In recent years, computing and communications systems have faced new design challenges based on their thermal, power and environmental requirements. Coupled with the increasing complexity and internal inefficiency of today’s designs and the inability of current architectures to keep up, these issues are now having a major impact on how engineers design and create new systems.

Put simply: the limited ability of today’s parallel systems to communicate internally – among their processors, memory and peripherals – has now begun to severely limit the performance of these systems.

A new generation of companies, such as Lightfleet, has been hard at work on developing solutions to these issues, including new technologies and innovative approaches that are dramatically advancing the capabilities and performance of next-generation computing and communications systems.

Pain points and trends

New computing applications and more than 1 billion computers connected to the worldwide Internet are creating an ever-growing flood of information that has grown from megabytes to gigabytes, gigabytes to terabytes, and now terabytes to exabytes – with no end in sight.

Despite continuing technical advancements, recent hardware solutions haven’t increased data center reliability. Rather, their increased complexity has required even higher levels of cost and complexity to keep them running! A poll of 100 IT managers conducted in 2006 by Morse Management Consulting, a business and IT consulting company, showed that fewer than 10% of large enterprises plan to adopt advanced computing architectures such as grid computing because of strong concerns about cost, complexity, and security.

Yesterday’s minor data center bottlenecks are rapidly turning into today’s major roadblocks. IT professionals often spend more time meeting data centers plumbing requirements than working on the solution for the original computing requirements themselves. Recent data from Symantec Corp. confirms this conclusion, reporting that as much as 70% of IT budgets are spent maintaining existing environments.

The oceanic research department at a major west coast university is typical of organizations facing an “exaflood” of data. The university is finding that its latest analytical projects are becoming choked with an unprecedented and growing amount of real-time data. In one application, the department’s video sensors produce a 4K by 4K (16MB per frame) data set. The number of input devices and data-sampling rate required to analyze and manage this data is beyond the capacity of current home-grown or commercially available systems.

One approach that some organizations have adopted to manage their data center workloads is to sub-divide applications by distributing them via high-speed networks to remotely located “embedded” systems. Examples of these embedded systems include high-end medical diagnostic machines; real-time geo-science data-acquisition systems; and field-based identity recognition for security systems.

The performance demands of these embedded solutions, which are often located outside of traditional data centers, are also experiencing data overload. These embedded applications present a particularly challenging situation, because they are often required to provide data center class computational performance in a compact physical size and battery-friendly design.

Traditional communication and performance approaches

For many organizations, grouping a number of compute modules into a workgroup or cluster has proven to be a very effective way to handle large computing loads. The simplest way to connect compute modules, or nodes, together is to use a direct, non-shared connection.

Enterprises can simplify the complexity of a computing or communications system with direct connections by using a communications switch, which reduces the number of physical cables and ports. Switches also provide built-in fault diagnosis and isolation – a benefit over directly wired nodes, which cannot isolate faults at the node itself.

More complex switch and cable arrangements, such as the current high-performance InfiniBand and Fibre Channel system interconnects, are known as “switched fabrics,” connoting the weaving of virtual cables from node to node. There are dozens of these open and proprietary interconnects in the industry today being used for general and specific applications; however, no single fabric interconnect has proven to be a universal solution.

Many data centers expect uninterrupted service availability. Because physical cabling can be such a common point of failure, clustered solution vendors have attempted to minimize risk by increasing the physical redundancy of interconnections and switches in case a single communication line breaks. The inherent unreliability of interconnect cabling has made them unattractive – but still necessary – for many mission-critical applications.

Many vendors have taken the traditional approach of adding nodes and/or processors to a system to increase its overall compute performance. However, this approach doesn’t generally deliver a linear performance improvement, because of the extra work of node synchronization, intra-system data communications, parallel environment creation, and cancellation.

In the current era of multi-threaded, multi-core processors, software programmers are becoming more adept at taking advantage of these hardware advances by parallelizing their workloads and implementing efficient, dynamic load-balancing techniques.

Despite these advances, however, node interconnect bottlenecks have become the most significant impediment to parallel computing performance gains. Put simply: the limited ability of today’s parallel systems to communicate internally – among their processors, memory and peripherals – has now begun to severely limit the performance of these systems.

Providing the solution with light

Up to this point, no vendor has been able to create a computing or communications system that maintains constant communications between its processors, without incurring interference, time delays, and congestion.

Lightfleet has created an interconnect solution that solves these problems, and unlocks the true potential of parallelism by removing the intra-system communication limitations that have existed up to now.

The company’s Corowave™ interconnect uses multi-channel, broadcast light – eliminating physical wires and optical fibers – to transmit data. This breakthrough enables simultaneous and continuous data transmission between a system’s nodes – and is scalable to dozens, hundreds or even thousands of nodes.

The Lightfleet Corowave interconnect is the world’s first continuous, all-to-all, broadcast optical interconnect. 

The Corowave optical interconnect utilizes laser transmitters and optoelectric receivers on a single plane. Each communicating node would have a corresponding transmitter and receiver.  The transmitters and optoelectric receivers are in a compact array matching the total number of communicating nodes supported. The use of a mirror enables reflection of each light-based data transmission back to all receivers.

Since all transmitters are always seen by all receivers, natural fault detection is easily achieved. If an originating node can’t receive its own transmitted data, it will immediately recognize there is a problem. If all the other receiving nodes also don’t see the transmission, it’s easy to isolate the interconnect problem to the originating node’s transmitter. If all the other receivers see the transmission, then the receiver on the originating node is in fault.

Corowave technology is applicable to a broad range of computing and communications applications. Although the technology will be initially integrated into enterprise-class commercial servers, it is applicable to a broad range of computing and communications equipment. Specifically, Corowave technology enables faster, less-expensive and more compact systems to address multiple commercial markets.