Jose M. Arco

Fast Path Ethernet Switching: On-demand, efficient transparent bridges for data center and campus networks

In this paper we propose Fast Path Ethernet, an evolution of the transparent bridges learning mechanisms to increase infrastructure utilization for campus and datacenter networks in a simple way. Fast Path Ethernet Switches reuse standard ARP Request and Reply packets to set up fast on-demand paths between hosts. This architecture uses the standard Ethernet frame format, so it is fully transparent to hosts and compatible with 802.1D bridging in core-island mode. A proof of concept has been implemented in Linux. Preliminary simulations in metropolitan and campus network topologies show superior performance to spanning tree and even to shortest path forwarding, at a fraction of the their complexity.

ARP-Path: ARP-Based, Shortest Path Bridges

This letter is a summary proposal for an evolution of the Ethernet transparent bridge paradigm that provides simple, shortest path bridging in campus networks. ARP-Path Ethernet Switches set up an on-demand path between two hosts just reusing and flooding the standard ARP request frame through all links and confirming the path reaching to the destination host with the ARP reply frame. ARP-Path uses the standard Ethernet frame format, is fully transparent to hosts and does not require spanning tree or link state protocol. Simulation results show superior performance to spanning tree and similar to shortest path routing, with lower complexity. Our implementations confirm backward compatibility, robustness and performance.

Evaluation of tree-based routing ethernet

Tree-based Routing (TRE) revisits Tree-based Routing Architecture for Irregular Networks (TRAIN)-a forwarding scheme based on a spanning tree that was extended to use some shortcut links.We propose its adaptation to Ethernet, using a new type of hierarchical Ethernet addresses and a procedure to assign them to bridges. We show that compared to RSTP, TRE offers improved throughput. The impact of transient loops in TRE is lower compared to the application of the classical shortest path routing protocols to Ethernet. Finally, TRE is self-configuring and its forwarding process is simpler and more efficient than in standard Ethernet and shortest path routing proposals.

HURP/HURBA: Zero-configuration hierarchical Up/Down routing and bridging architecture for Ethernet backbones and campus networks

Ethernet switched networks do not scale appropriately due to limitations inherent to the spanning tree protocol. Ethernet architectures based on routing over a virtual topology in which turns are prohibited offer improved performance over spanning tree, although in some cases suffer from excessive computational complexity. Up/Down routing is a turn prohibition algorithm with low computational complexity. In this paper we propose HURBA, a new layer-two architecture that improves Up/Down routing performance due to an optimization based on the use of hierarchical addressing, while preserving the computational complexity of Up/Down. The resulting architecture requires zero-configuration, uses the same frame format as Ethernet, allows upgrades by software update, and is compatible with 802.1D bridges by means of encapsulation. HURP protocol builds automatically a core with the interconnected HURP routing bridges and the standard bridges get connected to the edges in standard spanning trees. Simulations show that the performance of HURP, evaluated over various combinations of network topology and size, is close to the one of shortest path, is consistently better than that of Up/Down, and is equal or better than Turn Prohibition, with the advantage of having a lower complexity.

All-Path bridging

Display Omitted Today, link-state routing protocols that compute multiple shortest paths predominate in data center and campus networks, where routing is performed either in layer three or in layer two using link-state routing protocols. But current proposals based on link-state routing do not adapt well to real time traffic variations and become very complex when attempting to balance the traffic load. We propose All-Path bridging, an evolution of the classical transparent bridging that forwards frames over shortest paths using the complete network topology, which overcomes the limitations of the spanning tree protocol. All-Path is a new frame routing paradigm based on the simultaneous exploration of all paths of the real network by a broadcast probe frame, instead of computing routes on the network graph. This paper presents All-Path switches and their differences with standard switches and describes ARP-Path protocol in detail, its path recovery mechanisms and compatibility with IEEE 802.1 standard bridges. ARP-Path is the first protocol variant of the All-Path protocol family. ARP-Path reuses the standard ARP Request and Reply packets to explore reactively the network and find the fastest path between two hosts. We compare its performance in terms of latency and load distribution with link-state shortest-path routing bridges, showing that ARP-Path distributes the load more evenly and provides lower latencies. Implementations on different platforms prove the robustness of the protocol. The conclusion is that All-Path bridging offer a simple, resilient and scalable alternative to path computation protocols.

Implementing ARP-path low latency bridges in NetFPGA

The demo is focused on the implementation of ARP-Path (a.k.a. FastPath) bridges, a recently proposed concept for low latency bridges. ARP-Path Bridges rely on the race between broadcast ARP Request packets, to discover the minimum latency path to the destination host. Several implementations (in Omnet++, Linux, OpenFlow, NetFPGA) have shown that ARP-Path exhibits loop-freedom, does not block links, is fully transparent to hosts and neither needs a spanning tree protocol to prevent loops nor a link state protocol to obtain low latency paths. This demo compares our hardware implementation on NetFPGA to bridges running STP, showing that ARP-Path finds lower latency paths than STP.

Carrying ATM cells over Ethernet

This paper presents the design details and performance measurements of an implementation of the Cells In Frame (CIF) protocol over the Linux operative system. The CIF protocol allows to encapsulate ATM cells in Ethernet frames, bringing in this way the advantages of ATM technology over legacy Ethernet networks. Throughput, delay, and delay variation have been measured in order to determine the ability of this approach to suite properly the requirements of multimedia traffic. The obtained results point out that this scheme will provide the necessary ATM quality of service at the price of a minimum throughput reduction respect to native Ethernet. Also, a new queuing algorithm is being researched to optimize traffic management.

TCP-path: Improving load balance by network exploration

Bridging is widely used in Ethernet networks, but the spanning tree protocol blocks all redundant links to prevent loops, and therefore it cannot take advantage of multipath topologies to balance traffic load. Several multipath routing proposals use link-state protocols and Equal Cost Multi-Path routing (ECMP) to distribute the load over multiple paths. However, these proposals are complex and prone to flow collisions that may degrade performance. This paper proposes the new TCP-Path protocol, which selects the fastest available path to a destination host on the fly when a new flow is established. TCP-Path is a new multilayer switching paradigm based on exploring all network paths triggered by TCP-SYN packets, instead of link-state based route computation. Here, we evaluate TCP-Path on ns-3 and Mininet and compare its performance. TCP-Path improves on ECMP by up to 25% in terms of throughput and up to 60% in terms of flow completion time. Thus, TCP-Path offers a better, yet simple alternative to ECMP-based solutions for multipath topologies.

GA3: scalable, distributed address assignment for dynamic data center networks

Deployment and maintenance of current data center networks is costly and prone to errors. In order to avoid manual configuration, many of them require centralized administrators which constitute a clear bottleneck, while distributed approaches do not guarantee sufficient flexibility or robustness. This paper describes and evaluates GA3 (Generalized Automatic Address Assignment), a discovery protocol that assigns multiple unique labels to all the switches in a hierarchical network, without any modification of hosts or the standard Ethernet frames. Labeling is distributed and uses probes. These labels can be leveraged for shortest path routing without tables, as in the case of the Torii protocol, but GA3 also allows other label-based routing protocols (such as PortLand or ALIAS). Additionally, GA3 can detect miswirings in the network. Furthermore, control traffic is only necessary upon network deployment rather than periodically. Simulation results showed a reduced convergence time of less than 2 s and 100 kB/s of bandwidth (to send the GA3 frames) in the worst case for popular data center topologies, which outperforms other similar protocols.