Load balance is an important issue in datacenter networks. The flowlet-based algorithms can balance the traffic with fine granularity and does not suffer the packet mis-sequencing problem. But their performances are rather limited or require extra communication overhead. In this paper, we propose a local load-aware algorithm called Dynamic Roulette Wheel (DRW). In DRW, the roulette wheel is adopted to select a new path for the flowlet according to the local load. Each source of multipath balances the traffic to all its egress links without the communication overhead. Moreover, the granularity of flowlet can be dynamically tuned from a single packet to the whole flow. Finally, the Capacity Aggregation (CA) mechanism is designed for the case of link or switch failure. We prove in theory that DRW can achieve the optimal global load balancing. The simulation results also show that DRW provides almost the best delay performance and the least packet out-of-order proportion overall among all existing flowlet switching algorithms.

Multipath networks are widely used in data centers and load balancing is one of the most important technologies to improve their performances. Due to the ever-increasing in data center network size, existing load balancing algorithms face the challenges of efficiency and scalability. In this paper, we propose a new load balancing algorithm for large-scale, multi-tier fat-tree based data center networks. Different from the conventional architectures, a multi-tier fat-tree is divided into multiple routing domains according to the topology, and the routing processes in different domains are independent. The devices outside a routing domain can only access the specific interfaces provided by this domain. It is thus very convenient for deployment and modular upgrade. We also design a distributed and data-driven feedback mechanism, with which the routing decision is based on the global load information. We prove that the new algorithm can achieve perfect load balancing in multipath networks and show that the new algorithm outperforms all other load balancing algorithms in performance.

The emerging cloud applications require group communications. For these applications, multicast is a better choice than unicast, because it can significantly improve the performance by eliminating the duplicated packets generated by servers. However, existing multicast schemes for datacenters are either based on IP multicast or centralized scheduling. IP multicast is inefficient for datacenters as it cannot take full advantage of the multipath property. And centralized schemes suffer from single-point failure and scalability problems. To solve these problems, we propose MiniForest, a distributed multicast framework for large-scale datacenter networks. It consists of new routing algorithms and a dynamic group management mechanism. A new address mapping solution is then designed for compatibility to existing upper-layer applications. Based on the mapping solution, we propose an efficient load balancing strategy, with which a minimal forest is constructed for all multicast trees. To study the performance of the new multicast scheme in theory, we further provide an analytical model for Clos-based datacenter networks and analyze the overloading behaviors from a new perspective. We show that the distributed scheme can be used in any size of datacenters. It has much lower complexity and better performance than centralized schemes.

Multicast becomes essential in data center networks, since more and more applications require group communication. The existing multicast scheduling algorithms in data center are Internet-based or centralized, which are either not efficient or not scalable. In this paper, we propose a Distributed and Dynamic Multicast (DDM) solution for fat-tree data center networks. It includes multicast initialization, routing algorithm and load-balancing policy. As DDM does not need the central controller, it is more scalable and much simpler. Moreover, each host can dynamically join in or quit from an existing multicast group without suspending the live traffic in this group. Our simulation results show that DDM provides the better delay/throughput performance than the existing centralized multicast scheduling algorithm.