1. The hybird approach to programming clusters of multi-core architetures

2. High-performance computing  Cluster computing In order to gain high performance many computers are networked together to work on computation heavy problems as a whole  Multi-core computers Since the max single core computer speeds began to level out, multi-core computers became the popular solution

3. High-performance computing  Multi-core clusters With multi-core computers being popular, nodes in clusters began to make use of this style Having multi-core node would seem to make more powerful nodes and save lot of power and space Programming these multi-core nodes could possibly be different

4. Programming  Cluster computers The standard model for programing clusters is MPI MPI is the message passing interface that is used mostly with distributed memory  Multi-core computers Multi-core computers have shared memory An efficient way to program on shared memory would be the use of OpenMP

5. Multi-core Clusters  When programming for multi-core clusters MPI treats cores on different nodes and the same node in a similar manner  MPI completely ignore the fact that core on a single node work on shared memory  This causes problems because duplicate data is on on a single node with multiple cores

6. MPI  MPI uses communicator objects which connect groups of processes in the MPI session.  MPI supports point-to-point communication between two specific processes.  Collective functions are used to communicate all processes in a process group

7. Problems with MPI  Collectives Collectives take arguments that are arrays of size equal to the number of processes in the communicator ○ Example of collectives are MPI_Gatherv, MPI_Scatterv, and MPI_Alltoallv In the case of MPI_Alltoallv, it takes arguments for both sends and receives which add up to 4 MB on each process if there is a million processes

8. Solution  The combination of OpenMP and MPI is a worthy solution  MPI can handle the communication between nodes on distributed memory  OpenMP can handle communication within a single node on shared memory

9. Implementation  Since OpenMP is not an all or nothing model, it can be injected into certain parts of the program  One can identify the most time consuming loops and place OpenMP directives on them  One can also place directives on loops across undistributed dimensions.

10. Hybrid masteronly  This model is when only one MPI process is used per node  All communication within the node is done by OpenMP  Problem The idling of other threads in the node while communication my MPI is taking place MPI bandwidth not always fully used with a single communicating thread

11. Hybrid with overlap  Using this method is a way to avoid idling computed threads during communication  The communication is split into to one or more OpenMP threads to handle communication in parallel with useful calculation

12. Conclusion  Taking advantage of the hybrid approach has advantages over pure MPI in some cases  This is not valid for all cases, some cases will have reverse effects  If programmed well, applications can be very scalable and have great performance increase