Table 1 - uploaded by Ryan Friese
Content may be subject to copyright.
Source publication
As high performance computing systems contin-ually become faster, the operating cost to run these systems has increased. A significant portion of the operating costs can be attributed to the amount of energy required for these systems to operate. To reduce these costs it is important for system administrators to operate these systems in an energy-e...
Citations
... Similarly let AP C be a T × M matrix where APC i j is the average power consumption for a task of type i on a machine of type j. These matrices are frequently used in scheduling algorithms [4,[6][7][8]. ET C and AP C are generally determined empirically. ...
... 3, unless stated otherwise. All 1,100 tasks, 30 task types, 36 machines, and nine machines types are used and are described in [8]; the complete description of the system and output data files from the new algorithm are available in [12]. The hardware used for running the NSGA-II experiments is a 2013 Dell XPS'15 with an Intel i7-4702HQ 2.2 GHz CPU. ...
... Techniques for generating Pareto fronts have been well studied [3,7,8,15,21]. This work achieves huge gains in performance over prior algorithms by exploiting properties specific to this static scheduling problem. ...
The rising costs and demand of electricity for high-performance computing systems pose difficult challenges to system administrators that are trying to simultaneously reduce operating costs and offer state-of-the-art performance. However, system performance and energy consumption are often conflicting objectives. Algorithms are necessary to help system administrators gain insight into this energy/performance trade-off. Through the use of intelligent resource allocation techniques, system administrators can examine this trade-off space to quantify how much a given performance level will cost in electricity, or see what kind of performance can be expected when given an energy budget. A novel algorithm is presented that efficiently computes tight lower bounds and high quality solutions for energy and makespan. These solutions are used to bound the Pareto front to easily trade-off energy and performance. These new algorithms are shown to be highly scalable in terms of solution quality and computation time compared to existing algorithms.
... Techniques for generating Pareto fronts have been well studied (e.g., [4], [9], [10], [14], [20]). The LP-based approach in this paper achieves huge gains in run time and solution quality over prior methods by exploiting properties that are common to static scheduling problems. ...
... NSGA-II based approaches to find the energy and makespan Pareto front are in [4], [10] without the use of task and machine types. Other algorithms exist in the literature that may perform differently than NSGA-II such as the strength pareto evolutionary algorithm (SPEA2) algorithm [35]. ...
Resource management for large-scale high performance computing systems pose difficult challenges to system administrators. The extreme scale of these modern systems require task scheduling algorithms that are capable of handling at least millions of tasks and thousands of machines. These large computing systems consume vast amounts of electricity leading to high operating costs. System administrators try to simultaneously reduce operating costs and offer state-of-the-art performance; however, these are often conflicting objectives.
Highly scalable algorithms are necessary to schedule tasks efficiently and to help system administrators gain insight into energy/performance trade-offs of the system. System administrators can examine this trade-off space to quantify how much a difference in the performance level will cost in electricity, or analyze how much performance can be expected within an energy budget. In this study, we design a novel linear programming based resource allocation algorithm for a heterogeneous computing system to efficiently compute high quality solutions for simultaneously minimizing energy and makespan. These solutions are used to bound the Pareto front to easily trade-off energy and performance. The new algorithms are highly scalable in both solution quality and computation time compared to existing algorithms, especially as the problem size increases.
... Additionally, this research incorporates the dynamic voltage and frequency scaling (DVFS) property of the processor, and runs all the tests across a range of different voltages and frequencies. The results from this research can be used to provide highly accurate execution time and energy consumption information for use in the area of resource allocation in high performance computing systems, where application tasks are colocated on the cores of multi-core processors [FrB13], [FrK13], [OxP13]. ...
In this study, we analyze interference trends when co-running multiple applications possessing varying degrees of memory intensity on multi-core processors. We conduct tests with PARSEC benchmark applications and explore energy consumption, execution times, and main memory accesses when interfering applications share last-level cache. We also explore how co-running applications are impacted when the processor frequency is modified using dynamic voltage and frequency scaling (DVFS). A portable and lightweight testing framework is presented and results are shown for experiments conducted on an Intel i7 quad-core system. It is shown that the degree of degradation due to co-location interference on execution time is highly dependent on the types and number of applications co-located on cores that share the last-level cache.
... Similarly let AP C be a T × M matrix where APC ij is the average power consumption for a task of type i on a machine of type j. These matrices are frequently used in scheduling algorithms [1], [5], [6], [12]. ET C and AP C are generally determined empirically based on prior task execution times. ...
With the advent of energy-aware scheduling algorithms, it is now possible to find solutions that trade-off performance for decreased energy usage. There are now efficient algorithms to find high quality Pareto fronts that can be used to select the desired balance between make span and energy consumption. One drawback of this approach is that it still requires a system administrator to select the desired operating point. In this paper, a market-oriented technique for scheduling is presented where the high performance computing system administrator is trying to maximize the return on investment. A model is developed where the users pay a given price to have a bag-of-tasks processed. The cost to the system administrator for processing this bag-of-tasks is strongly related to the energy consumption for executing these tasks. A novel algorithm is designed that efficiently finds the maximum profit resource allocation and tightly bounds the optimal solution. In addition, this algorithm has very desirable runtime and solution quality properties as the number of tasks and machines become large.
... Similarly let APC be a T × M matrix where APC ij is the average power consumption for task type i on machine type j. These matrices are frequently used in scheduling algorithms [4], [6]- [8]. ...
... The system used for these experiments is the same as in Section III, unless stated otherwise. All 1100 tasks, 30 task types, 36 machines, and nine machines types are used as described in [8]; the complete description of the system and output data files from the new algorithm are available in [12]. The hardware used for running the NSGA-II experiments is a 2011 Sager NP7280 with an Intel Core i7 980 @ 3.33 Ghz. ...
The rising costs and demand of electricity for high-performance computing systems pose difficult challenges to system administrators that are trying to simultaneously reduce operating costs and offer state-of-the-art performance. However, system performance and energy consumption are often conflicting objectives. Algorithms are necessary to help system administrators gain insight into this energy/performance trade-off. Through the use of intelligent resource allocation techniques, system administrators can examine this tradeoff space to quantify how much a given performance level will cost in electricity, or see what kind of performance can be expected when given an energy budget. A novel algorithm is presented that efficiently computes tight lower bounds and high quality solutions for energy and makespan. These solutions are used to bound the Pareto front to easily trade-off energy and performance. These new algorithms are shown to be highly scalable in terms of solution quality and computation time compared to existing algorithms.
Trade-offs between energy and performance are important for energy-aware scheduling. Recently, a novel model, called energy-aware profit maximizing scheduling problem (EAPM), which combines energy and makespan into the objective of maximizing the profit per unit of time has been proposed. The user pay a given price to have a bag-of-tasks processed and the objective is to maximize the profit per unit time. In this study, we design a polynomial-time algorithm for the EAPM problem. The execution time of our algorithm is polynomial in the number of task types which is an improvement over the previous algorithm, whose execution time is polynomial in the number of tasks. Moreover, we demonstrate that the approximation ratio of our algorithm is close to 2 for a special case, which may be the best result we can obtain. Experimental results show that our algorithm can produce a feasible solution with better objective value than the previous algorithm.
In this paper, we address the problem of scheduling dynamically-arriving tasks to machines in an oversubscribed heterogeneous computing environment. Each task has a monotonically decreasing utility function associated with it that represents the utility (or value) based on the task's completion time. Our system model is designed based on the environments of interest to the Extreme Scale Systems Center at Oak Ridge National Laboratory. The goal of our scheduler is to maximize the total utility earned from task completions while satisfying an energy constraint. We design an energy-aware heuristic and compare its performance to heuristics from the literature. We also design an energy filtering technique for this environment that is used in conjunction with the heuristics. The filtering technique adapts to the energy remaining in the system and estimates a fair-share of energy that a task's execution can consume. The filtering technique improves the performance of all the heuristics and distributes the consumption of energy throughout the day. Based on our analysis, we recommend the level of filtering to maximize the performance of scheduling techniques in an oversubscribed environment.
The need for greater performance in high performance computing systems combined with rising costs of electricity to power these systems motivates the need for energy-efficient resource management. Driven by the requirements of the Extreme Scale Systems Center at Oak Ridge National Laboratory, we address the problem of scheduling dynamically-arriving tasks to machines in an oversubscribed and energy-constrained heterogeneous distributed computing environment. Our goal is to maximize total “utility” earned by the system, where the utility of a task is defined by a monotonically-decreasing function that represents the value of completing that task at different times. To address this problem, we design four energy-aware resource allocation heuristics and compare their performance to heuristics from the literature. For our given energy-constrained environment, we also design an energy filtering technique that helps some heuristics regulate their energy consumption by allowing tasks to only consume up to an estimated fair-share of energy. Extensive sensitivity analyses of the heuristics in environments with different levels of heterogeneity show that heuristics with the ability to balance both energy consumption and utility exhibit the best performance because they save energy for use by future tasks.
