HMM Procedure
Distributed Multistart
In the distributed multistart algorithm, the local optimizations that start from various initial points run concurrently. Each of these local optimizations requires evaluation of the objective or constraints as well as their first or second derivatives at different points, which are referred to as function evaluations. If function evaluations are performed in distributed mode, as in the case of PROC HMM, the evaluations necessary for each local optimization are performed in a subsession that stems from the current main CAS session. You can specify the SESSNODES=, FESESSIONS=, and FESESSNODES= options to achieve a balanced and more efficient usage of the computing resources, in particular the available nodes and the threads on each node.
Suppose that there are a total of W worker nodes in the current main CAS session and you specify SESSNODES=m, FESESSIONS=s, and FESESSNODES=n. Then m out of W worker nodes are dedicated to the local optimizations of the multistart algorithm. Hence the total number of local optimizations that run concurrently is
where 
is the number of threads on node i. In addition, s subsessions, each of which has n worker nodes, are created and dedicated to function evaluations. When a function evaluation is requested in one of 
local optimizations, the request is sent to one of the s subsessions. Use the following guidelines when you select the values of the SESSNODES=, FESESSIONS=, and FESESSNODES= options:
The value of m or n should never be greater than W.
The total number of local optimizations
should be as close to s as possible. This allows a good balance between the resources for local optimizations and function evaluations. If you have a limited number of worker nodes W, then in order to achieve the correct balance, preference is often given to a combination that favors a larger s and smaller n over one that favors a smaller s and larger n. This helps reduce the number of function evaluation requests that are waiting to be processed in a function evaluation subsession.
Avoid a scenario in which
is far greater than W. In such a scenario, competition for resources would occur and would degrade performance.