End-to-end high performance networking must continue to develop to keep pace with the computing ecosystem. While the community continues to make significant strides in high-performance computing, the ability to transfer the vast amounts of data generated by scientists has failed to increase commensurately. When scientists spend time dealing with data movement, or are limited by lack of timely access to data, the potential for scientific progress is hampered. Today, there are many application communities struggling with the well-known performance problems in TCP/IP in order to do the real work they need to do. This proposal would support the continued development and deployment of Phoebus, a protocol and service implementation for improving end-to-end transfers to gigabit speeds. Phoebus provides a network "inlay" that augments the network topology with an additional layer of protocol and functionality. Phoebus leverages the ubiquitous TCP/IPv4 implementation on the end hosts, while providing an adaptor to translate to a variety of alternate protocols over the wide area, including creation of virtual networks on the fly. Many data-intensive applications can use Phoebus without modification and without making any system-level changes. Phoebus makes use of a session-layer protocol called XSP, for eXtensible Session Protocol, for end-to-end communication and uses various transport layer protocol instances on different segments of the network, as appropriate. In many cases, performance can be improved by simply using existing protocols tuned based on network conditions and properties. Similarly, a high performance TCP alternative can be used when more assumptions about the underlying network can be made.
Phoebus is essentially a WAN accelerator for research and education networks. This project will work with various application communities to modify their software as necessary to make use of Phoebus and XSP. This will enable the creation of an automatically tuned data movement service, as well as the creation of appliances to act as "on-ramps" to dynamic circuit networks. Broader impact will therefore include significantly improved data transfer performance for scientific applications. Intellectual merit is found in the work's creative approach to adapting the concept of WAN acceleration to a multi-prototcol high bandwidth high delay environment.
