Cost-Effective Congestion Management for Interconnection Networks Using Distributed Deterministic Routing

Abstract

The Interconnection networks are essential elements in current computing systems. For this reason, achieving the best network performance, even in congestion situations, has been a primary goal in recent years. In that sense, there exist several techniques focused on eliminating the main negative effect of congestion: the Head of Line (HOL) blocking. One of the most successful HOL blocking elimination techniques is RECN, which can be applied in source routing networks. FBICM follows the same approach as RECN, but it has been developed for distributed deterministic routing networks. Although FBICM effectively eliminates HOL blocking, it requires too much resources to be implemented. In this paper we present a new FBICM version, based on a new organization of switch memory resources, that significantly reduces the required silicon area, complexity and cost. Moreover, we present new results about FBICM, in network topologies not yet analyzed. From the experiment results we can conclude that a far less complex and feasible FBICM implementation can be achieved by using the proposed improvements, while not losing efficiency.

An Effective and Feasible Congestion Management Technique for High-Performance MINs with Tag-Based Distributed Routing

“…Thus, the CAM size required at each port grows with the number of destinations in the network (i.e., with network size), thereby this strategy not being scalable (although its effectiveness is satisfactory). A second version of FBICM [18] reduced partially the CAM size required at each port by limiting the number of destinations per list, but actual scalability in medium-size or large networks was not achieved.…”

“…These techniques use either CAMs or T-CAMs, but the organization and requirements of the CAM/T-CAM fields that are not compared with the lookup word (i.e., those implemented with SRAM, see Fig. 2) are identical, requiring around 5 bytes per line [18]. Hence, we base our comparison on the number of cells required to implement the CAM/T-CAM fields involved in the searches.…”

“…Note in this case the lookup word to address the T-CAM (see Section 3.1.1) is built by filling its "address" field with the packet destination, the remaining fields of the lookup word being filled with X values. It is worth pointing out that all the destinations of packets contributing to a specific congestion root are controlled by a single mask, thereby DRBCM reducing the required amount of control memory with respect to previous approaches that use CAMs [17], [18].…”

“…Flow-Based Implicit Congestion Management. We consider the second, more efficient version [18] [7]. The RAM at each input port is divided into as many queues as output ports in the switch (i.e., eight in our experiments), so that each packet is mapped to the queue assigned to its next output port.…”

“…For the two FBICM versions, the number of cells per entry has been computed assuming five cells for the OP field to refer up to 32 switch ports, three cells for the Hops field to refer distances up to eight network stages, and two cells for the NextLine field to refer up to four entries [18]. Besides, six cells are required to store each destination for 64 Â 64 networks, eight cells per destination for 256 Â 256 networks, and 10 cells per destination for 1;024 Â 1;024 networks.…”