Load-balancing algorithms for data center networks

Abstract

Multipath load-balancing algorithms, which can be used for data center networks (DCNs), are provided. A multipath load-balancing algorithm can be, for example, a distributed multipath load-balancing algorithm or a centralized multipath load-balancing algorithm. Algorithms of the subject invention can be used for, e.g., hierarchical DCNs and / or fat-tree DCNs. Algorithms of the subject invention are effective and scalable and significantly outperform existing solutions.

Description

CROSS REFERENCE TO RELATED APPLICATION[0001]This application is a continuation application of U.S. application Ser. No. 14 / 297,125, filed Jun. 5, 2014, which claims the benefit of U.S. Provisional Application Ser. No. 61 / 832,458, filed Jun. 7, 2013, the disclosures of which are all hereby incorporated by reference in their entireties, including any figures, tables, and drawings.BACKGROUND OF INVENTION[0002]Data centers contain large numbers of servers to achieve economies of scale [19], and the number is increasing exponentially [13]. For example, it is estimated that Microsoft's Chicago data center has about 300,000 servers [1]. The huge number of servers has created a challenge for the data center network (DCN) to offer proportionally large bandwidth to interconnect the servers [30]. As a result, modern DCNs usually adopt multi-rooted hierarchical topologies, such as the fat tree [2], VL2 [11], DCell [13], and BCube [12], which offer multipath capability for large bisection bandwi...

AI Technical Summary

Benefits of technology

[0006]Embodiments of the subject invention provide multipath load-balancing algorithms, which can be used for data center networks (DCNs). In several embodiments, a multipath load-balancing algorithm is a distributed multipath load-balancing algorithm. In several embodiments, a multipath load-balancing algorithm is a centralized multipath load-balancing algorithm. Algorithms of the subject invention can be used for, e.g., hierarchical DCNs and / or fat-tree DCNs. Algorithms of the subject invention advantageously significantly outperform existing solutions. In addition, the designs of algorithms of the subject invention are effective and scalable.

Problems solved by technology

The huge number of servers has created a challenge for the data center network (DCN) to offer proportionally large bandwidth to interconnect the servers [30].

However, traditional link state and distance vector based [16] routing algorithms (e.g., for the internet) cannot readily utilize the multipath capability of multi-rooted topologies.

Although equal cost multipath (ECMP) [9] supports multipath routing, it performs static load-splitting based on packet headers without accounting for bandwidth, allows only paths of the same minimum cost, and supports an insufficiently small number of paths [14].

However, ECMP does not account for flow bandwidth in making allocation decisions, which may lead to oversubscription even for simple communication patterns.

The latter is a fast heuristic based on the topological feature of fat trees, but with the impractical assumption to split a flow among multiple paths.

Images