Andre Beliveau

Using Ethernet Commodity Switches to Build a Switch Fabric in Routers

Switch fabric in routers requires very tight characteristics in term of packet loss, fairness in bandwidth allocation, no head-of-line blocking and low latency. Such attributes are traditionally resolved using specialized and expensive switch devices. Motivated by the emergence of IEEE Data Center Bridging, we explore the possibility of using commodity Ethernet switches to achieve scalable, flexible, and more cost efficient solutions, while still guaranteeing the switch characteristics. In this context, we propose Ethernet Congestion Control & Prevention (ECCP), a novel concept to control and prevent congestion in switch fabrics. ECCP consists of (1) a method to estimate the available bandwidth along a given network path using a train of probes and (2) a rate control algorithm to adjust the sending rate of traffic along this path based on the estimated bandwidth. To prove ECCP and evaluate its characteristics, we present a first prototype based on the OMNEST simulator and conduct extensive experiments. Our analysis confirms that ECCP is a viable solution to (1) avoid congestion within the fabric, thus minimizing path latency and avoiding packet loss, (2) guarantee fair share of the link capacity between flows, and (3) avoid head of line blocking.

Optimal packet classification applicable tothe OpenFlow context

Packet Classification remains a hot research topic, as it is a fundamental function in telecommunication networks, which are now facing new challenges. Due to the emergence of new standards such as OpenFlow, packet classification algorithms have to be reconsidered to support effectively classification over more than 5 fields. In this paper, we analyze the performance offered by EffiCuts in the context of OpenFlow. We extended the EffiCuts algorithm according to OpenFlows context by proposing three improvements: optimization of the leaf data set size, enhancements to the heuristic used to compute the number of cuts, and utilization of an adaptive grouping factor. These extensions provide gains in many contexts but they were tailored for the OpenFlow context. When used in this context, it is shown using suitable benchmarks that they allow reducing the number of memory accesses by a factor of 2 on average, while decreasing the size of the data structure by about 35%.

Toward a semantic-based packet forwarding model for Openflow

Openflow is a step towards a more flexible data plane. In Openflow-based deployments, packet forwarding model relies on a set of pipelined tables. However, supporting multiple hardware tables is challenging while having serious limitations. In this paper, we investigate a novel approach to packet forwarding design that tailors Openflow forwarding paths according to the underlying hardware. Its main goal is to provide an optimal mapping of entries to physical search structures. To that end, conceptual abstractions are used in devising a semantically rich and high-performance packet forwarding.

Proactive Ethernet Congestion Control based on link utilization estimation

In the quest of providing a deterministic backplane-like behaviour a mechanism called ECCP (Ethernet Congestion Control & Prevention) was proposed where specialized and expensive devices were replaced by commodity Ethernet switches while still keeping the router properties, e.g. low-latency packet delivery, no packet loss within the fabric and fair share of the bandwidth. However, ECCP uses a self-induced congestion probing model which can cause queue length fluctuation in addition to the network bandwidth wastage due to the probes. Therefore, we propose enhancements to the algorithm used by ECCP to reduce probe packet overhead. In our solution NoP-ECCP (ECCP with no probes), we use a link utilization estimation technique instead of the estimation of the available bandwidth. The results obtained through simulations show that NoP-ECCP outperforms ECCP in terms of fairness, link utilization and queue length.

Extensions to decision-tree based packet classification algorithms to address new classification paradigms

Abstract The decision-tree based packet-classification algorithm field has seen many contributions since the first algorithm using a geometrical rule representation, HiCuts, has been proposed. While hardware reported implementations for this class of algorithms have proven that a high throughput can be reached, those algorithms are inherently facing a tradeoff between speed and memory consumption. This paper presents two extensions applicable to decision-tree based algorithms designed to tackle two of their common drawbacks. Applied together, they achieve a reduction of the number of memory accesses, while reducing the data structure size. The first contribution consists of a new rule-clustering method aimed for the reduction of the number of trees built. The second contribution relies on a leaf compression method that allows tackling the problem that stems from linear leaf traversal. Applied together, as shown by simulations, those two new methods improve the trade-off between search-time complexity and data structure size. These strategies provide gains in many contexts, although they are tailored for handling complex rule sets used in the context of Software Defined Networking. For sets of 100,000 and 10,000 rules, those two strategies reduce the number of memory accesses by a factor of 3 on average, while decreasing the size of the data structure by about 45% over EffiCuts, a well-known decision-tree based algorithm.