This chapter presents an extensive evaluation of a new contiguous allocation strategy proposed for 3D mesh multicomputers. The strategy maintains a list of maximal free sub-meshes and gives priority to allocating corner and boundary free sub-meshes. This strategy, which we refer to as Turning Corner-Boundary Free List (TCBFL) strategy, is compared, using extensive simulation experiments, to several existing allocation strategies for 3D meshes. In addition to allocation strategies, two job scheduling schemes, First-Come-First-Served (FCFS) and Shortest-Service-Demand (SSD) are considered in comparing the performance of the allocation strategies. The simulation results show that TCBFL produces average turnaround times and mean system utilization values that are superior to those of the existing allocation strategies. The results also reveal that SSD scheduling is much better than FCFS scheduling. Thus, the scheduling and allocation strategies both have substantial effect on the performance of contiguous allocation strategies in 3D mesh-connected multicomputers.
Top1. Introduction
Mesh interconnection networks have been extensively employed in large-scale multicomputers due to their structural regularity, simplicity, ease of implementation and scalability (Ababneh, 2001; Bani-Mohammad, Ould-Khaoua, Ababneh, & Machenzie, 2006; Bani-Mohammad, Ould-Khaoua, Ababneh, & Machenzie, 2009; Bani-Mohammad & Ababneh, 2013; Foster, 1995; Kumar, Grama, Gupta, & Karypis 2003; Lo, Windisch, Liu, & Nitzberg, 1997; Seo & Kim, 2003; Zhu, 1992).
Effective processor allocation and job scheduling are critical if the full computational power of multicomputers is to be exploited properly (Choo, Yoo, & Youn, 2000; Yoo & Das, 2002). Processor allocation is responsible for determining the set of processors on which a parallel job is executed, whereas job scheduling is responsible for selecting the order in which jobs are selected for execution (Choo et al., 2000; De Rose, Heiss, & Linnert, 2007; Yoo & Das, 2002). The job scheduler selects the next job to execute, and then the processor allocator assigns free processors for executing the selected job (Choo et al., 2000; Windisch, 1997).
In mesh multicomputers, a parallel job is typically allocated a distinct contiguous sub-mesh for the duration of its execution, and the sub-mesh has the same general shape as the mesh multicomputer itself. Most contiguous processor allocation strategies proposed in the literature are for 2D meshes (Ababneh, 2009; Chuang & Tzeng, 1994; Kim & Yoon, 1998; Zhu, 1992). Although the 2D mesh has been used in a number of parallel machines, such as the Cray XE6m (Cray, 2010b), the iWARP (Peterson, Sutton, & Wiley, 1991) and the Touchstone Delta system (Intel, 1991), current multicomputers, such as the K-computer (Riken & Fujitsu, 2014), the Cray XE6 (Cray, 2010a) and the IBM BlueGene/L (Horn et al., 2014) utilize the 3D mesh or the 3D tori (Cray, 2010a; Horn et al., 2014; Riken & Fujitsu, 2014) as interconnection network due to their lower diameter, lower average communication distance, and higher connectivity and network degree (Athas & Seitz, 1988).
Contiguous allocation can lead to high external processor fragmentation, which occurs when the allocation strategy cannot allocate available processors to an incoming job because the available processors are not contiguous or they do not have the requested shape. Non-contiguous allocation can reduce processor fragmentation and improve system performance because a job can execute on multiple smaller sub-meshes. However, contiguous allocation is used in highly-parallel systems such as the IBM BlueGene/L because it isolates jobs from each other, which is useful for security and accounting reasons (Aridor, Domany, Goldshmidt, Moreira, & Shmueli 2005). Contiguous allocation for 3D meshes has been investigated in the literature (Ababneh, 2001; Bani-Mohammad et al., 2006; Bani-Mohammad et al., 2009; Choo et al., 2000; Mao, Chen, & Watson, 2005; Qiao & Ni, 1995).