I think that you should separate the work from the database. MPI may be used for that, but perhaps that some SQL features could it for you, but I'm not enough expert on that. So let's try using your approcah, but in a different way.
First you should consider to isolate your computation workload from the database itself, for multiple reason, one is to leave SQL Server enough CPU power to do its own job, a second is from an architectural point of view, an independance from the datasource itself.
what you could eventually do is to query the data from your mpi application. the main MPI rank can start by querying the database to get a dataset, then create the MPI domains, and broadcast the initial set of data.
Then on each slaves, query the database on the domain you are working on, and do your matching job.
But in reality there is no relationship between each MPI tasks, because each is working on a separate subset. So maybe MPI is not the most appropriate for this, a parametric sweep job could be an easisest way to go.
HTH
Xavier
