2015 Informs Annual Meeting

SD20

INFORMS Philadelphia – 2015

SD20 20-Franklin 10, Marriott Queueing with Redundancy for Cloud Computing Cluster: Cloud Computing Invited Session Chair: Mor Harchol-Balter, Professor, Carnegie Mellon University, Computer Science Dept., 5000 Forbes Ave., Pittsburgh, United States of America, harchol@cs.cmu.edu 1 - Exact Queueing Analysis of Redundancy-d Systems Mor Harchol-Balter, Professor, Carnegie Mellon University, Computer Science Dept., 5000 Forbes Ave., Pittsburgh, United States of America, harchol@cs.cmu.edu, Kristen Gardner, Sam Zbarsky, Alan Scheller-wolf Recent cloud research has proposed using redundant requests to reduce latency by copying a request to multiple servers and waiting for only one copy to complete. We study the Redundancy-d system, in which each arriving job sends copies to d randomly selected servers. We provide the limiting state distribution. We also derive the exact mean response time in this system for any number of servers and any degree of redundancy. 2 - Scaling Redundancy to Many-server Systems Kristen Gardner, PhD Student, Carnegie Mellon University, Computer Science Dept., 5000 Forbes Ave., Pittsburgh, PA, 15213, ksgardne@cs.cmu.edu, Mor Harchol-Balter, Alan Scheller-wolf, Sam Zbarsky This talk is a continuation of the previous talk on the Redundancy-d system. Here, we study the system in the limit as the number of servers approaches infinity. We derive the full response time distribution and use this result to discuss Rhonda Righter, Professor, University of California, Berkeley, IEOR, UC, Berkeley, CA, 94720, United States of America, rrighter@berkeley.edu, Mor Harchol-Balter, Esa Hyytia We consider systems in which there is a mix of tasks that can be replicated and tasks that cannot. We explore the effect of the amount of replication on latency, and find the optimal service discipline in the presence of partial replication. 4 - Analysis and Routing in Parallel Queues with Class-based Redundancy Leela Nageswaran, PhD Candidate, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, United States of America, lnageswa@andrew.cmu.edu, Alan Scheller-wolf We study the performance of two parallel queues when some customers are redundant: a redundant customer joins both queues and is considered served when any one of his requests finishes service, instantly removing the other one. We examine the policy other (non-redundant) customers use to join a queue upon arrival. We find that while joining the shortest queue does not always minimize delay if the entire state information is available, it is optimal if only the queue lengths are observable. SD21 21-Franklin 11, Marriott Natural History Modeling for Medical Decision Making Sponsor: Health Applications Sponsored Session Chair: Julie Higle, Professor And Chair, University of Southern California, Epstein Dept of Indus & Sys Eng, Los Angeles, CA, 90089, United States of America, higle@usc.edu 1 - Developing and Validating Markov Decision Processes for Chronic Diseases Brian Denton, Associate Professor, University of Michigan, 1205 Beal Ave, Ann Arbor, United States of America, btdenton@umich.edu Chronic diseases are the leading cause of death in many countries including the United States. Building models for chronic diseases can be challenging because of the need to characterize severity of the disease, uncertainty in disease progression, and a potentially large number of strategies for screening and treatment. In this talk I will discuss my experiences in developing Markov decision processes (MDPs) for optimization of disease screening and treatment decisions in several contexts. the effect of the number of copies per job on response time. 3 - Optimal Scheduling of Partially Replicated Jobs

2 - Challenges and Opportunities for Developing Natural History Models

Oguzhan Alagoz, UW-Madison, 3242 Mechanical Engineering Building, 1513 University Aveneue, Madison, WI, 53706, United States of America, alagoz@engr.wisc.edu Natural history of a disease, which represents the onset and progression of a disease without an intervention, provides critical inputs for operations research models in health care. In this presentation, we will share our experiences in developing natural history models for various diseases including end-stage liver diseases and breast cancer. 3 - Robust Parameter Selection for Natural History Models Julie Higle, Professor And Chair, University of Southern California, Epstein Dept of Indus & Sys Eng, Los Angeles, CA, 90089, United States of America, higle@usc.edu, Suvrajeet Sen Natural history models are often used to facilitate an understanding of the potential impact of disease screening and/or treatment options. We consider a method for calibrating a natural history model that explicitly considers uncertainties in the calibration targets. The calibration model is designed to yield a robust parameter selection, especially with respect to medical decisions that result. 4 - Modeling Ductal Carcinoma in Situ (DCIS) Shadi Hassan Goodarzi, PhD Student, North Carolina State University, Fitts Dept of ISE, Raleigh, No, 27695, United States of America, shassan3@ncsu.edu, Julie Ivy Ductal Carcinoma In Situ (DCIS) is arguably a direct precursor of invasive breast cancer. Approximately 14–53% of DCIS turn into IBC, after long follow-up periods. So about 47%-86% of the DCIS cases are over diagnosed and as a result, treatment can only cause harm for these patients. This framework will allow us to study the progression of DCIS into IBC more clearly and as a result aid both patients and doctors in decision making. SD23 23-Franklin 13, Marriott Queues in Heavy-Traffic: Approximations and Control Sponsor: Applied Probability Sponsored Session Chair: Yunan Liu, Assistant Professor, North Carolina State University, 111 Lampe Drive, #400, Raleigh, No, 27695, United States of America, yunan_liu@ncsu.edu 1 - A Many-Server Heavy-Traffic Limit for the Overloaded G_t/gi/n+gi Queue Ahmet Korhan Aras, North Carolina State University, 307 Daniels Hall, Raleigh, NC, United States of America, akaras@ncsu.edu, Yunan Liu, Xinyun Chen, Ward Whitt We establish a many-server heavy-traffic FCLT for key performance processes such as potential waiting time, number of abandonment and queue length for the G_t/GI/n+GI queue in the overloaded regime. We obtain a stochastic differential equation driven by a Gaussian process in the limit for the scaled waiting time process. The Gaussian limit and Gaussian integral appear in the limit of the departure process which is not a Brownian motion when the service distribution is not exponential. 2 - Diffusion Approximation for Efficiency-driven Queues: A Space- time Scaling Approach Using a scaling approach in both space and time, we obtain a diffusion model for the virtual waiting time process in a GI/GI/n+GI queue in the ED regime. Besides the commonly used scaling in space by the number of servers, we also change the time scale by using the mean patience time as the factor. This approach leads to a simple one-dimensional diffusion limit, enabling us to obtain useful performance formulas such as the distributions of the steady-state virtual waiting time and queue length. 3 - Non-markovian State-dependent Networks in Critical Loading Chihoon Lee, Stevens Institute of Technology, Howe School of We establish a heavy traffic limit theorem for the queue-length process in a critically loaded single class queueing network with state-dependent arrival and service rates. A distinguishing feature of our model is non-Markovian state dependence. The limit stochastic process is a continuous-path reflected process on the nonnegative orthant. We give an application to a generalised Jackson network with state-dependent rates. Technology Management, Hoboken, NJ, United States of America, chihoon@stat.colostate.edu, Anatolii Puhalskii Shuangchi He, National University of Singapore, National University of Singapore, Singapore, heshuangchi@nus.edu.sg

126