Gigabit Ethernet Switch Gigabit Ethernet Switch

The Internet has created an opportunity for collaboration between scientists at unprecedented levels. Furthermore, in the late 80's, researchers developed general approaches to programmonitoring[1], [2], [3], followed by advances in program adaptation[4] and steering[5]. Speci cally, by coupling program monitoring and steering with online visualizations[6], [7] of scienti c data, it has now become possible for scientists to interact with their simulations at runtime and in a variety of ways that do not depend on simulation designs or implementations. As a result, intermediate program values may be inspected at will, parameters may be changed `on the y', and data values may be selected to `guide' applications into interesting data domains. Moreover, by extending the functionality of restart les, it has become possible to stop, rewind, and rerun applications when simulation output does not agree with observational data or when scientists deem ongoing runs uninteresting. In summary, such interactivity, also called `program steering', removes the separation in time between the scientist and the computational tool being employed. Program steering has the potential of improving end user productivity, yet its current realizations still con ne multiple scientists to interact with their applications via single visualization clients. This situation is not satisfactory, especially considering the explosion in network bandwidth of the early 90s and the broadening diversity of resources with which a researcher can participate in scienti c exploration, from high-end graphics workstations to home PCs with modem connections to the Internet. In response, research being conducted now aims to enable multiple scientists to interact with each other via shared complex scienti c applications, from geographically distributed locations, using diverse computing and networking resources, and such that each scientist is able to interact with the application via operations and data suited to his/her expertise. Recent directions in scienti c computing have also been in uenced by signi cant advances in compute power, making possible complex simulations that were heretofore unrealistic, including heterogeneous simulations simultaneously considering multiple, linked physical processes (e.g., atmospheric and oceanic modeling[8]) or simulations that consider multiple physical models at di erent time or length scales. For example, the atmospheric model discussed in Section IV couples a parallel spectral transport model with a grid-based chemical model. The environmental hydrology project [9] simulating the Chesapeake Bay integrates atmospheric models describing the physics of clouds and predicting rainfall, a river model predicting ow in stream channels, and a wind model describing surface ow patterns. Finally, our future work aims to couple a global atmospheric model of fairly low granularity with one or more regional models, perhaps attempting to understand the global e ects of pollution in a certain metro area. Therefore, these models are not only heterogeneous, but they also operate at highly di erent scales (i.e., kilometers versus meters). The opportunities in scienti c collaboration presented by the Internet coupled with signi cant advances in compute power suggest the inadequacy of existing models of heterogeneous parallel computing, like PVM or MPI. This inadequacy has already been recognized by projects like Globus[10] for running large-scale, distributed scienti c codes, interactively, across heterogeneous target systems. In addition, both the NCSA and PACI Supercomputer Centers in the U.S. are developing methods and tools for interactive use of large-scale data or simulations, and for scienti c collaboration via both high end immersive systems and low end, browser-based interaction media. Our work contributes to scienti c computing by improving the scientists' freedom to interact online with their applications and with each other, from remote locations, their o ces, and their laboratories, using interaction media suited to their locations and current needs. We call such research environments Distributed Laboratories; they are characterized by: multiple data sources, coupled models of varying levels of granularity, interactivity, multiple collaborating scientists, and