4:00 PM-6:00 PM
Room: Capitol South
The concept of a computational grid has emerged to encapulate the vision of a network computing system that provides broad access to massive information resources, and to massive computational resources. The grid is a powerful environment, but it is also very heterogeneous and its behavior is difficult to predict, making scheduling and running meta-computing programs challenging issues. There is therefore a clear need for appropriate frameworks through which domain scientists can benefit from the grid's power while being isolated from its complexity. A popular approach is the use of network-aware servers, i.e. servers that grant remote access to software components on various hardware computing resources. The end-user usually contacts the servers by the mean of simple and possible interactive interfaces; in addition scheduling is generally transparent. This approach has many advantages: the user is unaware of the grid's complexity and does not need to install or upgrade any software libraries, the pool of computational resources is dynamic, new interfaces can be easily developed to satisfy the needs of new users, and many tools are available to implement network-aware computational servers. However, there are also shortcomings, such as restrictions imposed on the programming model, management of data locality in conjunction with scheduling, lack of tools to predict grid conditions, and lack of standardization. The speakers in this minisymposium will discuss those shortcomings.
Organizer: Henri Casanova
University of California, San Diego
LMH, 1/19/99, MMD, 2/16/99