More or Less: Ensuring Tail Performance by Speculation and Approximation
Speaker: Dr Lydia Chen
In this talk, I will discuss some of our on-going efforts in minimizing the tail latency by introducing approximation and speculation. I will present both analytical and implementation results. To minimize the tail latency when real time processing big data, we develop AccStream, an approximation layer that enables input and task dropping policies on top of Spark streaming. We also derive stochastic models to guide the optimal tradeoff between the accuracy and the latency. To combat the performance instability due to the resource sharing and heterogeneous hardware in cloud, we develop sPARE, a variability-aware replication system that performs differentiated and partial replication for read workload.
Lydia Y. Chen is a research staff member at the IBM Zurich Research Lab, Zurich, Switzerland. She received a Ph.D. in Operations Research from the Pennsylvania State University in 2006. Her research interests include modeling, optimizing performance and dependability for big data applications and highly virtualized datacenters. She has published papers in international conferences and journals, e.g., Sigmetrics, DSN, INFOCOM, and IEEE transactions on Service Computing. She is the co-recipients of best paper awards at CCgrid'15, and eEnergy'15. She served on technical program committees for system and network conferences, e.g., DSN, ICDCS, Middlware and INFOCOM. She has lead and participated in Swiss National Science Fundation and European FP7 projects. She is a IEEE senior member.