Recent computer systems research has proposed using redundant requests to reduce latency. The idea is to run a request on multiple servers and wait for the first completion (discarding all remaining copies of the request). However, there is no exact analysis of systems with redundancy. This paper presents the first exact analysis of systems with redundancy. We allow for any number of classes of redundant requests, any number of classes of non-redundant requests, any degree of redundancy, and any number of heterogeneous servers. In all cases we derive the limiting distribution of the state of the system. In small (two or three server) systems, we derive simple forms for the distribution of response time of both the redundant classes and non-redundant classes, and we quantify the "gain" to redundant classes and "pain" to non-redundant classes caused by redundancy. We find some surprising results. First, the response time of a fully redundant class follows a simple exponential distribution and that of the non-redundant class follows a generalized hyperexponential. Second, fully redundant classes are "immune" to any pain caused by other classes becoming redundant. We also compare redundancy with other approaches for reducing latency, such as optimal probabilistic splitting of a class among servers (Opt-Split) and join-the-shortest-queue (JSQ) routing of a class. We find that, in many cases, redundancy outperforms JSQ and Opt-Split with respect to overall response time, making it an attractive solution.

Recent computer systems research has proposed using redundant requests to reduce latency. The idea is to run a request on multiple servers and wait for the first completion (discarding all remaining copies of the request). However there is no exact analysis of systems with redundancy.This paper presents the first exact analysis of systems with redundancy. We allow for any number of classes of redundant requests, any number of classes of non-redundant requests, any degree of redundancy, and any number of heterogeneous servers. In all cases we derive the limiting distribution on the state of the system.In small (two or three server) systems, we derive simple forms for the distribution of response time of both the redundant classes and non-redundant classes, and we quantify the "gain" to redundant classes and "pain" to non-redundant classes caused by redundancy. We find some surprising results. First, the response time of a fully redundant class follows a simple Exponential distribution and that of the non-redundant class follows a Generalized Hyperexponential. Second, fully redundant classes are "immune" to any pain caused by other classes becoming redundant.We also compare redundancy with other approaches for reducing latency, such as optimal probabilistic splitting of a class among servers (Opt-Split) and Join-the-Shortest-Queue (JSQ) routing of a class. We find that, in many cases, redundancy outperforms JSQ and Opt-Split with respect to overall response time, making it an attractive solution.

No abstract

Recent computer systems research has proposed using redundant requests to reduce latency. The idea is to run a request on multiple servers and wait for the first completion (discarding all remaining copies of the request). However there is no exact analysis of systems with redundancy. This paper presents the first exact analysis of systems with redundancy. We allow for any number of classes of redundant requests, any number of classes of non-redundant requests, any degree of redundancy, and any number of heterogeneous servers. In all cases we derive the limiting distribution on the state of the system. In small (two or three server) systems, we derive simple forms for the distribution of response time of both the redundant classes and non-redundant classes, and we quantify the "gain" to redundant classes and "pain" to non-redundant classes caused by redundancy. We find some surprising results. First, the response time of a fully redundant class follows a simple Exponential distribution and that of the non-redundant class follows a Generalized Hyperexponential. Second, fully redundant classes are "immune" to any pain caused by other classes becoming redundant. We also compare redundancy with other approaches for reducing latency, such as optimal probabilistic splitting of a class among servers (Opt-Split) and Join-the-Shortest-Queue (JSQ) routing of a class. We find that, in many cases, redundancy outperforms JSQ and Opt-Split with respect to overall response time, making it an attractive solution.

The M/M/k/setup model, where there is a penalty for turning servers on, is common in data centers, call centers and manufacturing systems. Setup costs take the form of a time delay, and sometimes there is additionally a power penalty, as in the case of data centers. While the M/M/1/setup was exactly analyzed in 1964, no exact analysis exists to date for the M/M/k/setup with k > 1.In this paper we provide the first exact, closed-form analysis for the M/M/k/setup and some of its important variants including systems in which idle servers delay for a period of time before turning off or can be put to sleep. Our analysis is made possible by our development of a new technique, Recursive Renewal Reward (RRR), for solving Markov chains with a repeating structure. RRR uses ideas from renewal reward theory and busy period analysis to obtain closed-form expressions for metrics of interest such as the transform of time in system and the transform of power consumed by the system. The simplicity, intuitiveness, and versatility of RRR makes it useful for analyzing Markov chains far beyond the M/M/k/setup. In general, RRR should be used to reduce the analysis of any 2-dimensional Markov chain which is infinite in at most one dimension and repeating to the problem of solving a system of polynomial equations. In the case where all transitions in the repeating portion of the Markov chain are skip-free and all up/down arrows are unidirectional, the resulting system of equations will yield a closed-form solution.