Difference between revisions of "Consolidate Domain Decomposition"
From MohidWiki
(→Topology) |
|||
| Line 9: | Line 9: | ||
[[File:Ddc.rcm.png|thumb|left|alt=Diagram of the architecture for the consolidation of domain decomposition HDF result files.]]There are 2 kinds of MPI processes running: | [[File:Ddc.rcm.png|thumb|left|alt=Diagram of the architecture for the consolidation of domain decomposition HDF result files.]]There are 2 kinds of MPI processes running: | ||
*a Parser; and | *a Parser; and | ||
| − | * | + | *any number of workers. |
The Parser identifies result files to be consolidated. It keeps 2 lists: | The Parser identifies result files to be consolidated. It keeps 2 lists: | ||
Revision as of 11:34, 3 February 2015
The program Consolidate Domain Decomposition uses the actor model to parallelize file processing tasks. For instance, suppose there are hydrodynamic and water quality files from 2 sub-domains:
- hyd_submod_1.hdf and wq_submod_1.hdf; and
- hyd_submod_2.hdf and wq_submod_2.hdf.
The desired consolidated files are hyd.hdf and wq.hdf, resulting form the concatenation of:
- hyd_submod_1.hdf and hyd_submod_2.hdf to produce hyd.hdf; and
- wq_submod_1.hdf and wq_submod_2.hdf for the wq.hdf
Topology
There are 2 kinds of MPI processes running:
- a Parser; and
- any number of workers.
The Parser identifies result files to be consolidated. It keeps 2 lists:
- a list of jobs to be done; and
- a list of workers.
The list of workers will be used in the end, when there are no more pending jobs. A message will be send to every worker notifying them that they should terminate. The list of jobs works like this:
- a Worker is idle it sends a request to the Parser;
- the Parser sends the first unfinished job to the worker and moves it from the first to the last position of the list;
- when a Worker terminates a job sends a processed message to the Parser;
- the Parser removes the completed job from the list.
This strategy for the jobs list ensures fault tolerance. Suppose a job is sent to a Worker that, for some reason, does not complete it. That job will eventually be send to another Worker. It may occur that a job is executed twice but the program is idempotent so no harm occurs...
