Features:

INTERVIEW WITH STEVE WALLACH, CHIARO
By Alan Beck, Editor-in-Chief, HPCwire

HPC: Exactly what is supernetworking, and why is it important?

STEVE WALLACH: Supernetworking is an infrastructure that has the following attributes:

• The bandwidth of the network is the same whether the network elements are 100 meters away, or 10,000 kilometers away -- a difference in distance of five orders of magnitude. Bandwidth is the same, only the latency is different.
• There is a convergence of computer and communications that translates into the convergence of LAN/WAN/SAN. New protocols and standards, like iSCSI, IPv6 and XCP, will greatly facilitate this convergence.
• High-speed and cost-effective optical transport is the underlying technology that is facilitating this movement.

Supernetworking is important because networking is worldwide. We must not distinguish nor discriminate the ability to do science due to the location of the researcher or the location of the research apparatus. Ubiquity and satisfying our nomad needs are required. We live in a world that makes timeliness a high priority.

HPC: How is the OptIPuter relevant to the evolution of supernetworking? Are OptIPuter technologies sufficient to realize the supernetworking vision?

SW: The OptIPuter is a "virtual" supercomputer in which individual compute elements are widely distributed clusters, where the backbone network is provided by IP packets transported over single or multiple dedicated lambda's (DWDM). Each lambda (or wavelength) is 1 to 10 Gbps. Mass storage is large (scaling to Petabytes) distributed scientific data repositories, fed by scientific instruments as near real-time devices. Another OptIPuter goal is to make interactive visualization of remote multi-gigabyte data objects as easy in the future as the web makes manipulation of megabyte data objects today.

The OptIPuter and the TeraGrid represent the two extremes of supernetworking. The TeraGrid's WAN bandwidth is currently 40 Gbps, connecting Linux clusters containing 1000 to 2000 nodes, each with a GigE I/O. Thus TeraGrid is optimized for computing, with a bisection bandwidth of a few percent. The OptIPuter will connect smaller clusters, but will be optimized for WAN bandwidth, approaching a bisection bandwidth of 100 percent. Thus, the OptIPuter WAN bandwidth can be equal to or greater than end-user I/O bandwidth. In practice, much of the time the 10 Gbps light pipes between TeraGrid sites will be used to couple smaller "virtual" clusters of the entire TeraGrid. In this case the WAN to I/O ratio would be much higher. One can view TeraGrid as a hosting environment for "Virtual OptIPuters."

Systems like the OptIPuter or TeraGrid and supernetworking are technologies that are symbiotic. OptIPuter systems need very high-speed networks and connectivity to fulfill their mission. Supernetworking needs new services and technologies to both achieve economies of scale and provide the latest scientific and research tools.

A question frequently posed is, "what is the next killer app?" Conventional wisdom would like the identification of one or two very specific applications. However, if one examines what has really happened, it is more of a "build it and they will come" technology cycle. The OptIPuter and Supernetworking are contemporary examples of this. We are in the "build it" phase.

HPC: Please describe your own work with supernetworking and where you intend to take this concept in the future.

SW: Chiaro has developed the Enstara, the first and only router that uses nano-second optical switching, programmable network processors and is scalable to in excess of 6 Terabit/sec of bandwidth. It has been chosen as the basis for the OptIPuter system.

My view is that a router to support the supernetworking tenets must be a general-purpose packet processor. As new requirements are placed on routers, upgrades must occur (whenever possible) in software. As the co-CTO of Chiaro, I made sure that the Enstara system was designed from the outset to meet the needs of the supernetworking world.

Going forward I see several thrusts. One thrust is to provide cost-effective interfaces for 40 Gbps interfaces. It is unclear when this will be required. The availability of cost-effective, 10 Gbps lambdas and DWDM has pushed out the development of OC768c (40 Gbps). Another direction is to increase the router bandwidth to Petabits or more. While I spoke about the convergence of LAN/WAN/SAN, I also can see the day when the internal switches used to create clusters are converged in the LAN/WAN/SAN network. That would mean one general-purpose internal network that optically interconnects SMP's, attached storage, visualization systems and ports to the global grid.

HPC: What are the principal challenges to efficient supernetworking, and how do you foresee these being overcome?

SW: There are many challenges. We must first recognize the economics of the network. While much of the technology is available, costs must be reduced by an order of magnitude for widespread deployment. Much like we witnessed "Moore's Law" making supercomputing available to the masses (including a cluster of Apple's and 64-bit address space processors at the consumer desktop), we will see a similar reduction in the cost of the basic supernetworking building blocks.

The next major challenge will be an efficient and stable software environment. To fully utilize supernetworking, we must bee able to support and implement the TeraGrid objective of: develop at A, execute at B and store at C. This means a common middleware or software stack and the strict adherence to the software interfaces established. Having Linux as a base-level operating system, and open source, will make it easier to achieve these objectives. But by no means is this a slam-dunk.

Fundamentally, I believe that efficient supernetworking won't happen simply by tweaking existing technologies and hoping they do the trick. As with most revolutions in technology, it's going to require thinking about the challenges from new and innovative perspectives. Our work at Chiaro is one example of this: creating a next-generation routing platform by combining expertise in the fields of networking, supercomputing, telecommunications and photonics. This kind of synergistic approach will be an absolute necessity for supernetworking, TeraGrid and whatever succeeds them in the future. One thing's for sure -- it'll be an exciting ride.