Doctor of Philosophy (Ph.D.)
This dissertation presents an analysis of performance effects of burstiness (formalized by the autocorrelation function) in multi-tiered systems via a 3-pronged approach, i.e., experimental measurements, analytic models, and policy development. This analysis considers (a) systems with finite buffers (e.g., systems with admission control that effectively operate as closed systems) and (b) systems with infinite buffers (i.e., systems that operate as open systems).;For multi-tiered systems with a finite buffer size, experimental measurements show that if autocorrelation exists in any of the tiers in a multi-tiered system, then autocorrelation propagates to all tiers of the system. The presence of autocorrelated flows in all tiers significantly degrades performance. Workload characterization in a real experimental environment driven by the TPC-W benchmark confirms the existence of autocorrelated flows, which originate from the autocorrelated service process of one of the tiers. A simple model is devised that captures the observed behavior. The model is in excellent agreement with experimental measurements and captures the propagation of autocorrelation in the multi-tiered system as well as the resulting performance trends.;For systems with an infinite buffer size, this study focuses on analytic models by proposing and comparing two families of approximations for the departure process of a BMAP/MAP/1 queue that admits batch correlated flows, and whose service time process may be autocorrelated. One approximation is based on the ETAQA methodology for the solution of M/G/1-type processes and the other arises from lumpability rules. Formal proofs are provided: both approximations preserve the marginal distribution of the inter-departure times and their initial correlation structures.;This dissertation also demonstrates how the knowledge of autocorrelation can be used to effectively improve system performance, D_EQAL, a new load balancing policy for clusters with dependent arrivals is proposed. D_EQAL separates jobs to servers according to their sizes as traditional load balancing policies do, but this separation is biased by the effort to reduce performance loss due to autocorrelation in the streams of jobs that are directed to each server. as a result of this, not all servers are equally utilized (i.e., the load in the system becomes unbalanced) but performance benefits of this load unbalancing are significant.
© The Author
Zhang, Qi, "The effect of workload dependence in systems: Experimental evaluation, analytic models, and policy development" (2006). Dissertations, Theses, and Masters Projects. Paper 1539623507.