For the purposes of a discussion of the Analyst Drive program, we will discuss two types of cores, Physical and Virtual. A physical core is one that is actually present on a CPU, while a virtual core represents a thread process which can use a physical CPU core. CPUs that employ hyperthreading can have more virtual cores than physical CPU cores. This can be beneficial as the virtual core can perform a computation on a physical core while another virtual core may be busy fetching data and is not ready to execute and would otherwise leave the core idle. We will also discuss two types of memory, shared and distributed. Shared memory is local to a processor (or several processors) and has cache coherence, meaning that all CPU cores that use the same shared memory can access all shared data. You can think of your desktop or laptop computer as a shared memory computer. On the other hand, a distributed memory computer has disjoint memory sets; thus a CPU core on such a system cannot share information with another core without using some type of message passing.

Most modern computing clusters have hybrid memories; in other words: shared memory within some processors of a single compute node, and several disjoint compute nodes which make up the cluster. This is important information to consider when running a modern parallel program. The Analyst Drive program employs simple threading for shared memory systems so that the benefits of parallelism with cache coherence can be realized, providing large benefits even on a simple desktop or laptop computer with a multicore CPU. At a higher level, Analyst Drive is built using a message passing library so that it can use the parallelism available to distributed memory clusters, but at the same time can be used on a shared memory system. While the shared memory threads defined by the NCORES parameter are lightweight and dynamic (they are created and destroyed as needed), the distributed memory threads of the MPICORES parameter are heavy and static. These threads launch separate processes and will remain active throughout the life of the program. When running Analyst Drive for a single user class problem it is optimal to set NCORES to the number of virtual cores present on the computer. When running Analyst Drive with the MPICORES parameter in addition to NCORES, the optimal choice depends on the computer setup and the number of user classes to be run.

This choice can be a challenging decision to the inexperienced user. In the next section, Performance metrics with NCORES and MPICORES, guidance is provided through the examination of some performance metrics for a set of cases that use the two parameters.