Efficient processor allocation strategies for mesh-connected multicomputers

Abstract Efficient processor allocation and job scheduling algorithms are critical if the full computational power of large-scale multicomputers is to be harnessed effectively. Processor allocation is responsible for selecting the set of processors on which parallel jobs are executed, whereas job scheduling is responsible for determining the order in which the jobs are executed. Many processor allocation strategies have been devised for mesh-connected multicomputers and these can be divided into two main categories: contiguous and non-contiguous. In contiguous allocation, jobs are allocated distinct contiguous processor sub-meshes for the duration of their execution. Such a strategy could lead to high processor fragmentation which degrades system performance in terms of, for example, the turnaround time and system utilisation. In non-contiguous allocation, a job can execute on multiple disjoint smaller sub-meshes rather than waiting until a single sub-mesh of the requested size and shape is available. Although non-contiguous allocation increases message contention inside the network, lifting the contiguity condition can reduce processor fragmentation and increase system utilisation. Processor fragmentation can be of two types: internal and external. The former occurs when more processors are allocated to a job than it requires while the latter occurs when there are free processors enough in number to satisfy another job request, but they are not allocated to it because they are not contiguous. A lot of efforts have been devoted to reducing fragmentation, and a number of contiguous allocation strategies have been devised to recognize complete sub-meshes during allocation. Most of these strategies have been suggested for 2D mesh-connected multicomputers. However, although the 3D mesh has been the underlying network topology for a number of important multicomputers, there has been relatively little activity with regard to designing similar strategies for such a network. The very few contiguous allocation strategies suggested for the 3D mesh achieve complete sub-mesh recognition ability only at the expense of a high allocation overhead (i.e., allocation and de-allocation time). Furthermore, the allocation overhead in the existing contiguous strategies often grows with system size. The main challenge is therefore to devise an efficient contiguous allocation strategy that can exhibit good performance (e.g., a low job turnaround time and high system utilisation) with a low allocation overhead. The first part of the research presents a new contiguous allocation strategy, referred to as Turning Busy List (TBL), for 3D mesh-connected multicomputers. The TBL strategy considers only those available free sub-meshes which border from the left of those already allocated sub-meshes or which have their left boundaries aligned with that of the whole mesh network. Moreover TBL uses an efficient scheme to facilitate the detection of such available sub-meshes while maintaining a low allocation overhead. This is achieved through maintaining a list of allocated sub-meshes in order to efficiently determine the processors that can form an allocation sub-mesh for a new allocation request. The new strategy is able to identify a free sub-mesh of the requested size as long as it exists in the mesh. Results from extensive simulations under various operating loads reveal that TBL manages to deliver competitive performance (i.e., low turnaround times and high system utilisation) with a much lower allocation overhead compared to other well-known existing strategies. Most existing non-contiguous allocation strategies that have been suggested for the mesh suffer from several problems that include internal fragmentation, external fragmentation, and message contention inside the network. Furthermore, the allocation of processors to job requests is not based on free contiguous sub-meshes in these existing strategies. The second part of this research proposes a new non-contiguous allocation strategy, referred to as Greedy Available Busy List (GABL) strategy that eliminates both internal and external fragmentation and alleviates the contention in the network. GABL combines the desirable features of both contiguous and non-contiguous allocation strategies as it adopts the contiguous allocation used in our TBL strategy. Moreover, GABL is flexible enough in that it could be applied to either the 2D or 3D mesh. However, for the sake of the present study, the new non-contiguous allocation strategy is discussed for the 2D mesh and compares its performance against that of well-known non-contiguous allocation strategies suggested for this network. One of the desirable features of GABL is that it can maintain a high degree of contiguity between processors compared to the previous allocation strategies. This, in turn, decreases the number of sub-meshes allocated to a job, and thus decreases message distances, resulting in a low inter-processor communication overhead. The performance analysis here indicates that the new proposed strategy has lower turnaround time than the previous non-contiguous allocation strategies for most considered cases. Moreover, in the presence of high message contention due to heavy network traffic, GABL exhibits superior performance in terms of the turnaround time over the previous contiguous and non-contiguous allocation strategies. Furthermore, GABL exhibits a high system utilisation as it manages to eliminate both internal and external fragmentation. The performance of many allocation strategies including the ones suggested above, has been evaluated under the assumption that job execution times follow an exponential distribution. However, many measurement studies have convincingly demonstrated that the execution times of certain computational applications are best characterized by heavy-tailed job execution times; that is, many jobs have short execution times and comparatively few have very long execution times. Motivated by this observation, the final part of this thesis reviews the performance of several contiguous allocation strategies, including TBL, in the context of heavy-tailed distributions. This research is the first to analyze the performance impact of heavy-tailed job execution times on the allocation strategies suggested for mesh-connected multicomputers. The results show that the performance of the contiguous allocation strategies degrades sharply when the distribution of job execution times is heavy-tailed. Further, adopting an appropriate scheduling strategy, such as Shortest-Service-Demand (SSD) as opposed to First-Come-First-Served (FCFS), can significantly reduce the detrimental effects of heavy-tailed distributions. Finally, while the new contiguous allocation strategy (TBL) is as good as the best competitor of the previous contiguous allocation strategies in terms of job turnaround time and system utilisation, it is substantially more efficient in terms of allocation overhead