Distributed Processor Allocation in Large PC Clusters

Current processor allocation techniques for highly parallel systems are based on centralized front-end based algorithms. As a result, the applied strategies are restricted to static allocation, low parallelism and weak fault tolerance. To lift these restrictions we are investigating a distributed approach to the processor allocation problem in large distributed memory machines. A contiguous and a noncontiguous version of a distributed dynamic processor allocation strategy are proposed and studied in this paper. Simulations compare the performance of the proposed strategies with that of well-known centralized algorithms. We also present the results of experiments on a Simens hpcLine Primergy Server with 96 nodes that show distributed allocation is feasible with current technologies. 1. The processor allocation problem Processor allocation involves the selection of a processor partition for a given parallel job, with the goal of maximizing throughput over a stream of many jobs. Because allocation operations have to be fast, allocation techniques used by the majority of commercial parallel machines, as well as the research community, restrict the feasible shapes of partitions to achieve some regularity, which facilitates their management. We call a partitioning scheme structure preserving if it generates partitions that are of the same topological graph family as the entire processor graph (subcube allocation in hypercubes and submesh allocation in meshes). In addition, many systems also require that the allocated processors are constrained to be physically adjacent (contiguous allocation). 2. Distributed processor allocation Several approaches to deal with the processor allocation problem can be found in the literature [3]. In spite of the fact that they apply different policies in the resource management, all the schemes have one in common: the control of allocated resources is done with a global data structure localized mostly in a host machine.The main problems of such centralized management are lack of scalability, the incompatibility with adaptive processor allocation schemes (dynamic allocation), and its weak fault tolerance. Figure 1 shows a global view of the proposed distributed allocation [1] and the distributed Processor Managers involved in the allocation operation. The main differences to the centralized management are (i) the absence of a central data structure with information about the state of all processors, and (ii) the execution of allocation operations directly in the processor mesh in a distributed way, and not in a data structure localized in the host. The host machine is now only responsible for queuing the incoming requests and forwarding them to the processor mesh. The communication between host and mesh is done through a direct channel to a boundary node. This node is called an entry point and due to the distributed environment there is no restriction concerning the number of entry points. Each node in the mesh has a local Processor Manager (PM) responsible for the processor allocation. The PM’s cooperate to solve the allocation problem in a distributed way. Additional allocations Releases Allocation jobs Initial allocations Processor mesh Host machine PM