Online Caching with Convex Costs: Extended Abstract

Modern software applications and services operate nowadays on top of large clusters and datacenters. To reduce the underlying infrastructure cost and increase utilization, different services share the same physical resources (e.g., CPU, bandwidth, I/O, memory). Consequently, the cluster provider often has to decide in real-time how to allocate resources in overbooked systems, taking into account the different characteristics and requirements of users. In this paper, we consider an important problem within this space -- how to share memory between users, whose memory access patterns are unknown in advance. We assume that the overall performance (or cost) of each user is a non-linear function of the total number of misses over a given period of time. We develop an online caching algorithm for arbitrary cost functions. We further provide theoretical guarantees for convex functions (which capture plausible practical scenarios). In particular, our algorithm is α^α k^α-competitive, where k is the memory (cache) size, and α is a constant which depends on the curvature of the cost functions. We also obtain a bi-criteria result which trades-off the performance and the memory size. Finally, we give a lower bound on the performance of any online deterministic algorithm which nearly matches the upper bound of our algorithm.