workloads differ at most 5% compared to the trace and that they are highly correlated at 70% on average.