Service Deployment System For MPI Applications In Cloud Computing Environments

With development of virtualization and cloud computing, more and more scientificcomputing applications run in the cloud computing resources. The MPI application is akind of typical parallel computing application, which is constituted by multiple componentstogether, and most of the scientific computing application is the MPI application. InInfrastructure-as-a-Service (IaaS) mode, the physical machine (PM) is divided into multipleindependent virtual machines (VM) with using virtualization technology, but theseindependent VMs can’t satisfy the parallel MPI application. It is an important issue how todynamically configure multiple VMs into the computing environment which meets the MPIapplication. It is another important issue how to schedule VMs which are cooperative worktogether.We develop the service deployment system for MPI Applications in cloud computingenvironments based on IaaS, which solves the problem of dynamical configuration parallelcomputing environment. Aiming at the virtual machine collaborative working characteristic,we design an I/O network load balancing virtual machine placement algorithm for oursystem. The algorithm basic principle is as follows: predict the network I/O consumption ofthe application based on monitoring information; search the network I/O load upper limit ofthe PM by using the binary search algorithm; with the network I/O load upper limit of thephysical node, establish the directed weighted graph model of the VM and the PM as thevertex and solve the VM placement matrix based on the model.Based on Linux operating system and Apache platform, adopting Python, HTML andJavaScript, we implement the system. It effectively solves the problems of dynamicalconfiguration parallel computing environment for the MPI application. Compared with thegreedy algorithm, several experiments have been analyzed in the real environmenthighlighting the performance of our algorithm.