Monday, August 14, 2006
Internet2 Speed Record Team Launches Startup
For the last three years, a group of researchers have been breaking the Internet world network speed records every year at the SuperComputing Bandwidth Challenge. Last year, that group showed a demonstration of an astounding 100 gigabit per second sustained speed, and 131 gigabits per second peak transfer speeds -- enough for transmitting five feature-length DVD movies on the Internet from one location to another in a single second, or the equivalent of serving 10,000 MPEG2 HDTV movies simultaneously in real time. The Caltech team that has been responsible for the FAST TCP protocol that makes that speed record possible today launched a new company to commercialize that technology, FastSoft (www.fastsoft.com). Steven Low is CEO of Monrovia, California-based FastSoft, and also a Professor at Caltech, and I spoke to him ahead of the launch to hear more about the company and the technology.
Ben Kuo: Tell me the story behind FastSoft and how this technology came about?
Steven Low: The company was founded to commercialize research we have been working on at Caltech for the last 5 years. We started in 2000, on a very broad, mathematical model of control on the Internet, looking at the problems with current protocols and to address them. In 2002, the theory has developed to the point that we had implemented, and tested it in a real network. Our first public demonstration was in November of 2002. That experiment broke world records, and surprised many people who didn't think it was possible. A large part of the scientific community at that time had given up on TCP/IP--there were too many problems, they had tried different solutions, and they didn't work well. Our experiment and subsequent experiments changed their view. We also started to work very closely with the scientific community to refine the technology. Rich Wolf, who is at the Office of Technology Transfer at Caltech, strongly urged us to form a company and commercialize the technology. In 2003, we started thinking about the commercialization issues and problems we could address with the technology. Meanwhile, we were devising the technology. In 2004, we decided to commercialize it. Usually, with a project, we do the research, publish the results, move on, and nothing happens. For this project, we thought the technology was interesting, and it had good potential to address a difficult problem. It took us a year to understand as much as we could about business, and we finally formed the company in April of 2005, after incubating at Caltech for a few years. Our first angel funding was in November of 2005, when we had a very preliminary prototype demonstrated to a few companies.
Ben Kuo: Tell me a little bit about the problem you are addressing, and why you thought it was interesting?
Steven Low: There is convergence to TCP/IP for all networks, and that has been the trend for more than twenty years. Ninety percent of Internet traffic runs on TCP. However, the protocol was originally designed in 1970. It really does two things -- one is congestion control--how fast a server sends data into a network. If the network is not congested, things speed up, otherwise, it slows things down. It's not easy to handle this in a distributed manner. The other thing TCP does is loss recovery. TCP will detect and re-transmit data and ensure that a receiver receives it. These functions were developed in 1988. Although the protocol was designed in the 1970's, these two functions were developed in 1988. However, the protocol makes several assumptions, and networks are now several orders of navigation bigger than the assumptions now--those assumptions are being violated. The problem is, when a network becomes higher speed, and larger distance, TCP cannot fully utilize the network capacity. One experiment which we did--transmitting data from Chicago, across the Atlantic to Geneva, we saw that TCP only had 20 percent efficiency. With FAST, we saw ninety five percent efficiency. There are many such examples. With TCP, at high speed, you cannot use the capacity. The second problem, is at low speeds, for example over a DSL link, you create huge latencies. For example, from my home to Caltech, which is a short distance, during a transfer the latency goes from 20 milliseconds to more than a second. Imagine, if I am transfering a file, and my son is using Voice over IP or playing a game or watching a video, the video becomes unwatchable. We address both issues, both the high throughput and reducing latency.
Ben Kuo: Hasn't there been some work on these kinds of issues with Jumbo Frames--how is this different?
Steven Low: It is radically different, but complementary. Jumbo Frames are independent of what we do. Jumbo Frames will help, but the issue with Jumbo frames is that you need ever device in your path to support Jumbo Frames. If one device in your end to end connection does not support Jumbo frames, you can't use it. Indeed, this was one of the solutions scientists had tried before. If you have Jumbo Frames support--which is nonstandard--it mitigates the issues, but in practice, it's difficult to deploy. What Jumbo Frames does, versus what our solution does, is it does not address the TCP issues head on. With our solution you don't need Jumbo frames for performance problems, but it does help reduce CPU load at the server.
Ben Kuo: Where did the idea to look at this problem come from?
Steven Low: From the mid to late nineties, network engineers knew about this problem. There were lots of proposals to address it, with mixed success. To us, it was a curiosity which we could propose academic solutions to. For the scientific community, it's an issue they are facing daily. They have to send huge files around the world, and use TCP every day. We got together with the Caltech High Energy Physics community to work on a project. They send huge files to people in 30 countries, and need a high speed global network to support collaboration. They had built some hardware infrastructure, and faced the problems with TCP, and knew where it becomes a bottleneck. We developed a mathematical model of large scale networks, funded by the National Science Foundation. We quickly got interest from the Army research lab and other programs in the NSF. Our project was jointly funded by Caltech, UCLA, SLAC (Stanford Linear Accelerator), and the University of Cambridge.
Ben Kuo: What made you decide to commercialize the technology?
Steven Low: Our November 2002 experiment, plus our experiments in 2003, generated lots of interested--not only in the research and networking area, but also in the market and industry. Several companies, both in Northern California and Southern California, came to understand what we had. There was a lot of interest, which is why we wanted to commercialize the technology. What we weren't sure of--because we knew what the technology could and couldn't do--was if the scientific community was enough to sustain the company. We were not convinced the scientific community was a good business case, so we weren't interested in commercializing it. The question we did not know was if there was an important commercial pain. It took us awhile to decide we should commercialize, but fortunately, that gave us three things. One, it gave us lots of time to refine the technology and test it with collaborators, two, it gave us time to learn about business and management, and three, in retrospect, we hit the right timing. IP video is going to explode, and although no one knows how it will play out, it is like VoIP was four or five years ago. It will evolve, but there will be a fundamental transformation with video. We have been talking about digital video and network video since the eighties, but the difference in the last year or two is the business mode. The iPod really made a huge difference. Two or three years ago, a group of people from Warner Bros. came to Caltech to understand our research. The thing that dominated the discussion back then was privacy and copyright protection--they were very scared of digital video. People's mindsets have changed, and the question now is how do they actually make a business that works. The iPod has convinced people that there is a legitimate model to beat piracy, and that it can be extremely profitable. So we have been working on our business model--the technology has never been a bottleneck.
Ben Kuo: How does this go to market--are you selling hardware, or licensing technology?
Steven Low: It is an appliance. Our value is in the software, but we sell it in a box. You plug the appliance into your network, and you don't need to integrate this or that, you don't have to change applications, or your workflow, or your network. It's just one box at the sender, which accelerates file transfer.
The product is the Aria 1000. It allows you to use this in any environment, not just in your own or across different sites of the same enterprise. You can use it for file transfers to customers, vendors, and so on.
Ben Kuo: Let's talk a bit about your funding. How are you funded?
Steven Low: It's relatively unconventional. We started to raise funds late last year, and at that time we were not ready for venture capital. We raised angel funds to figure out what we should do. We got the funding very quickly, and we started to operate and hire people, so that things could become clearer and we knew what we had to do. When we raised an angel round, more and more angels were interested that even though we raised the round we had to double it. The round was oversubscribed just by the angels. Therefore, our thinking now is to take that angel money, and deliver market traction before venture capital. We do have an institutional investor considering putting in a small amount in this round. They are interested in the next round--the venture capital round--but want to put in a small amount to help the company.
Ben Kuo: So what are your thoughts on how this technology will influence the industry?
Steven Low: The revolution that has started means that the way video is being distributed in the future will be very, very different from today. That revolution is ongoing, whether or not we participate. We want to participate and contribute a lot to that transformation. One of the areas is in B to B -- for example, production and post production. There are about 300 editorial production houses just in the LA area. If you broaden the definition of post production houses to include ad agencies and related companies--companies that need to work on video within their ecosystem--there are 1400 companies just in the LA area. All of them need to send large files around. All of these are potential customers we can help. Another natural example is Apple iTunes, or Google Video. They have content partners--hundreds now, growing to thousands--who have to send large files to the iTunes store every day. Those people need to send large files, video files that you can't compress, and face problems with TCP daily. That's one of the market verticals we can help. The bigger media market, which will be slower to develop, is B to C, which we can help in the future. Even though we can potentially develop solutions into the horizontal, we think we have to focus on a few vertical, to validate our technology and validate our business case.
Ben Kuo: Thanks, and good luck!