Balancing power consumption in multiprocessor systems

Abstract

Actions usually taken to prevent processors from overheating, such as decreasing the frequency or stopping the execution flow, also degrade performance. Multiprocessor systems, however, offer the possibility of moving the task that caused a CPU to overheat away to some other, cooler CPU, so throttling becomes only a last resort taken if all of a system's processors are hot. Additionally, the scheduler can take advantage of the energy characteristics of individual tasks, and distribute hot tasks as well as cool tasks evenly among all CPUs.This work presents a mechanism for determining the energy characteristics of tasks by means of event monitoring counters, and an energy-aware scheduling policy that strives to assign tasks to CPUs in a way that avoids overheating individual CPUs. Our evaluations show that the benefit of avoiding throttling outweighs the overhead of additional task migrations, and that energy-aware scheduling in many cases increases the system's throughput.

Profiling, Prediction, and Capping of Power Consumption in Consolidated Environments

“…Previous work has looked at mechanisms to enforce power budgets both at the server level [17] and at the cluster level [31]. Typically power budgets are enforced either by throttling resource usage [17,31] and/or migrating workloads [18,27]. In our opinion, these techniques often operate with inadequate information about the power consumption behavior of the hosted applications.…”

“…Chase et al considered energyaware resource provisioning in a data center in accordance with negotiated QoS agreements [6]. Merkel et al [18] did 4 This is done using the z-transform. The z-transform of a random variable U is the polynomial Z(U) = a 0 + za 1 + z 2 a 2 + · · · where the coefficient of the i th term represents the probability that the random variable equals i (i.e., U(i)).…”

TProf: An energy profiler for task-parallel programs

“…Prior work on energy accounting between software components uses models to distribute the energy budgets between coarse-grain system-level abstractions, such as virtual machines [27], kernel threads [26], or whole programs [38]. TProf follows the paradigm of EProf [29], which accounts energy at the finer granularity of system calls and applicationlevel library calls [30].…”

Profiling Software for Energy Consumption

“…CPU and Memory Energy: The CPU and memory energy model is a linear model based on hardware performance counters [5]; it is trained using the SPEC CPU 2006 benchmark suite. We select the set of performance counters that yields the best energy model; this optimization also determines the coefficients for the model [2].…”