Glossary


application group
The gang to which a process currently belongs.

assertion
An argument to a one-sided synchronization call for making a statement about the context of the call, for example, that no epoch will follow a fence operation. Assertions are used by the implementation to optimize the synchronization operation. An assertion must represent a true statement, otherwise the behavior of the call is undefined.

attribute
Optional information that can be assigned to a communicator. An attribute is local to an MPI process and is specific to the communicator to which it is assigned.

attribute key
A key that identifies an attribute that is assigned to a communicator.

barrier synchronization
An operation that waits for the arrival of all MPI processes in a communicator. Until all MPI processes arrive, the MPI_BARRIER procedure is blocked from returning from the procedure call.

blocking
 Type of communication procedure call in which control does not return from the procedure until either the send or receive operation is completed. When a send operation is completed depends on the particular transmission mode that is in effect.

broadcasting
An operation that transfers data contained in any MPI process (root process) to all other MPI processes.

buffered mode
A send request that can be issued regardless of whether the corresponding receive request has been issued. In a buffered-mode send, the contents of the send area are transferred through a user-defined buffer to the receive area. MPI defines a buffered-mode send as complete when all contents of the send area have been transferred to the user-defined buffer.

canceling a communication
Requests the cancellation of a pending (not completed) nonblocking communication operation. Some cancellation operations succeed while others fail. A cancellation operation is a local operation, and control returns from the cancellation-issuing function regardless of whether or not the cancellation was successful.

Cartesian topology
The mapping of an MPI process to the coordinate axes. A topology is expressed by the number of dimensions, the size of each dimension, and an attribute that contain information indicating whether or not a given dimension is periodic.

C-binding
A C call format.

collective (communication) operation
A (communication) operation in which all processes in a communicator are involved (barrier synchronization, broadcast, reduction, etc.)

communicator
An opaque object that indicates the range of communication operations to be performed. There are two types of communicators; intracommunicator and intercommunicator.

communication mode
One of four modes of send operations: standard, buffered, synchronous, and ready. See the definition of each mode, elsewhere in this glossary.
Receive operations are not subdivided into modes.

communication request (MPI_Request)
An opaque object that identifies a given communication operation during nonblocking communication. It is also used to mark the beginning and the end of a communication operation.

data map
A generic representation of data types. It is a list of sets of (basic data type, displacement).

data shift
The transfer of data along a coordinate axis in the Cartesian topology.

data signature
A list representation of basic data types.

data type
An opaque object that defines the type of data that is stored in the send/receive area. Pre-defined and user-defined data types exist. Examples of predefined data types are MPI_CHAR and MPI_INT. A user-defined data type is created when the user specifies the desired data structure and executes an appropriate MPI procedure.

error class
A standard error code that is independent of the MPI/SX processing system specifications. An error class is a subset of an error code.

error code
MPI procedures return information, as an error code, indicating whether or not an abnormal condition occurred during their execution. If no abnormal conditions occurred, an agreed upon value is returned to the MPI-reserved name MPI_SUCCESS.

error-handling function, error handler
This function is called and executed when an error occurs during the execution of an MPI procedure. An error-handling function is assigned to a communicator. User-defined error-handling functions can also be assigned. Error-handling functions are opaque objects.

extent of a data type
The size of the data type, which can be determined by the formula (upperbound - lowerbound).

FORTRAN-binding
A FORTRAN call format.

gather
An operation in which all MPI processes in a communicator send data to the root process and the root process receives data from all processes. gather is the inverse of the scatter operation.

graph topology
The mapping of an MPI process to a graph. A graph is expressed by the number of nodes and attributes that contain information on edges that exist between the nodes.

group
A collection of MPI processes. Each of the MPI processes that belongs to a given group has a unique rank within the group.

handle
Permits the user to access opaque objects for comparing or substituting them.

info object
Information that can be provided to certain MPI calls like creation of a window for one-sided communication or opening of a file. The info object provides further information to the operation to take special actions, or that certain optimizations are possible.

intercommunicator
A communicator that indicates a range consisting of two groups (local and remote groups).

intracommunicator
A communicator that indicates the range of a single group.

leader
The MPI process that represents a given group. The leader of group can communicate with the leader of the other group.

local group
One of two groups that constitute an intercommunicator; it is the group that contains the MPI process that issues a communication operation using the intercommunicator. In other words, the local group is the group to which the sender process belongs during a transmission operation and the group to which the receiver process belongs during a receive operation.

local operation
An operation in which completion of a procedure depends solely on the execution of the local process. An example of this operation is buffered-mode send.

lower bound of a data type
The minimum displacement within a data map.

nonblocking
In a nonblocking communication procedure call, control returns from the procedure before the transmission or reception (send/receive) operation is completed. When using a nonblocking communication, the program should not change the contents of the transmission or receive area before the send/receive operation is completed or an error can occur. The completion of a send/receive operation can be detected by using a communication completion confirmation procedure such as MPI_WAIT or MPI_TEST.

nonlocal operation
An operation in which the completion of a procedure depends on the status of other processing tasks. An example of a nonlocal operation is synchronous-mode send.

one-sided communication
A communication operation in which a process communicates with another process without the explicit involvement of the other process.

opaque object
MPI provides objects that are not directly accessible to users. These objects are called opaque objects. Access to an opaque object requires the use of a handle.

packing data
Compressing noncontiguous data into a contiguous area.

peer_comm
In the two groups that constitute an intercommunicator, the peer_comm is a communicator that contains the leaders of the groups. Membership in the common communicator permits the leaders to communicate with each other.
point-to-point communication
A communication operation in which two processes communicate with each other where both are explicitly involved in setting up the communication.

periodic
In Cartesian topology, identifies whether the coordinate axis of a given dimension is continuous or discontinuous.

probe
Inspection of transmitted data. By probing, the program can obtain information regarding received data (transmission source, tag value, data length, etc.) without actually receiving the data.

rank
The identification number for an MPI process in a given group or a communicator. In the case of a communicator, the rank is associated with the group that represents the range of the communicator. In the case of an intercommunicator, no ranks for the entire communicator exist; ranks exist only in local and remote groups (ranks that are local to a given group).

ready mode
A ready-mode send request can be issued only when the corresponding receive request has been issued. Issuing a ready-mode send request before a corresponding receive request is issued results in an error. MPI defines a ready-mode send as complete when all contents of the send area have been transferred.

reduction operation
An operation that calculates the sum, total area, maximum, and minimum of the data contained in all MPI processes that belong to a given communicator. There are two types of reduction operations predefined and user-defined reduction operations.

remote group
One of two groups that constitute an intercommunicator; it is the group to which the destination process belongs in a communication operation using the intercommunicator. In other words, the remote group is the group to which the receiver process belongs during a transmission operation and the group to which the sender process belongs during a receive operation.

root process
MPI process with the rank value that is specified in the root argument of the MPI procedure.

scatter
An operation in which the root process divides the contents of the transmission area and sends the divided components to all MPI processes in the communicator. scatter is the inverse operation of gather.

size
The number of MPI processes that constitute either a group or a communicator. In the case of a communicator, size indicates the size of the group that represents the range of the communicator. In the case of an intercommunicator, however, a size for the whole communicator does not exist; instead, only the sizes of a local group and a remote group exist.

standard mode
A send request that can be issued regardless of whether the corresponding receive request has been issued. In a standard-mode send, whether the contents of the send area are transferred through the system buffer or directly from the send area to the receive area depends on the implementation.
Regardless of whether a send occurs directly or indirectly, MPI defines a send as complete when all contents of the send area have been transferred to either the system buffer or the receive area.
for details, see Section 2.4.

status (MPI_Status)
An object consisting of the source of a received communication message, a message tag, and an error status.

synchronous mode
A synchronous-mode send request can be issued regardless of whether the corresponding receive request has been issued. MPI defines that the synchronous-mode send is complete when a corresponding receive request has been issued.

tag
Optional information used to identify a communication message. Also called a message tag, a tag is attached to a communication message in a transmission operation. In a receive operation, it is used to identify a communication message.

topology
A special attribute that is assigned to a communicator. It is a naming mechanism for MPI processes and can facilitate mapping of MPI processes to hardware. Two types of topologies exist Cartesian topology and graph topology.

unpacking data
Restoring compressed data in the buffer to its original state.

upper bound of a data type
The maximum displacement within a data map. The upper bound of a data type must be defined by taking memory boundary alignment into consideration.

VIEW
File access pattern for MPI-IO expressed by any basic MPI datatype or derived MPI datatype. This may be specified by a user.

window
A registered memory region accessible by the one-sided communication operations. Window must be registered by MPI_WIN_CREATE procedure.

Contents Previous Chapter Index