A faster Internet lane
- By William Jackson
- Sep 28, 2008
MANY OF THE WORLD'S research and education networks face a conundrum: After upgrading network backbones to multigigabit speeds, performance for users often lags.
So several of the nation's high-performance networks are experimenting with new protocols to make better use of network capacity for large science projects that require moving gigabytes, and even terabytes, of data.
'The problem we are addressing is more effective end-to-end utilization of existing resources,' said Martin Swany, assistant professor of computer and information sciences at the University of Delaware and lead developer of Phoebus.
Phoebus is a platform through which applications communicate with networks to better allocate resources, based on the application's needs and network conditions. It is being deployed experimentally on the Internet2 backbone and also is being studied by the Energy Department's Energy Sciences Network and the New York State Education and Research Network, NYSERNet. When appropriate, a file transfer can be shunted off the traditional TCP/IP portion of a network and moved onto a dedicated path between Phoebus gateways, eliminating the latency associated with TCP.
'Our Phoebus testing has shown dramatic performance increases, even while using well-tuned applications that were already achieving good performance over the routed IP network,' said Bill Owens, NYSERNet's director of advanced technology and networking.
One NYSERNet project, the Laser Interferometer Gravitational Wave Observatory (Ligo), illustrates the challenges and opportunities in moving massive amounts of data. 'One Ligo researcher in our region has been able to increase his application performance by over 10 times, reducing the time needed to transfer a dataset from 40 days to less than four,' Owens said.
Raw bandwidth is not enough to provide high performance on large-scale networks. The Energy Department's Office of Science, which runs ESnet, heard about that when it held a series of workshops to determine requirements for its science community, said Brian Tierney, leader of DOE's Advanced Technologies Group at Lawrence Berkeley National Laboratory.
'One of the scientists told us about a year ago, 'We don't use the network because it's too slow; we use portable USB drives,' ' Tierney said.
And yet ESnet provides some big pipes for its scientists through a partnership with Internet2, a network operated by a consortium of companies, universities and research organizations. Internet2 recently built a new 13,500-mile fiber-optic backbone with Level 3 Communications (GCN.com/1198). Some Internet2 segments are operating at 100 gigabits/ sec now, and plans call for speeding that rate to 200 gigabits/sec.
With current technology, the network can easily scale to 400 gigabits/sec, but TCP imposes restrictions on using that bandwidth over long distances. TCP was designed for shared networks such as the Internet, and it acknowledges receipt of packets so that errors can be corrected. With large data flows over long distances, the latency associated with this process interferes with performance, creating a gap between network capacity and performance experienced by users.
'From our end of the network, we're not using what we've got,' Tierney said.
As a way around this, Internet2 is using Dynamic Circuit Networking (DCN), a technique for segmenting traffic and providing dedicated channels that combines the best qualities of packet-switched and circuit-switched networks. But users need a way to access these alternative circuits, and even nuclear physicists do not necessarily understand networking protocols. That is where Phoebus comes in.
'The main motivation wasn't just to take traffic off the IP network, but to deal with the problems associated with TCP over long distances,' said Internet2 Chief Technology Officer Rick Summerhill. 'You essentially take a TCP flow and break it into three pieces.'
One segment goes from the host to the Phoebus gateway, one long-haul section travels between gateways, and a final hop goes from the second gateway to the destination. 'If this is a very large flow, Phoebus will take it off the normal IP network, which is shared, and put it on a dynamic long-distance link that is dedicated.'
The two shorter links using TCP at the beginning and end of the journey typically would not be long enough for latency to create problems.
Phoebus is more than just a switch between IP and DCN segments. Phoebus provides an additional service layer in the protocols that lets it choose and negotiate the best transport technology for an application from the available options.
The Phoebus gateway ' a server running Phoebus software ' does the negotiating. It communicates with the application through a software 'shim' installed on the user's system.
'Our software inserts itself fairly transparently' so that even a physicist can use it, Swany said.
The Phoebus code is mature, Swany said, but work with the new platform on a live network still is in the experimental stages. Deployment of the gateways on Internet2 was delayed while the new network was built, but nine are in the process of being installed on the network.
Research networks in Ireland, Poland, Greece and Switzerland also are installing experimental Phoebus gateways.
'None of these are in production yet,' Swany said. 'We need a lot of feedback from the community and more experience using them.'