Icpc2011 syer
- 1. Supporting the Comprehension of System Behaviour Under Load Mark D. Syer, Bram Adams and Ahmed E. Hassan http://sail.cs.queensu.ca/~mdsyer How To Understand ULS Systems?How To Understand ULS Systems? Case Study: Thread PoolsCase Study: Thread Pools Ultra-large-scale (ULS) systems dominate the fields of e-commerce and telecommunications. !Highly concurrent and highly distributed. Behaviour is hard to understand! !Complex architectures built on complex sub-architectures We propose: • an iterative methodology • for automatically identifying and ranking deviating behaviour • based on the level of dissimilarity between the resource usage metrics of threads. Iterative MethodologyIterative Methodology 2. Distance Calculation2. Distance Calculation1. Metric Abstraction1. Metric Abstraction 4. Ranking4. Ranking 3. Hierarchical Clustering3. Hierarchical Clustering Conclusions and Future WorkConclusions and Future Work Our methodology has enabled us: • Develop a better understanding of the subject system • Identify and rank similar and deviating behaviour • Tackle different problems where resource usage _metrics are collected Threads pools face a variety of challenges: • Synchronization errors • Deadlock Thorough understanding of the system is needed to address these challenges The subject system in our case study was a commercial ULS system that implemented the typical thread pool architecture. Abstracted Metrics Distance Matrix Clusters Service Request Service Request Handler Thread 1 Thread 2 Thread n … Load Balancer Risks? => New Level of Abstraction Threads Within A Load TestThreads Within A Load Test letters a b aa ab ba bb aaa aab aba abb baa bab bba bbb aa ab bbba aa ab bbba Covariance Matrices - - - - - - - - - - - - - - - Individual threads - - - - - - - - - - - - - - - Threads Between Load TestsThreads Between Load Tests • Resource thrashing • Thread leakage T1W1 T1W8 T1W2, T1W3, T1W4, T1W5, T1W6, T1W7. Future work: • More systems • More design patterns and architectural styles • Improved ranking algorithm (e.g. load testing) .5 Ramp up Ramp down Iterative aa ab ba bb Height 00.51.01.52.02.5 Height aa ab ba bb 00.51.01.52.02.5