Special Report: Network Provisioning
5. Workshop Findings in Network Transport Protocols Area The current dominant Internet transport protocol, TCP, was originally designed and optimized for what we now consider low-speed data transfers over low bandwidth connections. It lacks the performance and scalability to meet the challenges of DOE large-science applications in terms of large throughputs as well as agile and stable dynamics. To achieve high throughput data transfers, TCP methods on shared IP networks can be adapted and scaled to Gbps and Tbps rates. But, this is a challenging task that requires investigations into various parts of TCP, including sustained slow-start and robust congestion avoidance, to achieve the require throughput levels. At the other extreme, one could provide dedicated high bandwidth channels from source to destination nodes wherein a suitable rate control method can be used for transport. This approach avoids the complicated problem of optimizing TCP by avoiding congestion altogether, but still requires mechanisms to account for non-congestive packet losses and suitable flow control to optimally utilize the provisioned bandwidth. Recently there have been several UDP-based methods that attempt to fill the available bandwidth but such methods can have very negative effects on the TCP transfers simultaneously taking place on shared links.
Current transport methods are massively inadequate to meet the multitude of DOE largescience networking requirements. The required throughput levels are unattainable except with significant efforts from teams of technical experts, that too often only in demonstration scenarios and typically for small periods of time. But such throughputs are needed at the application level on a daily basis. Furthermore, it is also important to sustain the throughputs during the entire execution of the application rather than ephemerally achieving the peak bandwidth. Currently, TCP methods are not able to provide sustained and stable streams for control operations particularly in networks with heavy traffic loads. Since TCP has provably complicated dynamics, it might be difficult to use it for control operations. The transport protocols needed to support control operations on dedicated channels must be developed particularly for long haul connections. Another important consideration is that the time-to-solution in the protocols area is currently too high; for example TCP tuning for Gbps throughputs took several years. In view of impending DOE needs, it is important to develop the needed protocols in a much more timely manner. It is also important to develop and integrate the functionalities between middleware and transport, and these aspects are discussed in the next section.
Previous Next Table
of Content for report: Network Provisioning Home
