Filling the Gap: A Tool to Automate Parameter Estimation for - - PowerPoint PPT Presentation
Filling the Gap: A Tool to Automate Parameter Estimation for Software Performance Models Weikun Wang, Juan F. P erez, Giuliano Casale Department of Computing Imperial College London weikun.wang11@imperial.ac.uk September 1, 2015 1/13 W.
September 1, 2015
1/13
◮ DevOps - a recent trend in Software
◮ Bridges the gap between software development
◮ Use performance models for QoS analysis ◮ Accurate parametrization is challenging
2/13
◮ Continuous performance model parametrization ◮ Advanced estimation algorithms ◮ Statistical inference from monitoring data ◮ Application QoS report generation
3/13
◮ FG Local DB: monitoring data storage ◮ FG Analyzer: statistical analysis ◮ FG Actuator: performance model update ◮ FG Reporter: application performance report
4/13
FG Analyzer FG Report FG Actuator Deployment Module Monitoring Platform QoS Model Monitoring History DB FG Local DB Computing Cluster
5/13
◮ An important parameter of queueing models ◮ Difficult to obtain directly ◮ Extensive monitoring poses overhead
6/13
◮ Complete Information (CI) ◮ Gibbs sampling with Queue Lengths (GQL) ◮ MINPS/FMLPS ◮ Extended Regression-Based approach (ERPS) ◮ FCFS ◮ Utilization-Based Regression/Optimization
7/13
◮ Uses full trace: ts. of arrivals and departures ◮ Poses additional overhead for intensive
◮ Requires queue length samples, i.e. number of
◮ Estimates demand with Bayes’ theorem ◮ Uses Gibbs sampling to obtain demand
8/13
◮ MINPS: a maximum likelihood (ML) method ◮ FMPLS: ML method with fluid approximation ◮ Both requires response times and queue lengths
9/13
◮ Requires response time and queue length
◮ Linear regression to obtain demand
◮ Estimation for FCFS servers ◮ Requires response time and queue length
◮ Linear regression to obtain demand
10/13
11/13
Algorithms CI MINPSERPS GQL UBR UBO FMLPS Error (%) 10 20 30 40 50 60 70
(a) Error (%)
Algorithms CI MINPSERPS GQL UBR UBO FMLPS Execution time (s) 10-3 10-2 10-1 100 101 102 103
(b) Execution time (s)
12/13
◮ i) The feedback for different demand
◮ How much monitoring information can be timely
◮ Which metric is the easiest or most readily
◮ Which metric poses the least overhead?
◮ ii) How to correlate the resource IDs as well as
13/13