Giuliano Casale - Imperial College London

A Quantitative Approach to Sizing IT Applications Processing Time-Varying Workloads

An important challenge of current research in capacity planning is to understand the performance impact of burstiness on computer systems.

Burstiness arises due to flash crowds, caching, and dependence between consecutive requests, and it is well-known to be a key performance driver. In this presentation, I will introduce a methodology to size enterprise systems processing workloads affected by burstiness.

I will introduce a model-based technique for the automatic and controlled generation of bursty benchmarks. Here, Markov models are constructed in an automated manner to model the distribution of service demands placed by user sessions on various system resources. These models are then used to derive session submission policies that result in user-specified levels of service demand burstiness for resources at the different tiers in a system. Our approach can also predict under what conditions these policies can create dynamic bottleneck switching among resources.

I will also summarize the main features of a new family of analytic methods that can accurately describe burstiness and temporal locality effects. An approximate analysis technique will be presented, together with a methodology to parameterize these models from limited measurements. Experiments on a real-world TPC-W testbed will be shown to demonstrate the good accuracy of the proposed techniques compared to those of classic approaches that ignore burstiness and temporal dependence.