Energy Management for Parallel Real-Time Systems

 

This research is based on the fact that the processor frequency has a linear relation with its supply voltage, while the power-consumption is a strictly increasing convex function of the supply voltage. The idea is to lower the supply voltage to some extend such that the reduced processor frequency can still meet the deadlines of tasks. In this case, we will get considerable energy-savings. 

Consider a very simple example: to execute one task with full processor frequency which needs 10 time units, while the deadline is 20 time units. If there is no power management, the task will be executed at full processor frequency with the energy-consumption of: power*time = 1*10 = 10. When power management is employed, the processor frequency can be slowed to 1/2 with supply voltage also reduced to 1/2. The time needed will be 20 which just meet the task's deadline. The power consumption will be 1/4, so the energy consumption will be: power * time = 1/4 * 20 = 5. We get the energy saving of 5 while still meet the task's deadline of 20. 

  • Efficient Energy Management Schemes for Parallel Systems

    First, we considered a task set of { Ti , i = 0..n } that will be executed on parallel systems with a common deadline D, each task has a WCET (worst case execution time) and the dependence of tasks is represented by a directed acyclic graph (DAG). If tasks can finish well before the deadline and there is some static slack time, or some tasks do not use up their WCET and dynamic slack exists, we explore the optimal schemes to schedule the tasks in a parallel systems using the voltage/frequency scaling techniques to manage the power consumption of systems and obtain the maximum energy savings. 

With the observation that different schedule sections may have different parallelism due the dependence between tasks and the schedule sections with higher parallelism should get more slack for energy efficiency, one static power management with parallelism (SPM-P) has been proposed, which can save more energy compared with simple static power management (S-SPM) where the static slack is proportionally allocated over the whole schedule regardless the different parallelisms within the schedule. See [3] for more details.

Considering the dynamic slack that appears online due to the early completion of tasks, a slack sharing algorithm has been proposed to share slack between processing units which could balance the workload among processing units and achieve more energy savings [1]. The result was further extended to AND/OR model applications where only a subset of tasks maybe executed during any executions [2].

Although processors normally consume a significant amount of power in the computing system (especially embedded systems), other power consuming components may also be managed (such as low power memory). Considering the power consumption in a whole system, we have proposed a simple power model that could be easily incorporated in previous voltage/frequency scaling algorithms [4]. Considering static/leakage power in a system, it may not be always energy efficient to operating systems at a lower speed. For clusters of multiple servers, considering the current/expected workload, it could be more energy efficient to power off some servers and let others operate at appropriate speed for maximal energy savings. Based on this observation, energy efficient schemes have been studied for clusters. See [5] for more details.

  • Energy Efficient Fault Tolerance in Parallel Systems

For reliable systems, fault tolerance will be an important concern in addition to energy consumption. Considering the interplay between power consumption and system reliability for different degree of redundancy, we have studied the speed setting problem for an optimistic-TMR scheme, where one unit of a traditional TMR system is slowed down for saving energy provided that other two units do not have an agreement. See [4] for more details.

Moreover, for parallel reliable servers, different redundancy configurations will lead to different level of system reliability as well as different amount of energy consumption. For independent unit-size requests, we analyzed the optimal redundant configurations of the servers to minimize energy consumption for a given reliability goal, or to maximize system reliability for a given energy budget. See [6] for more details.

Future work: We have ignore the communication and task migration overhead in our previous study, which maybe significant, especially in distributed systems. For systems that communicate through wireless channel, the combined power management for computation and communication would be interesting to explore. Considering tasks in an application may have separate deadline, extending our work for applications with tasks having different deadlines (or periodic task) on parallel real-time systems could also be the next step. 

 

[1]  Dakai Zhu, Rami Melhem, and Bruce R. Childers. Scheduling with Dynamic Voltage/Speed Adjustment Using Slack Reclamation in Multi-Processor Real-Time Systems, in IEEE Real-Time Systems Symposium (RTSS'01), London, England, Dec 2001;

[2] Dakai Zhu, Nevine AbouGhazaleh, Daniel Mossť and Rami Melhem. Power Aware Scheduling for AND/OR Graphs in Multi-Processor Real-Time Systems, in International Conference on Parallel Processing (ICPP'02), Vancouver, B.C. Canada, Aug., 2002;

[3] Ramish Mishra, Namrata Rastogi, Dakai Zhu, Daniel Mossť and Rami Melhem, Energy Aware Scheduling for Distributed Real-Time Systems, in International Parallel & Distributed Processing Symposium (IPDPS'03), pages 21 - 29, Nice France, Apr. 2003.

[4] Dakai Zhu, Rami Melhem, Daniel Mossť and E. (Mootaz) Elnozahy. Analysis of an Energy Efficient Optimistic TMR Scheme  in International Conference on Parallel and Distributed Systems (ICPADS), Newport Beach, CA, Jul. 2004;

[5] Ruibin Xu, Dakai Zhu, Cosmin Rusu, Rami Melhem and Daniel Mossť, Energy Efficient Policies for Embedded Clusters, in Proc. of the Conference on Language, Compilers, and Tools for Embedded Systems (LCTES'05), pages 1 - 10, Chicago, Jun. 2005.

[6] Dakai Zhu, Rami Melhem and Daniel Mossť, Energy Efficient Configuration for QoS in Reliable Parallel Servers, in Proc. of the Fifth European Dependable Computing Conference (EDCC'05),pages 122 - 139, Budapest, Hungary, Apr. 2005

Last updated: 09/10/2006 10:55:51 PM