Practical PACE for embedded systems

Abstract

In current embedded systems, one of the major concerns is energy conservation. The dynamic voltage-scheduling (DVS) framework, which involves dynamically adjusting the voltage and frequency of the CPU, has become a well studied technique. It has been shown that if a task's computational requirement is only known probabilistically, there is no constant optimal speed for the task and the expected energy consumption is minimized by gradually increasing speed as thetaskprogresses [11].Itispossibletofindtheoptimal speed schedule if we assume continuous speed and a welldefined power function, which are assumptions that do not hold in practice. In this paper, we study the problem from a practical point of view, that is, we study the case of discrete speeds and make no restriction on the form of the power functions. Furthermore, we take into account processor idle power and speed change overhead, which were ignored in previous similar studies. We present a fully polynomial time approximation scheme (FPTAS), which has performance guarantees and usually obtains solutions very close to the optimal solution in practice. Our evaluation shows that our algorithm performs very well and generally obtains solutions within 0.1% of the optimal.

A Theoretical Foundation for Scheduling and Designing Heterogeneous Processors for Interactive Applications

“…This assumption is validated extensively by both analytical models and measurement studies [14,23,36]. In practice, if a slower core consumes more power and thus energy than a fast one, it wont be built.…”

“…Theorem 1 formalizes this property. While prior studies [23,27,36,39] show to use the "slow to fast" property under the maximum latency constraint in different contexts such as DVFS, in contrast, Theorem 1 is the first formal result that applies it to the more general case of any latency norm constraint and with multiple latency norm constraints. Theorem 1.…”

“…This assumption is not restrictive. In practice, a job's service demand cannot be continuous and is typically grouped into a finite number of bins [36]. Let {f1, f2, · · · , fN } and {F0, F1, · · · , FN } be the probability distribution and cumulative distribution of the job's service demand, respectively: fi = Pr(w =wi) and Fi = i j=1 fj, for i = 1, 2, · · · , N .…”

“…Many prior studies mainly focused on single and simple latency constraints, such as maximum latency (deadline) or average latency [23,36]. Motivated by recent work that addresses latency requirements in contexts such as load balancing [2,26], we introduce the Lp norm to generalize latency constraints.…”

“…Unfortunately, the service demand of individual requests in interactive applications is usually unknown and difficult to predict [23]. (2) For unknown service demand, prior work only optimizes for a single simple latency constraint [23,36,39], such as average latency or maximum latency, and is inadequate for two reasons. First, many applications strive for consistency by reducing tail latency (e.g., 95th-and 99th-percentile) or variance [13,17], which average and maximum latency do not model.…”