Feature Based Representative Workload Construction from Traces

Abstract

Storage system traces are rich in information as it contains real-world behavior. Replaying already recorded traces is used to reproduce the realistic behavior of systems as accurately as possible. The growing popularity of object storage systems and less focus on creating precise trace replay workload leads this work to focus on identifying components core to currently existing object stores. In this work, a novel features-based method for constructing representative traces from original traces is proposed, where the former serves as an approximation with information from later. Time-related patterns that occurred in the original trace are kept in the representative trace to exhibit the trigger time of operations and application profile. Users can configure the proposed pipeline to evaluate object storage systems and any other storage or file system that uses trace replay workload. A quantifying metric is proposed, which helps analyze the created workload accurately. This also assists in investigating the coverage of features in the representative workload from the original workload. This work tries to solve a computationally complex problem with the help of a greedy algorithm that uses submodularity and guarantees the representation of distinct attributes in the final set of operations. Using the comparison of different configuration types and varying settings of parameters, we show that the method proposed can be easily extended to other applications or for benchmarking of any other system. This work facilitates understanding traces and contributes to improving links to the factors of the system under evaluation.