Strudying About Topology-Aware Collective Operations In The Grid

The Grid is a important information technology that has risen in recent years. It aims at sharing resource with high-performance , cooperating in virtual network environment, and eliminating isolated island of information and resource .In the grid environment, an application is accomplished by cooperation of some processes, but these processes may distribute over resource in different sites. In order to accomplish the application, collective communication operations between processes is absolutely necessary. The quality of a collective communication algorithm significantly affects the performance of application. So, in recent years, the problem of how to building "optimal" communication trees for collective operations has received much attention. In fact, producing ''optimar trees for collective operations mentioned in this paper is searching optimal algorithm for collective operations.Broadcast is the most typical collective communication operation in an MPI library, it directly affects the performance of MPI applications. So, this paper has focused on broadcast. Some existing topology-unaware collective operations algorithms don't make use of the information about network topology. Firstly, these approaches construct a model of network topology and make use of a set of parameters to describe the communication characteristics of a network model, then build the optimal trees based parameter values and their models to realize collective communication. But these approaches have some problems. Such as the models assume that the communication times between all process pairs (a sending process and a receiving process) in the computation are equal. It is a reasonable approximation when the entire computation is performed on a single machine. But when the computation is executed on a Grid environment, it is not reasonable for communication costs over different communication links is differ.And then, some people presented two-level topology-aware collective communication operation algorithm (e.g. MAGPIE). The approach can reduce total communication costs by minimizing the communication across slower channels (e.g. a wide-area network). However, based on the distance between processors, these approaches only limited the depth of network stratification to only two levels : local-area communication and wide-area communication.However, in the practical grid application, because communication link between processes is various, for example, wide-area-nerwork TCP, local-area network TCP, intra-machine TCP and vendor-supplied MPI and so on, communications speed over these links are different. So as to one collective operation, we will stratify the network and produce a multilevel topology tree byraexploiting knowledge concerning the structure of a multilevel network. Then we take advantage of the differences on communication cost at every level in the network multilevel topology tree to minimize communication across slower channels (e.g., a wide-area network). In the way. completion time of the collective communication operation is optimal in the grid environment.In this paper, we gave an example about broadcast, analyzed and compared a topology-unaware broadcast algorithm (e.g. a binomial tree broadcast), a two-level topology-aware broadcast (e.g. MagPIe) and a multilevel topology-aware (e.g.. MPICH-G2). then emphasized on the study of a multilevel topology-aware broadcast in the grid. But the algorithm of MPICH-G2 only considered the special case of process PO as the root of a broadcast, however, when the root of a broadcast is an arbitrary process Pi. what is the algorithm? In the paper, we modified the algorithm's condition that the root process of broadcast is po is, it can satisfy the instance that the root of a broadcast is an arbitrary process p, only with simple change.The theoretic result mentioned before need to be proved by experiment. So. this paper studyone of the most representative grid items------Globus, build an cluster Grid environment byexploiting Globus Toolkit 2.2 and MP...