A novel approach for distributed application scheduling based on prediction of communication events
Future Generation Computer Systems
| Hi-index | 0.00 |
Recent researches on parallel and distributed systems present drawbacks and limitations regarding to the analysis, design, implementation and execution of high performance parallel applications. Complex issues related to the development of such high performance parallel applications vary from the modularization, in the design stage, to the task allocation and distribution. Such difficulties have motivated the design and implementation of MidHPC (Middleware for High Performance Computing), a middleware which automatically and transparently balances process workloads using knowledge obtained from techniques to extract and predict application behaviors. Through this middleware, application tasks transparently communicate among each other through a distributed shared memory system. In order to better take advantage of the MidHPC middleware, parallel applications are written using the POSIX threads standard. Applications developed following this programming model can be executed in SMPs, Clusters and Grids without any source code modification nor any additional change. Experimental evaluation of the proposed middleware shows not only good scalability and performance can be achieved, as also efficient process migration.