MaxPass: Credit-Based Multipath Transmission for Load Balancing in Data Centers

Abstract

Various applications require the data center networks to carry their traffic efficiently. The data center networks usually have a hierarchical topology and exhibit distinct traffic patterns, which is different from the traditional Internet. These features have driven the data center networks to reduce the flow completion time (FCT) and to achieve high throughput. One of the possible solutions is balancing network loads across multiple paths by leveraging transport mechanisms like equal-cost multipath (ECMP) routing. ECMP allows flows to exploit multiple paths by hashing the meta-data of the flows. However, due to the random nature of hash functions, ECMP often distributes the traffic unevenly, which makes it hard to utilize the links' full capacity. Thus, we propose an adaptive load balancing mechanism for multiple paths in data centers, dubbed MaxPass, to complement ECMP. A sender adaptively selects and dynamically changes multiple paths depending on the current network status like congestion. To monitor the network status, the corresponding receiver transmits a probe packet periodically to the sender; its loss indicates a traffic congestion. We carry out the quantitative analysis on the ns-2 simulator to show that MaxPass can improve the FCT and the throughput.