Data Distribution and sharing during cluster computing using OpenMP and MPICH2

For cluster computing, using MPICH2 to connect and control each node, and using OpenMP to fully parallel the CPU and each CPU core within the node is a low-cost and tens of thousands of Essential Oils solution. (It is estimated that heterogeneous computing requires OpenCL or CUDA participation, but it has never been implemented ). MPI (CH2) is a parallel technology applied to distributed computing facilities.

For cluster computing, using MPICH2 to connect and control each node, and using OpenMP to fully parallel the CPU and each CPU core within the node is a low-cost and tens of thousands of Essential Oils solution. (It is estimated that heterogeneous computing requires OpenCL or CUDA participation, but it has never been implemented ). MPI (CH2) is a parallel technology applied to distributed computing facilities.

For cluster computing, using MPICH2 to connect and control each node, and using OpenMP to fully parallel the CPU and each CPU core within the node is a low-cost and tens of thousands of Essential Oils solution. (It is estimated that heterogeneous computing requires OpenCL or CUDA participation, but it has never been implemented ). MPI (CH2) is a parallel technology applied to distributed computing facilities. It corresponds to OpenMP and is used to share storage facilities. When MPI cooperates with OpenMP, tasks are generally divided into first-level granularity (coarse), and then MPI is sent to the computing node, and OpenMP is used in the node for algorithm-level parallelization.

For Computing-intensive tasks that do not store much data, use MPI to distribute and process data, and merge the data.

If the input is computing-intensive and the input is very large and the output is very small, the input data is prepared on a machine in the cluster and shared and published using the LAN file system linked with optical fiber. MPI distributes data indexes instead of the data itself, and then merges the processing results.

If it is computing-intensive, the input is large, and the output is large, it is considered to be implemented in all distributed file systems.

When considering whether to copy data to each node, we mainly consider the high cost of replication or the high traffic for shared access. Copy the data that requires repeated random access to each node. Data that can be processed after reading the data in sequence can be directly shared.