Tool interfaces (MPI-T), MPICH parameters and instrumentation
This page describes the design of the MPI Tool (MPI-T) Information Interfaces in MPI-3. MPI-T provides a set of interfaces for users to list, query, read and possibly write variables internal to an MPI implementation. Each such variable represents a particular property, setting or performance measurement from within the MPI implementation. MPI-T classifies the variables into two parts: control variables and performance variables. Control variables correspond to the current MPICH parameters, through which MPICH tunes its configuration. Performance variables correspond to the current MPICH internal instrumentation variables, through which MPICH understands its performance.
Through MPI-T, a user can,
for control variables (cvar),
- Get the number of cvars by MPI_T_cvar_get_num();
- Get attributes of each cvar, which includes its name, verbosity, datatype, description, bind and scope;
- Allocate a handle for a cvar;
- Read / write a cvar through its handle.
for performance variables (pvar),
- Get the number of cvars by MPI_T_pvar_get_num();
- Get attributes of each pvar, which includes its name, verbosity, datatype, description, bind, class;
- Create a session so that accesses to pvars in different sessions won't conflict;
- Allocate a handle for pvar in a specific session;
- Start / stop / read / write / reset / readreset a pvar through its handle.
for cvars and pvars,
- Know their categorization, i.e., how an MPI implementation categorizes its variables, which category contains which variables and which sub-categories.
We wish to have a framework through which components of MPICH can add their parameters and instrumentation uniformly. And it is easy to expose variables to MPI-T interfaces and it is efficient to access variables through MPI-T interfaces.
Current MPICH Parameter Implementation and Its connection to MPI-T
The document Parameters_in_MPICH describes requirements and a potential design of MPICH parameters. However, the current MPICH code doesn't follow this design. Currently, all MPICH parameters are declared in mpich/src/util/param/params.yml in a markup language, which then is parsed and the results are dumped into two files: mpich_param_vals.h/c. The main data structure is a static array MPIR_Param_params, which is initialized to hold info for each parameter, such as its name, data type and a pointer to the parameter.
The current code connects a cvar handle to its corresponding element in MPIR_Param_params to facilitate cvar accesses.
Problems of the current design