70 likes | 260 Views
Workload Characterization and Prediction. Joshua Ferguson. Overview. What is a workload?. How are workload patterns matched?. What is the cutting edge?. Characterization Metrics. Workload metrics depend heavily on the type of application
E N D
Workload Characterization and Prediction Joshua Ferguson
Overview • What is a workload? • How are workload patterns matched? • What is the cutting edge?
Characterization Metrics • Workload metrics depend heavily on the type of application • Webservers may consider requests per second, database access rates • HPC environments may consider file access rates, file lifetimes, job length, job submission patterns, job file accesses • Metrics can have different views also, such as device view, data view, or software view. • Device view: A CPU heavy application may use a metric such as the frequency of a specific opcode from its ISA. • Data view: file access rates, or file access relations • Software view: statement branching rates Common technique for choosing relevant metrics: Run numerous benchmarks of that workload type, and note any metric that has high variation amongst benchmarks
Analysis for Prediction • Simple statistical metrics: • Shifting window of Mean, Mode • Last Seen Value • Table Based Prediction: • Record history of metric values, of limited window sizes • For each entry, supply a lifespan counter • At each occurrence, record the next value witnessed and reset the lifespan • Perform prediction actions based on expected value • Probabilistic Technique: • Markov chain of witnessed patterns leads to probability distribution of next possible metric value: • P(s) = P(s1)P(s2|s1)...P(sk|sk−1...s1), for S1, S2, . . .Sk • Metric sample can be a single metric value, or a metric value pattern
Cutting Edge • Since metrics are heavily dependent on workload, they’re constantly being invented as new applications are developed. • SPEC benchmark suite at http://www.spec.org/ has a large variety of benchmarks ranging from desktop application emulation to HPC and virtualization • Statistical Metric Modeling, a probabilistic prediction technique by Sarikaya et al. 2010. • Adept at predicting long-term patterns • International Symposium on Workload Characterization • http://www.iiswc.org