For example,Бобцов

METHOD OF PARALLEL-PIPELINE-PARALLEL PACKET COMMUTATION IN MULTIPROCESSOR

Annotation

The problem of increasing the speed and throughput of multiprocessor communication networks using input FIFO-queued switches with an output register matrix is under consideration. A packet switch-ing method is proposed featuring a parallel packet transfer pipeline which makes it possible to load packets from the input buffers to the register matrix with no delay needed to spin until the matrix is empty. The proposed method is shown to provide parallel and concurrent packet processing in the input and output circuits of the packet switch. A structural model of a packet switching unit based on the proposed approach is presented. A packet switching algorithm is formulated based on the representation of the set of packets loaded into the register matrix in the form of a packet consistency graph reflecting the packet set ability of being issued in parallel. A graph vertex weight assignment rule is stated taking into account the idle time packets spend in the register matrix. A maximum total weight clique of the consistency graph is shown to be searched for to pick up a proper subset of packets that can be issued currently which makes it possible to reduce the idle time. A formula is deduced to calculate the average time needed for a packet to be transferred through the register matrix of a switch based on the proposed method. The average packet transfer time versus the number of input/output terminals graphs are investigated and the comparison is made for the parallel-sequential switching method and the proposed approach. The developed method is demonstrated to decrease the average packet transfer time by 41 % for all cases of practical significance.

Keywords

Постоянный URL

Articles in current issue

CONTENT © ITMO UNIVERSITY 1993-2024