ModelGraft: Accurate, Scalable, and Flexible Performance Evaluation of General Cache Networks

Abstract

Large scale deployments of general cache networks, such as Content Delivery Networks or Information Centric Networking architectures, arise new challenges regarding their performance evaluation for network planning. On the one hand, analytical models can hardly represent in details all the interactions of complex replacement, replication, and routing policies on arbitrary topologies. On the other hand, the sheer size of networks and content catalogs makes event-driven simulation techniques inherently non-scalable. We propose a new technique for the performance evaluation of large-scale caching systems that intelligently integrates elements of stochastic analysis within a MonteCarlo simulative approach, that we colloquially refer to as ModelGraft. Our approach (i) leverages the intuition that complex scenarios can be mapped to a simpler equivalent scenario that builds upon Time-To-Live (TTL) caches; it (ii) significantly downscales the scenario to lower computation and memory complexity, while, at the same time, preserving its properties to limit accuracy loss; finally, it (iii) is simple to use and robust, as it autonomously converges to a consistent state through a feedback-loop control system, regardless of the initial state. Performance evaluation shows that, with respect to classic event-driven simulation, ModelGraft gains over two orders of magnitude in both CPU time and memory complexity, while limiting accuracy loss below 2%. In addition, we show that ModelGraft extends performance evaluation well beyond the boundaries of classic approaches, by enabling study of Internetscale scenarios with content catalogs comprising hundreds of billions objects.

Fairness in online social network timelines: Measurements, models and mechanism design

“…The proposed News Feed model relies on TTL counters. TTL-based caching mechanisms are versatile and flexible, and can be used to reproduce the behavior of traditional caching policies such as LRU and FIFO [37,81,21,36,64]. In this paper, we leverage the analytical tractability of TTL-based caches, showing how to adapt them for the purposes of modeling and optimization of the News Feed.…”

A hybrid methodology for the performance evaluation of Internet-scale cache networks

“…As a consequence, the pivotal role of inversion rejection sampling in studying cache dynamics at Internet-scale (i.e., with trillion objects [33]) becomes clear. Formally (proof available in [45]):…”

“…where W is the size of the sample window, and ε CV is a user-defined convergence threshold (a sensitivity analysis of ε CV is extensively reported in [45] and briefly summarized in Sec.8.5). To avoid biases, new samples are collected only if (i) the cache has received a non-null number of requests since the last sample, and (ii) its state has changed, i.e., at least a new content has been admitted in the cache since the last sample 2 .…”

“…In the remainder of this paper, of which a preliminary version appeared at [44], we first formalize the modeling background at the base of ModelGraft (Sec. 2).…”

New Caching System Under Uncertainty for Mobile Edge Computing

“…In case of a full cache, the system evicts the content which was earliest stored and replaces it by the new content. For the distribution of requests and contents, studies have confirmed Zipfs law as an appropriate distribution model for access pattern to contents on the Internet such as videos hosted on YouTube [23] or peer-to-peer file sharing systems (Bit-Torrent, Gnutella) [24] [25]. According to Zipfs law, a small fraction of popular web objects attracts most user requests, which is favourable for the efficiency of small caches [25].…”