Non-Blocking Message Passing Operations

• In blocking protocols, the overhead of guaranteeing semantic correctness was paid in the form of idling (non-buffered) or buffer management (buffered).
• It is possible to require the programmer
  • to ensure semantic correctness,
  • to provide a fast send/receive operation that incurs little overhead.
• This class of non-blocking protocols returns from the send or receive operation before it is semantically safe to do so.
• Consequently, the user must be careful not to alter data that may be potentially participating in communication.

• Non-blocking operations are generally accompanied by a check-status operation,
• which indicates whether the semantics of a previously initiated transfer may be violated or not.
• Upon return from a non-blocking operation, the process is free to perform any computation that does not depend upon the completion of the operation.
• Later in the program, the process can check whether or not the non-blocking operation has completed,
• and, if necessary, wait for its completion.
• Non-blocking operations can be buffered or non-buffered.

• In the non-buffered case, a process wishing to send data to another simply posts a pending message and returns to the user program.
• The program can then do other useful work.
• At some point in the future, when the corresponding receive is posted, the communication operation is initiated.
• When this operation is completed, the check-status operation indicates that it is safe to touch this data.
• This transfer is indicated in Fig. 4.3Left.
• The benefits of non-blocking operations are further enhanced by the presence of dedicated communication hardware.
• In this case, the communication overhead can be almost entirely masked by non-blocking operations.
• However, the data being received is unsafe for the duration of the receive operation.
• This is illustrated in Fig. 4.3Right.

Figure 4.3: Non-blocking non-buffered send and receive operations Left: in absence of communication hardware; Right: in presence of communication hardware.

• Comparing Figures 4.3Left and 4.1a, it is easy to see that the idling time when the process is waiting for the corresponding receive in a blocking operation can now be utilized for computation (provided it does not update the data being sent).
• This removes the major bottleneck associated with the former at the expense of some program restructuring.

• Typical message-passing libraries such as Message Passing Interface (MPI) and Parallel Virtual Machine (PVM) implement both blocking and non-blocking operations.
• Blocking operations facilitate safe and easier programming.
• Non-blocking operations are useful for performance optimization by masking communication overhead.
• One must, however, be careful using non-blocking protocols since errors can result from unsafe access to data that is in the process of being communicated.

Figure 4.4: Space of possible protocols for send and receive operations.