Predicting Intermediate Storage Performance for Workflow Applications

TL;DR

This paper introduces a performance prediction mechanism for storage systems tailored to workflow applications, enabling efficient configuration tuning by estimating application turnaround times with high accuracy and significant speedup.

Contribution

We designed an end-to-end prediction system that models storage performance at data and control levels, supporting various optimizations and configurations for workflow applications.

Findings

Achieves over 200x speedup in predicting application performance.

Provides accurate configuration recommendations for storage systems.

Scales to model entire cluster workflows effectively.

Abstract

Configuring a storage system to better serve an application is a challenging task complicated by a multidimensional, discrete configuration space and the high cost of space exploration (e.g., by running the application with different storage configurations). To enable selecting the best configuration in a reasonable time, we design an end-to-end performance prediction mechanism that estimates the turn-around time of an application using storage system under a given configuration. This approach focuses on a generic object-based storage system design, supports exploring the impact of optimizations targeting workflow applications (e.g., various data placement schemes) in addition to other, more traditional, configuration knobs (e.g., stripe size or replication level), and models the system operation at data-chunk and control message level. This paper presents our experience to date with designing and using this prediction mechanism. We evaluate this mechanism using micro- as well as synthetic benchmarks mimicking real workflow applications, and a real application.. A preliminary evaluation shows that we are on a good track to meet our objectives: it can scale to model a workflow application run on an entire cluster while offering an over 200x speedup factor (normalized by resource) compared to running the actual application, and can achieve, in the limited number of scenarios we study, a prediction accuracy that enables identifying the best storage system configuration.