Diagnosability Evaluation of the Data Center Network DCell

Abstract

With the rapid development of cloud computing, many large-scale data centers are being built to provide increasingly popular online application services. This leads to the proposal of data center networks (DCNs) supporting millions of servers with high-network capacity by using only commodity switches. The k-dimensional DCell with n-port switches and tk,n servers, Dk,n⁠, has been proposed as a model for a large-scale DCN with a server-centric structure. In this paper, we study the diagnosability and the g-good-neighbor conditional diagnosability of Dk,n⁠. We prove that: (i) Dk,n is (n+k−1)-diagnosable under the precise diagnosis strategy and (2k+n−2)/(2k+n−2)- diagnosable under the pessimistic diagnosis strategy; (ii) the g-good-neighbor conditional diagnosabilities of Dk,n under the PMC model and the MM* model are both (g+1)k+n−1 (resp. (n+k−g)tg−n+1,n−1) with 0≤g≤n−1 (resp. n≤g≤n+k−2⁠), which is almost (g+1) (resp. tg−n+1,n⁠) times of the traditional diagnosability. These results provide a quantitative evaluation for a large-scale DCN’s reliability and availability.