PVM daemons and tasks can compose and send messages of arbitrary lengths containing typed data. The data can be converted using XDR [12] when passing between hosts with incompatible data formats. Messages are tagged at send time with a user-defined integer code and can be selected for receipt by source address or tag.
The sender of a message does not wait for an acknowledgment from the receiver, but continues as soon as the message has been handed to the network and the message buffer can be safely deleted or reused. Messages are buffered at the receiving end until received. PVM reliably delivers messages, provided the destination exists. Message order from each sender to each receiver in the system is preserved; if one entity sends several messages to another, they will be received in the same order.
Both blocking and nonblocking receive primitives are provided, so a task can wait for a message without (necessarily) consuming processor time by polling for it. Or, it can poll for a message without hanging. A receive with timeout is also provided, which returns after a specified time if no message has arrived.
No acknowledgments are used between sender and receiver.
Messages are reliably delivered and buffered by the system.
If we ignore fault recovery,
then either an application will run to completion or,
if some component goes down, it won't.
In order to provide fault recovery,
a task (
) must be prepared for another task
(
,
from which it wants a message) to crash,
and must be able to take corrective action.
For example,
it might reschedule its request to a different server,
or even start a new server.
From the viewpoint of ,
it doesn't matter specifically when crashes
relative to messages sent from .
While waiting for ,
will receive either a message from or notification
that has crashed.
For the purposes of flow control,
a fully blocking send can easily be built using the semi-synchronous
send primitive.