Allocation optimization for reducing resource fragmentation in Angara high-speed interconnect

This paper considers a high-speed interconnect with a multidimensional topology. The paper is devoted to the optimization of fragmentation resulting from sequential allocation of compute nodes in a supercomputer provided that network traffic from different user’s tasks should not overlap. This paper proposes a method for searching nodes with an evaluation of the fragmentation of the system based on ideas from multidimensional container loading problem. For such an evaluation, the concept of rectangles is introduced, which can be inscribed into the system after placing the next user task. Each set of nodes that is suitable for placing the task is evaluated by the proposed function taking into account the size and the number of found rectangles of maximum size. The proposed method was evaluated using computer system model. A set of different computer systems with three-dimensional and four-dimensional topologies was considered. The minimum system size is 32 compute nodes and the maximum is 144. A synthetic queue of tasks is set for each system. The parameters of the synthetic queues are close to a real ones. The average utilization of the resources of the computer system and the average waiting time for the tasks in the queue is chosen as a method quality criterion. The study showed that the increase of the resources utilization for the proposed method averaged 11 % compared to the base method, and the average time spent in queue is reduced by 45, 3 %.