Load Unbalance in k-ary n-cube Networks J. Miguel-Alonso 1* , J.A. Gregorio 2* , V. Puente 2* , F. Vallejo 2* and R. Beivide 2* 1 The University of the Basque Country, Department of Computer Architecture and Technology, P.O. Box 649, 20080 San Sebastián, Spain miguel@si.ehu.es 2 University of Cantabria, Computer Architecture Group, ETSIIT, Av. de Los Castros s/n, 39005 Santander, Spain, {jagm,vpuente,fernando,mon}@atc.unican.es Abstract. This paper studies the effect that HOL (Head-of-Line) blocking in the packet injection queue has on the performance of bidirectional k-ary n- cubes, for values of k over a certain threshold (around 20). The HOL blocking causes an unbalanced use of the channels corresponding to the two directions of bidirectional links, which is responsible for a drop in the network throughput and a rise in the network delay. Simulation results show that this anomaly only appears in those rings where most injections are performed (normally, those in the X axis), and that the elimination of the HOL blocking in the injection queue enables the network to sustain peak throughput after saturation. 1 Introduction The performance of the interconnection network of a parallel computer has a great impact in the system’s performance as a whole. K-ary n-cubes, are the most common direct interconnection network topologies, encompassing rings, meshes, and tori. A central element of this kind of networks is the packet router that injects packets from (and delivers packets to) the compute node to which it is connected, and also routes packets coming from other routers which have to be delivered to other nodes. The architecture of the router has a fundamental impact on the execution time of applications running in the parallel machine [4],[7]. Typical design objectives are to keep it simple (to reduce cycle time) while getting as much functionality as possible. Simplicity leads to the use of input transit queues and injection queues with FIFO policy. It is well known, however, that this policy has negative effects, HOL (Head- of-Line) blocking among the most harmful of them. Several works have dealt with ways of reducing this effect on transit queues [5]. However, HOL blocking in the injection, to the best of our knowledge, has not been reported in the literature, and as this paper will prove it also has a negative impact on performance that severely af- fects the scalability of this kind of networks. In fact, HOL at the network interface is one of the key reasons why performance suddenly drops when the network surpasses * This paper has been done with the support of the Ministerio de Ciencia y Tecnología, Spain, under grants TIC2001-0591-C02-01 and TIC2001-0591-C02-02, and also by the Diputación Foral de Gipuzkoa under grant OF-758/2003.