Information Technology Reference
In-Depth Information
mapping table between application memory space and the conceptual
MSR model. Each memory block that may be referenced in the MSR,
including a dynamic memory block, has an entry in the table. To represent
a pointer, which contains a machine-specii c address, the MSRLT is
searched for the memory block that contains the address. The pointer is
then represented in MSR by an edge to the referenced memory block. The
preinitialized process restores the pointer to the correct address allocated
to the referenced memory block.
16.3.3.2 Communication State Transfer
State transfer of point-to-point communication is supported by HPCM
[26]. Recently, group communication protocols to handle the state transfer
of group communications during a migration and implemented under
MPI have been designed and developed [27].
The difi culty of migration with group communications is that all the
processes in a group need to be synchronized for a new process or an
updating of the group membership information. The basic idea of our
protocols is to divide the synchronization process into two phases: collec-
tive synchronization and point-to-point synchronization. We dei ne a
superstep as the execution block between any two collective operations.
Within a superstep, processes can send messages only through point-to-
point (pt2pt) communication channels. After receiving a migration signal,
collective synchronization protocol brings all the processes in a group to
the same superstep. The pt2pt synchronization protocol wakes up all pro-
cesses waiting for messages through the old communication group, drains
the communication buffer, and preserves the messages in transmission.
After synchronization, all the processes coordinate to spawn a new pro-
cess, create a new group, or update the group information. The preserved
communication state, together with local process states, is transmitted to
the new process for continuous execution.
In these protocols, processes are allowed to execute asynchronously for
better performance. Processes are synchronized only when a process in
the group is migrating. After broadcasting the migration signal, the migrat-
ing process checks the current superstep of each process and determines
the maximum step as the global superstep. The global superstep is sent to
each process and all processes keep execution until all of them reach the
global superstep. After all processes are within the same global superstep,
the migrating process initiates the distributed pt2pt synchronization
protocol. After all processes are synchronized, the local communication
channels are drained and all pending messages are stored as the local
communication state. Then, all processes collectively spawn a new pro-
cess and update their local group information. The migrating process
transmits memory and communication states to the new process, and
i nalizes its communication channels. The new process then resumes
Search WWH ::
Custom Search