In [1], AbdelBaky et al. proposed the concept of HPC as a Service, aiming to transform traditional HPC resources into a more convenient and accessible service. They focused on the issues related to elastic provisioning and dynamic scalability, which are concerned in malleable jobs [6]. In this thesis, we focus on the moldable property [6] in most modern parallel applications. Most existing HPC systems usually treat parallel jobs as rigid and require that users specify the amount of processors to use when submitting jobs, even though the parallel jobs have the moldable property. According to treating parallel jobs as rigid or moldable, and whether a job has a deadline or not, we divide most of the research works on parallel job scheduling into the four quadrants as shown in Figure 2.1.
Fig. 2.1. Classification of parallel job scheduling research
Scheduling rigid jobs without deadline, quadrant I, has long been an important research field in high-performance computing and parallel processing. Many research efforts have been spent on developing backfilling job scheduling approaches for resolving the resource fragmentation issues under the FCFS policy [14][15][16][17]. The authors of [15] proposed an EASY-backfilling approach where a job can bypass other jobs with earlier arrival time than it provided that such bypassing would not delay the expected start time of the first job in the
Rigid Moldable
no
deadline I II
deadline III IV
waiting queue. The work in [16] investigates the impact of under-estimation of job execution time on the performance of the EASY-backfilling approach. Various backfilling scheduling approaches are compared and evaluated in [14].
For moldable jobs without deadline, quadrant II, previous research [18] has shown potential performance improvement achieved by adaptive processor allocation. The proposed adaptive processor allocation methods in [18] dynamically determine the number of processors to allocate just before job execution according to the amount of currently available resources and job queue information. In [19][20], Srinivasan et al. proposed a schedule-time aggressive fair-share strategy for moldable jobs, which adopts a profile-based allocation scheme. This strategy thus needs to have the knowledge of job execution time.
Scheduling jobs with deadline is an important issue in many fields, such as real-time systems, and thus there are many research works in the literature discussing related problems [21][22][23][32][37]. Earliest-Deadline-First (EDF) has been one of the most commonly used heuristics for scheduling jobs with deadline. In [24] EDF was applied to maintain timeliness and data freshness while minimizing imposed workload in real-time database research. In [38]
EDF has been applied to scheduling real-time tasks on multicore processors. Another popular heuristic, called Least-Laxity-First (LLF), was used in [23] for scheduling jobs with deadline in distributed systems. For moldable jobs, it is hard to apply the LLF heuristic since the required execution time of a moldable job depends on the amount of processors used and thus is not a fixed value available for calculating the laxity.
The work in [21] presents algorithms with priority strategies for scheduling sequential tasks on a network enabled server (NES) environment. Many scheduling approaches concerning deadline in real-time systems were reviewed and compared in [25], including EDF,
LLF, and Rate-Monotonic algorithms. As cloud computing emerges, some research works, such as [22][26], begin to consider the scheduling issues of tasks with deadline constraints in cloud environments. In [22], an adaptive resource management policy was proposed to handle requests of deadline-bound applications with elastic clouds. The work in [26] addresses the problem of remote scheduling of a periodic and sporadic tasks with deadline constraints in cloud environments. In this thesis, we deal with scheduling online moldable jobs with deadline in HPCaaS environments.
As parallel job scheduling is concerned, most of the related work on deadline-constrained jobs deals with rigid jobs, quadrant III. Few attention has been paid on moldable jobs with deadline. In [27], Saule et al. proposed a moldable EDF method to schedule parallel jobs with deadline using the well-known Earliest Deadline First (EDF) heuristic. However, deadline is not the focus in [27] and the moldable EDF method was actually proposed to deal with optimizing the stretch of moldable jobs without deadline.
In [28], Ligang et al. addressed the problem of dynamically scheduling moldable jobs with QoS demands in multi-clusters. The QoS demand concerned in [28] is soft-deadlines. In addition to soft-deadlines, our work in this thesis also deals with jobs concerning hard-deadlines. In [29], Kwon et al. considered the problem of scheduling independent parallel tasks with individual deadlines so as to maximize the total work performed by the tasks which complete their executions before deadlines. The work deals with static scheduling and the speedup model of parallel jobs is assumed to be linear, while our work is for dynamic scheduling in HPCaaS environments and the speedup of a job with different numbers of processors is calculated using Amdahl’s Law [34].
In our previous work [35], three dynamic scheduling methods were proposed for deadline-constrained moldable jobs, aiming to minimize the deadline-miss rate. In the work of this thesis, we have developed a reservation-based dynamic scheduling approach, which can further improve the overall system performance, and the goal is to maximize HPCaaS providers’
profits instead of minimizing deadline-miss rate.