Upgrading the Internet with PlanetLab

Like many great ideas, the next generation of the Internet was born out of frustration.

In 2002, Princeton Computer Science Chair Larry Peterson had a problem: The computer science community was buzzing with a new generation of applications and services – and nowhere to run them. These services – Web content distribution, media distribution, an adaptive form of Webcasting called multicasting, efficient file transfers, wide-area storage systems, worm tracking, and network monitoring – had to run everywhere, all the time, and nobody had access to a computational services network that ran all over the world.

So Peterson decided to build one.

“Everybody contributed a computer or two, creating an instant, shared services network where we could test and deploy this new generation of network services," he recalls.

And thus PlanetLab was born in June 2002. HP threw its support behind the initiative in April 2003, joining the consortium as a charter industrial sponsor.

"PlanetLab is collaborative research at its best. Neither we, nor Intel, nor Berkeley, nor Princeton alone could create the next-generation Internet services that will be born in PlanetLab," says Patrick Scaglia, director of the Internet and Computing Platforms Research Center in HP Labs and a vice president for research, HP. “More than being just a research facility, PlanetLab has spurred us in entirely new directions."

Worldwide computing utility

Today, PlanetLab consists of 530 nodes at more than 200 sites worldwide. It’s become a backbone of research and education networks in Canada, Brazil, China and the United States. PlanetLab membership is expected to double by the end of 2006 as universities and companies join the consortium. Google, AT&T, France Telecom and the U.S. Public Broadcasting Service (PBS) have joined Intel and HP as corporate members.

With a worldwide computing utility like PlanetLab, researchers -- from HP, Intel, Google, Princeton University, the University of California at Berkeley, Carnegie-Mellon University, the University of Washington and elsewhere -- can do completely new things and perform existing tasks much more reliably and robustly. Serving as a deployment platform, collaboration vehicle and world-class test bed, PlanetLab has helped researchers to speed up file transfer, reduce the cost of webcasts, track the spread of worms, develop stronger file protection, and much more.

“Even more than the specific projects, PlanetLab has achieved two key miracles,” says Rick McGeer, HP program manager for PlanetLab. “It’s the first democractic world-class experimental facility in any field -- any university can join just for the price of a couple of PCs and by allocating some bandwidth."

What's more, adds McGeer, PlanetLab has provided a platform for cross-institutional collaboration. "Previously, if people wanted to collaborate on some project, they had to go through a maze of logistics to make that happen," he says. "Now, it’s easy: Just set up a shared slice on PlanetLab."

Seamless global collaboration

Such global collaborations form the heart of two major projects on PlanetLab. In one, HP is heading a three-company, four-university consortium designed to make Internet applications more robust and secure by using a common distributed program to route data around the Internet. This program can make routing decisions based on remote information, avoiding hot spots, congestion and other problems. So if some parts of the Internet fail, system operators could still be able to route data without users experiencing slowed service or disruption.

The project requires that all seven institutions in the consortium can easily interact, and that technologies contributed by each institute can work together seamlessly.

"Not long ago, just putting all that together would have been a nightmare," recalls HP Labs researcher Jack Brassil, HP's technical lead for PlanetLab. "Today, putting a big project involving many organizations together in a few months is not only possible, it’s quite tractable.”

The other project, co-led by HP, is an effort to distribute high-definition television signals for PBS. It, too, is highly complex, involving numerous technologies and institutions. Instead of worrying about logistics, says Brassil, "we can focus on solving the problem – in this case, getting nine channels’ worth of TV programming from PBS in Washington to 180 affiliates around the United States.”

Recent projects

• Better Web content distribution: Existing Web servers can be overwhelmed by unusually heavy traffic. CoDeeN, an experimental content distribution network on PlanetLab, aims to avoid this problem by storing frequently requested pages throughout the network, replicating popular pages and serving up the stored pages where possible. By using many sites throughout the network, CoDeeN can potentially store and replicate many more pages than conventional Web proxies; its network of Web proxies, therefore offers much faster Web performance and prevents servers from being overwhelmed.
• Stronger file protection: Currently, a computer user’s files are stored in one location only, making users and their data vulnerable to various local disruptions such as power failures, a local loss of connectivity, or a disk crash. A new Internet-wide file system known as OceanStore is designed to eliminate vulnerabilities by operating as an Internet-wide Redundant Array of Independent Disks (RAID) system. Instead of using multiple disks on a single machine to make redundant copies as RAID systems do, the system spreads those disks over multiple machines. The goal is of the project to make it impossible to corrupt a file and to allow system operators to do backups by copying from disk to disk -- so that, potentially, a file could last for a thousand years.
  
• End-system multicast: Currently, Webcasts are extremely expensive and difficult to set up. End-system multicast, from Carnegie-Mellon, sets up an ad-hoc multicast tree that any new user can join, dropping the cost of Webcast to the cost of a conference call.
  
• Worm tracking: Because PlanetLab is global, researchers at Intel learned they could track the spread of worms globally, simply by logging requests to a PlanetLab server. By studying this data, they discovered that 60 percent of all worms originated from just 10 IP addresses. This knowledge could lead to development of an early-warning monitoring service; network administrators could subscribe and then refuse connections from infected IP addresses. No more worms.
  
• More efficient file transfer: File transfer is currently handled point-to-point, even when large files are transferred to many recipients simultaneously. Transmitting the whole file, one copy at a time, to each recipient, overuses bandwidth at the source without ever tapping the recipients' bandwidth. Lucy Cherkasova of HP Labs proposed a solution called FastReplica, in which the sources break the file into numerous pieces and transmit one piece to ach recipient. The recipients each transmit their piece to every other recipient, resulting in potentially much faster and more efficient file broadcast.
  
• Faster, more efficient network monitoring: All the above applications are dependent on network conditions both locally and remotely. As a result, sensing applications –applications that discover and report on network conditions – are widespread on PlanetLab. One of these is a system developed by Puneet Sharma, Sujata Banerjee, Sung-Ju Lee and Sujoy Basu of HP Labs that is designed to efficiently determine nearest neighbors and bandwidth on the network. The conventional approach to measuring distance between two sites is for one to send a short message, called a ping, to the other, and measure the time it takes the second site to respond. This method is inefficient because it requires messages between each pair of sites -- so measuring the distance between all pairs of sites in a 1,000-site network would, for instance, generate a million messages. Most new methods that require a small number of measurements from each node suffer from lack of accuracy and are adversely affected by incomplete or erroneous measurements. The HP Labs system uses novel clustering techniques to find nearest neighbors with low-measurement overhead. This system is not only efficient, but also designed to be highly accurate and robust to measurement errors, making it practical to use in very large networks.
  
• Wide-area caching services: The Logistical Computer Infrastructure out of the University of Tennessee, Knoxville, offers wide-area caching services, liberating people from the need to provision and cache local storage.

The power of many

Some new HP projects on PlanetLab include an effort to build an adaptive wide-area distributed system management platform, a project involving Dejan Milojicic of HP Labs and Intel’s Robert Adams. Kevin Lai of HP Labs and Rob Knauerhause of Intel are working on a joint project to provide sophisticated, market-based resource allocation, and provable isolation and security to distributed applications.

“PlanetLab has become a centerpiece of our relationship with HP,” says Mic Bowman, Intel distinguished engineer and the architect of the PlanetLab kernel.

In addition to two joint innovation programs, HP and Intel teamed to build out the Brazil and Chinese national research networks, and they are partners in an initiative to develop commercial services on PlanetLab. Says Bowman, “PlanetLab has shown us that we can do things together that we simply couldn’t have done separately.”