Industry Summary

Market Opportunity

The Internet is a nexus for communications convergence, in time becoming the sole communications technology used in networks as small as those between chips on a printed circuit board, up to networks of global size. From consumer products to phone to PC to storage devices to supercomputers, all will use the same mechanism. This will create enormous growth in Internet service demand, as well as vast new product sales.

Realistically, however, communications convergence using the Internet is elusive to achive - it’s very evolution impeded by the residues of past technologies devised in the 1970’s during it’s "ad hoc" development period. As volume customer demands for universal and reliable service that can provide voice/data, multimedia, and electronic commerce increase, cracks in the current infrastructure have arisen, manifesting themselves in bandwidth "bottlenecks" obvious to everyone. Delays, unpredictability, and loss of efficiency stemming from these untimely bandwidth bottlenecks leave the end customer feeling helpless and frustrated. Radical solutions through use of IP Multicast, RSVP, and other related pure technical approaches by their nature have not remedied this problem because they run counter to customer acceptance - they don’t improve web, e-mail, and file transfers capability, which are the staples that the customer uses today.

InterProphet’s Solution

Our mission is to remove these intractable "bottlenecks" to convergence, regardless of application. Rather than contort many proprietary bus/networking technologies that could cope with the vast demand to work with the Internet, we chose a different path. We chose instead to creatively refine the fundamental, interoperable technology of the Internet itself in a way to both 1)-scale bandwidth and 2) can be applied to the broad range of products that will be a part of the Internet.

In effect, we are reversing the process - instead of the traditional approach of adapting data technology to the Internet, we use the Internet itself to derive the data technology necessary to extend the Internet most effectively.

This technology is a refinement and replacement of the "Internet stack" found in every Internet system. It embraces and extends the vast existing Internet infrastructure, where it acts as a volume driver for existing and new Internet products ranging from servers to infrastructure to embedded applications. No significant economic or technology advantage now stands in the way of universal communications convergence via the Internet.

InterProphet’s Transport Protocol Technology Advantage

InterProphet has determined that the best course of action for resolving these issues is not to replace TCP as dominant transport layer protocol of the Internet, but to instead complete its function through the use of novel hardware and software techniques. This is an evolutionary view, which follows many earlier accomplishments. Earlier NICs used hardware boosts (e.g. Parallel Tasking II, 3Com) to increase NIC performance at the link layer. Software interface specifications (e.g. NDIS, 3Com) opened up industry opportunities to deliver more value-added features to the customer, raising the value proposition of network interconnects. Reduction of context switches, system calls, and interrupts in the operating system through software bypass mechanisms (e.g. VIA, Intel) allow a server to focus less on the operational issues of the operating system and more on the performance of applications.

Silicon TCPTM technology is a combination of hardware and software mechanisms which result in the increased bandwidth and decreased latency of TCP. It comprises a unique invention, with numerous patents (pending) and trade secrets that provide continuous flow switching technology complimentary to Internet standards that scales protocol bandwidth with signalling rate. It can provide a benefit in any product that passes an Internet packet. It makes possible scalable products that can handle any-sized bandwidth demands.

Silicon TCP provides a conservative path to reduction of Internet performance bottlenecks, and an upgrade path to increase existing product transmission capacity. Silicon TCP amplifies Internet serving by increasing per server processor bandwidth capacity by orders of magnitude. Internet processor clusters can be formed that can handle tens of thousands of transactions per second between systems in metropolitan-sized areas. Network-connected storage can cost effectively switch information via the Internet at the "wire speed" of networks. An enterprise application program can now make use of orders of magnitude greater bandwidth than before, making gigabit Internet applications programs a practical possibility.

Current Market Solutions and Trends

Other technical solutions to these bottlenecks have been broached – front-end processor (FEP) protocol engines, and software-only approaches – and have reappeared in many guises over the last decade, but have met with only limited success. FEP solutions are vexed by the contradictions inherent in the design of the microprocessors on which they depend yet is deeply adversarial to in their aims. This has resulted in unsatisfactory results, and a road strewn with failures (e.g. Protocol Engines Incorporated). Software approaches have had more success at refining the current protocol and systems processing bottlenecks themselves, but now run up against fundamental design issues which cannot be resolved short of obviating the current infrastructure with an entirely new architecture.

From a customer’s point-of-view, the best way to reduce these bottlenecks and personal stress levels has always been to purchase newer and faster computers, modems, and the like. The computer industry, in fact, has benefited significantly by feeding customers with these "better, faster, cheaper" solutions. While this approach has been successful in the controlled environment of the desktop, the Internet itself presents unique challenges.

Software and hardware approaches which rely on the processor speed advantages to keep up with traffic, for example, negate this advantage because the processor has to synchronize itself with the protocol stream, causing it to fall too far behind to make up the deficit. In addition, the dominant trend of modular operating systems architecture by the leading supplier, Microsoft Corporation, implies processing delays, which cannot be minimized through software techniques. Thus, hybrid techniques such as hardware buffer management used to redirect data before delivery still rely on the operating system software to determine the destination of delivery. In sum, faster processors cannot scale faster protocol processing.

Thus, bottlenecks continue to exist, despite the wishes of the customer and the assurances of the salesman that they have been resolved. Is it no surprise then that sales of high-end computers and peripherals have slowed, while low-end sales increase? How can customers rely on "better, faster, cheaper" solutions when they fail to deliver reliably in the obvious case -- the Internet?

Communications convergence is no longer a postulate of a few "visionaries". It is now feeding the demands of customers and driving the business strategies of the entire computer and telecommunications industries. The continued instabilities exacerbated by bandwidth bottlenecks will magnify as usage and expectations increase, until a tenable economic resolution palatable to all lies at hand.

Overview of Competitive Landscape

The Internet has redefined the competitive landscape in recent years, which has resulted in the launch of a new series of strategic initiatives and products by those companies, which have decided to compete in its sphere. Development of an Internet strategy and business model, which leverages its unique aspects, has thus become a primary focus of both start-ups and major established players. As such, the strategic partnerships and alignment must be incorporated into the business and technology directions of a smaller player from the company’s inception.

An understanding of the mutable competitive landscape and how to tread lightly as an Internet company is the pivotal strategic element of any company entering the fray.

Limits to Current Solutions

These interests represent current solutions that may be challenged by InterProphet’s technologies. The five competitive groups include:

Faster processors, Newer Processors, Multiprocessors

The common answer is that faster processors are the answer – add more Megahertz and the problem will vanish. However, while TCP bandwidth has improved, it’s an ever-losing battle, as one always needs more Megahertz than a processor can provide. At some point, the processor must be able to run applications as well as the network – perhaps even a database as well. The most successful NIC products today take this into account, attempting to reduce the processor demand as much as they can. But they’ve also reached the limits as far as that goes – you still need too many interrupts, processor cycles, system calls and context switches.

So let’s turn to newer (e.g. VLIW) processors than can accomplish more per Megahertz. Unfortunately, these are currently even less successful at increasing network bandwidth. Partly it may be the lack of experience with the newer architecture, but it may also be that processors of this kind fare less well with interrupts, system calls, context switches, and unpredictable code instruction paths that are common in network protocol processing.

Well, if we need more cycles, why not just string together more processors? Microsoft tried this with NT, and while NT can handle 2-8 processors in symmetric shared memory multiprocessing, the costs outweigh the benefits fairly rapidly, as each of the processors compete for common resources. Worse, as processors go faster, sharing becomes more troublesome. Processor clustering (shared nothing) can minimize sharing costs, but it requires low level networking in order to scale networking processing itself – which gets complicated and costly rather fast. Many believe that networks are the level where clusters should be integrated with to begin with.

Protocol Engines

Protocol engines are yet another variation on multiprocessor architectures. The fallacy is assuming one can always offload any type of processing of the main processor to another processor. The degree to which this can be successful is the degree to which one can optimize a processor to reduce the steps to process packets. Unfortunately, the protocol engine needs to amortize the communications cost (interrupts and context switches) with the host processor, as well as compete with the host processor vendors for performance. Large transmission, or "fat" packet handoffs are about the best strategy to minimize costs, but still the latency of getting a packet out the door is always higher. Each bit of latency postpones another session the server can handle – so you can get some improvement in bandwidth, but at the cost of fewer clients that a server can handle.

OS improvements

Well, if we can’t toss more conventional silicon at the problem, perhaps we can mitigate the cost. Various improvements to network stacks and interface libraries in the operating system can squeeze 10%-20% improvements out of specific portions of the operating system, but usually this comes at a cost of lost features, or deferring the costs to hardware. With the Internet’s TCP protocol, the largest costs of processing are the copying of data, checksum calculations, and message handling (Intel recommends that servers have NIC’s that support Microsoft’s extensions to do just these functions). The difficulty in exporting these functions to hardware is that it’s equally difficult for hardware to do them as well, so unless one can arrange them differently than sequential software implementations allow, one doesn’t get much advantage for the additional hardware cost. The root of the evil is that the software presumes successive processing after the packet arrives – and then it’s too late to recover.

Interconnects

Ok, we’ve got limits with the processor and the operating system. Why don’t we just spit the information out over some special purpose hardware, which doesn’t need to obey the rules that frustrate our efforts in the first place? Special purpose interconnect hardware does exactly this quite well. Very large data rates can be had, with low processor overhead as well. But we have a problem – the information spit between the systems isn’t compatible with the rest of the computers participating in the Internet, because it doesn’t obey the rules necessary to interoperate. Never fear, we’ll just add some software in the application when necessary to make it compatible enough to have limited interoperability where we need it. But eventually, when we add all of the parts back in for complete industry qualification, we’re back where we started, with the same performance limitations and hardware cost.

Server NICs

Server NICs attempt to make the best of a bad situation by optimizing all of the details of network protocol processing architectures with a hodge podge of the above mentioned techniques. Interrupts are costly; so many ports are shared on a single board to consolidate many context switches off of one interrupt for a statistical gain. Interrupts are also postponed on the odd chance that they will not be required. A microprocessor may be added to offload TCP message segmentation and checksum calculation, so long as the microprocessor outruns the host processor. A bus master DMA interface eliminates the second or third data copy operation at an increased setup cost, so long as a small packet (like an acknowledgement, the second most popular packet on the Internet) isn’t being sent. PCI bus management is optimized to reduce the total bus transfer time of a packet by predicting how long to seize the bus for without fouling disk controller transfers. The best description of this strategy is the "I hope it works sometimes" approach.

General Operating Systems: Microsoft

The primary software player in the Internet space and greatest competitor to all is Microsoft Corporation. Currently with the final release of NT 4.0, Microsoft does not offer any solution to cope with the demand for general scalable low-latency applications vending – a failing that Sun Microsystems has been able to exploit with its scalable UNIX systems. Utilizing its Wolfpack clustering software, Microsoft intends to introduce scalable applications vending as part of NT 5.0 in 2000, with debuts of the software in beta form available to sites projected by mid-1999, but low-latency issues are not resolved within the operating system itself. As a consequence, Microsoft is unlikely to achieve a clear dominance over Sun Microsystems in this arena over the next two years. While high-end serving is not the primary focus of Microsoft, this is of significant concern to platform vendors who are competing in the enterprise space with Sun Microsystems.

While Silicon TCP can be used with any operating system (or even as a stand-alone component in other infrastructure), the market dominance of Microsoft is very compelling. Using InterProphet’s Silicon TCP technology, Microsoft could demonstrate an NT "Hot Box" server offering significant performance and scalability over an existing NT system. This would be very interesting to the growing electronic commerce and portal companies. Since InterProphet is a small company, however, the best approach to dealing with such a dominant player might be to structure a marketing partnership for NT 4.0, with a strategic partnership developed for NT 5.0 and Windows 2000 as the technology becomes accepted in general use.

Server Platforms

The competition to hold the high-end server platform is currently dominated by three players: Compaq, Sun, and IBM.

Compaq is an X86 based platform vendor which sells high-end servers based on Microsoft’s NT offering. As cited above, the inherent limits to NT as supplied by Microsoft (e.g. modular design of NT coupled with scalability issues) makes it difficult to meet the demands of a competitive enterprise server environment. Special hardware (e.g. interconnects, server NICs) and software packages (e.g. load balancing, protocol acceleration) mitigate but do not completely obviate these problems, and in the case of software solutions often simply move the bottleneck around. Finally, non-standard proprietary hardware imposes additional costs (unproven emerging technologies, custom maintanance, migration) which precludes its rapid acceptance outside of rarified channels. (N.B. Compaq’s recent acquisition of the alpha processor line from DEC may provide it additional opportunities to expand in the enterprise arena in the future, but its true impact has yet to be felt.) Only a server which compensates for NT’s shortcomings in a standards-based manner acceptable to Microsoft would be able to compete with leader Sun Microsystems. An InterProphet partnership with such a platform vendor could provide Compaq with the additional qualities required comparing favorably with Sun.

Sun is a Sparc-based platform vendor, which sells high-end servers based on its custom UNIX operating system software, Solaris. The combination of it’s own internally developed hardware and software have allowed it to achieve maximal performance in terms of high bandwidth, low-latency, and scalability. Sun’s lithe ability to modify its own software and hardware to take advantage of new demands, coupled with its long involvement in networking and the Internet make it a formidable competitor to other platform vendors.

IBM, in contrast to both, offers a multitude of platforms based on UNIX (AIX), NT, and proprietary technologies (e.g. AS/400, mainframes). It is focussing on a unification of these disparate platforms using a Java-based environment. The complexity of IBM’s product roadmap, in addition to its continued dedication to the mainframe and AS/400 lines, present interesting opportunities for partnerships, but also could impose additional costs for maintanance of this relationship.

Server Interconnects and Server NICs

The high-bandwidth server market consist of many types of products ranging from server NICs (e.g. 3-Com, Intel, SMC,) to protocol engines (e.g. Packet Engines, Xtaqi,) to server interconnects (e.g. GigaNet, ServerNet,)

Both protocol engines and server interconnects attempt to achieve, through specialized hardware solutions, high-bandwidth and, in the case of interconnects, low-latency, results. Protocol engines attempt to "off-load" the server by processing the protocols separately, while server interconnects attempt to provide "fast" path messaging of the data. Server interconnects have achieved the best data throughput performance, and are used by platform vendors to support specialized applications such as Oracle Parallel Server (OPS). However, the very act of attempting to reach the required benchmarks by both groups has resulted in their continued investment in expensive proprietary emerging technologies which do not leverage interoperable standards. This in turn increases the company’s risk without providing any assurance that this technology will become widely accepted and economic. Thus, the customer acceptance risks shackle these companies to a narrow base of demanding customers with little hope of breaking free to an economically driven wider audience.

The large players in the server NIC market have the established channels of distribution and manufacturing. However, they have had little success in extending their technology to achieve demonstrably greater bandwidth with reduced delays as the commonly-accepted costly proprietary techniques used by interconnect and protocol engine vendors are antithetical to their focus on reduced cost standards-driven products. Hence, companies, which attempt to partner with these players, must demonstrate the wide customer acceptance and economic viability prior to establishing an agreement.

All three groups converge as an element used by a platform vendor to support server applications. Thus, instead of approaching these groups directly, a better approach might be to partner with established platform players who need to supply this technology in established channels, but would prefer to supply a standard interoperable solution rather than a plethora of proprietary specialized hardware devices. In addition, a careful justification of the performance and utility of the solution through the use of accepted vertical applications benchmarks would reduce the risk of acceptance, allowing dominant server NIC vendors to comfortably consider follow-on partnerships.

Leading Product Comparisons

Two extant products which we believe demonstrate the best available comparison to InterProphet’s Silicon TCP in each of the following product categories: interconnects (GigaNet GNN-1000) and server NICs (3Com 3C984, Intel EtherExpress).

GigaNet is an Intel-backed company which is aggressively entering the interconnect space. Due to its backer, it provides good product integration biased towards Intel’s VIA architecture (which has not yet achieved acceptance). While it appears to be cost-competitive with interoperable Internet-based devices, its requirement of a very expensive switch to do actual messaging between systems is a significant detractor. InterProphet in contrast permits messaging between systems through the purchase of an Internet switch of one-tenth the cost, and can even use hubs available at commodity prices of less than $100. Since it is not hogtied to its switch counterpart, InterProphet’s focus on economic integration with NT and OPS, for example, provides a lower cost entry product than GigaNet can provide.

3Com and Intel are examples of well-established server NIC players with excellent channels to market. They also can afford to develop markets for their products without return for an extended period of time if tactical needs dictate the establishment of a beachhead. These devices usually provide only percentage improvements in nonscalable bandwidth, however, while InterProphet provides orders of magnitude scalable improvement at lower cost. InterProphet’s focus on direct sales to e-commerce sites, platform partners, and third party manufacturers leverages these other groups channels to the customer at reduced cost and reduced risk.

.