If there is a technology moving as quickly as SDN, it has to be photonics. The emergence and growth of affordable optical networking is changing how we think about network architectures, making possible a new class of low-latency, high-bandwidth transport within and between data centers. SDN, coupled with Layer 1 optical circuit switching technology allows the physical network topology to be reconfigured on demand to support changing traffic patterns and application demands. One particular use case applies to the handling of highly persistent traffic flows.
Modern warehouse-scale data centers are essentially huge computing resource pools where workloads are shuffled between server clusters to maximize efficiency. This concept also extends to interconnected cloud facilities where Virtual Machines migrate across data centers based on the utilization levels. The net result of these movements of virtual machines is large highly-persistent east-west traffic flows, popularly know as Elephants. Imagine moving a VM instance and associated storage consisting of tens or hundreds of gigabytes of data and you have a classic elephant flow.
VM movements however are not the only source of elephants. Other generators include movements of large data sets in Hadoop frameworks, and data replications and backups. Typically, the majority of traffic flows within a data center are the opposite of elephants – short-lived, bursty, latency-sensitive “mouse” flows, but the majority of the traffic is contained within a smaller number of elephants. Both flow types are extremely important and need to be handled efficiently.
This presents a problem in homogeneous data center networks because highly persistent TCP flows fill network buffers, and this adds queuing delay to anything that shares these buffers. The result is that the mouse flows that are more sensitive to network latency are negatively impacted. At the same time, the network capacity limit in the data center may delay the delivery of the elephant flows resulting in multiple minutes or tens of minutes to move a VM or data set. Attempting to accommodate elephant and mouse flows within a common network potentially degrades the performance of both flow types.
There have been several proposed solutions to allow elephants and mice to coexist harmoniously. One very promising approach is to move elephant flows onto a different physical network using policy-based routing. Elephants can be detected in real-time via flow detection algorithms monitoring flow data from Top of Rack switches or virtual switches and rerouted onto an alternate uplink connected to an offload network. Another question arises however – what sort of offload network should the elephants be routed to?
Optical circuit switches that operate in a pure photonic domain using 3D MEMS micro-mirrors offer interesting advantages in the handling of offloaded elephant flows.
Imagine allocating dedicated, high-bandwidth, single-hop paths to elephant flows when and where needed. No multiple network hops, no queuing constraints – just pure TOR to TOR optical bandwidth whose only constraint is the speed of the TOR transceiver. This means VM migrations, data replications, and Big Data jobs at light speed while freeing up the L2/3 network to deliver mouse flows efficiently and with minimal latency.
At 40G or above, optical circuit switches are considerably less expensive than their L2/3 equivalents and consume orders of magnitude less power – a 320 port OCS typically consumes less than 50 watts compared with more than 5KW for a similarly configured L2/3 switch running 40G Ports. The switches also scale to 100G, 400G and beyond as the network grows so the only future upgrades required are faster TOR transceivers.
Is there a downside to routing the elephant flows through an optical circuit switch? Until SDN, the answer was definitely yes, the challenge being that a circuit switch switches – well, circuits. Since there are not an unlimited number of circuit paths available between large numbers of TORS, this “hybrid” packet-optical network needs a controller to monitor the traffic matrix and make intelligent policy-based decisions regarding when to open circuit paths between specific TORS and move flows from the L2/3 network to the OCS and vice versa. In a nutshell, the light speed circuit paths must be allocated where they are most needed at any given time.
The separation of control and data planes ala SDN together with new controller frameworks and flow-monitoring analytics is allowing this intelligent allocation of physical layer capacity to be realized. Flow statistics at the TOR or virtual switch are collected and made available to a network controller that has traffic engineering and topology configuration capabilities. When an elephant flow is detected the controller sets up a path through the optical circuit switch and subsequently moves the elephant flow at the TOR to an alternate uplink port connected to the OCS.
Elephant flows can thus be quickly detected and offloaded to pure optical express routes where they are delivered to their destinations at near light speed. This leaves the L2/3 aggregation network essentially free to efficiently deliver mouse flows with minimal latency. The mice and elephants coexist happily on the hybrid network.