Special Report: Network Provisioning

Networking Infrastructure for Large-Scale Science In June 2003, a roadmap was formulated for DOE networks, which envisions a seamless, high- performance network infrastructure to facilitate the collaborations among researchers and their access to remote experimental and computational resources [2]. Such an infrastructure can eliminate resource isolation, discourage redundancy, and promote rapid scientific progress through the interplay of theory, simulation, and experiment. For example, a timely distribution of multi Petabytes of Hadron Collider data produced at CERN in Switzerland, can eliminate the bottleneck experienced by US physicists today due to inadequate bandwidth in the trans-Atlantic and US networks. Also the ability to access remote, complex, scientific instruments such as SNS or High Flux Isotope Reactor (HFIR) in real time will enable interactive collaborations among geographically dispersed researchers, without the need for coordinated travels and duplications of specialized experimental instruments. Figure 5. Paradigm for DOE networking for large-science applications and network research Two important classes of high-performance networking capabilities are critical to a successful execution of the above tasks. First, large volumes of data must be transported to various end nodes over networks of disparate and varying capacities and traffic. Such data transports might be required in an off-line mode for data archival or post processing operations, or in an on-line mode for interactive visualization tasks. Second, the visualizations and computations must be controlled remotely over wide-area networks to ensure the responsiveness as well as the stability of control loops. This task requires that the higher-order moments of transport delays be kept suitably bounded: high levels of jitter in the control signals can destabilize and steer the remote process into unwanted regions. This problem is particularly acute when the computation is guided by a number of remote experts, each with a different process view, different parameters to control, and with different network connections. The overall network requirements of DOE large-science applications range from the routine to extreme.

The network capabilities to address the DOE large-science needs include the following: 1. 1. Reliable and sustained transfers of terabyte scale data at Gbps to Tbps rates, 2. 2. Remote interactive and collaborative visualization of large datasets of Petabyte scale, 3. 3. Steering of computations on supercomputers and experiments at user facilities, 4. 4. Interactive collaborative steering of computations visualized through multiple perspectives, and

5. Securing scientific cyber environments with minimal impact on applications. In particular, it is essential that these capabilities be transparently available to the application scientists with little or no additional demands on their time and effort to utilize them. In particular, it is not very effective if these capabilities require sustained efforts from teams of network and application experts just to use them.

To adequately cover the broad spectrum of DOE large-science networking requirements, several network research areas have been identified at the workshop and are listed as follows: ? High-Performance Data Transport: For high performance data transfers there are two distinct approaches. At one extreme, TCP methods on shared Internet Protocol (IP) networks can be adapted and scaled to Gbps to Tbps rates. The challenges here include investigating various parts of TCP, such as sustained slow-start and robust congestion avoidance, to achieve the require throughputs. At the other extreme, one could provide dedicated high bandwidth pipes or channels from source to destination nodes wherein a suitable rate control method can be used for transport. In this method, both provisioning and transport methods must be developed (unlike the first method which can be executed on the current IP networks). Nevertheless, this approach circumvents the complicated problem of optimizing TCP congestion control by avoiding it altogether. Note that the network is still be shared (albeit not simultaneously) in this mode by allocating paths on- demand into time-slots for applications. In either case the networking modules must be suitably interfaced and integrated with the middleware and applications.

Previous    Next    Table of Content for report: Network Provisioning    Home

Network Provisioning

 

 

 

Photuris.com - Optical Data Networking