Getting to know the grid to use it better. |
|
|
|
|
|
|
| The Grid is often presented as the next eldorado of scientific computing. Thanks to this technology, it will be very easy to use the power of the biggest machines from hand-held computers. World-wide computers will unite to sell their services to the highest bidder. Unfortunately, the reality is a bit different. The use of this platform still comes up against a lot of problems. In contrary to the previous parallel systems, the Grid presents dynamic, heterogeneous and even not dedicated performances. Upstream from any scheduling problem, the main difficulty is then to acquire pertinent, recent and accurate information about the system state, and about the expectable performance in the next minutes or hours. We will present several complementary tools to solve this problem. The first one is NWS, developed at the University of California Santa Barbara. It constitutes a coherent system to monitor the platform availability in term of CPU and memory load as well as in term of bandwidth between hosts. The second tool is FAST, a home-developed library which aims at predicting the needs of routines to schedule in terme of memory space, computation time and communication amount. The main method for that in FAST is to benchmark routines at installation time. The third tool is ALNeM, also developed at the ENS-Lyon. Its goal is to get the topology interconnecting several hosts to detect contention points and predict the bandwidth achieved by concurrent data stream. This presentation will be concluded by a short presentation of the Grid middleware DIET (also developed at the ENS-Lyon) and how it uses the presented tools to get the needed information about the platform. |
|
|
algweb@cerfacs.fr Last Update: Mar 27, 2003 |