High performance numerical computing with YML (Application)
Conducted byOlivier Delannoy, Nahid Emad
The context of this experiments is detailed here.
Numerous problems can be modeled using linear algebra application such as Linear system solving and Eigenvalue problems. In order to implements solver for large problem, the CNI team of PRiSM laboratory has evaluated various approach to implements hybrid method such as Multiply Explicitly Restarted Arnoldi Method(MERAM). This approach includes : Sequential and Parallel (MPI) object oriented library on cluster and parallel HPC computers, Problem Solving Environment using Netsolve and more recently workflow application description using YML.
An iterative method computes its result by creating a succession of partial results. Those results are approximation of the solution of the problem. Each iteration try to produce a new solution based on the results acquired at the previous stage. The performance of an iterative method is measured based on the number of iteration needed to reach the solution. A way to improve the performance is to combined several numerical method together in order to decrease the number of iteration needed to converge. MERAM is such a method and use multiple instance of the ERAM process which exchange intermediate results asynchronously. With the previous approaches we have been using based on MPI we have not been able to implement scalable version of MERAM and to use them with a high number of collaborating ERAM process. The use of YML as our main application description allowed us to easily change the number of ERAM processes collaborating.
The aim of these experiments is to evaluate the effect of a large number of ERAM process on the number of iteration needed to converge. We also want to evaluate our methods on huge problem with sparse matrix of order 10^6 and more. Results of this experiments will hopefully gives us more knowledge on the scalability of hybrid methods. We will also be able to evaluate YML with complex graph and realistic applications on real problem. For those experiments we will use matrix generator from Matrix Market in order to be able to use arbitrary huge matrices.
- Nodes involved: 500
- Sites involved: 3
- Minimum walltime: 8h
- Batch mode: no
- Use kadeploy: yes
- CPU bound: yes
- Memory bound: yes
- Storage bound: yes
- Network bound: yes
- Interlink bound: yes
Tools usedNo information
More information here
Shared by: Olivier Delannoy, Nahid Emad
Last update: 2007-02-28 16:48:16