برنامه نویسی انتگرال خطی و تکنیک های اکتشافی برای برنامه ریزی قدرت کم در سطح سیستم در معماری چند پردازنده تحت محدودیت توان عملیاتی
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|25153||2007||29 صفحه PDF||سفارش دهید||محاسبه نشده|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Integration, the VLSI Journal, Volume 40, Issue 3, April 2007, Pages 326–354
The increased complexity and performance requirements of embedded systems has led to the advent of programmable multiprocessor architectures. The paper presents system-level design techniques for minimizing the power consumption of throughput constrained periodic applications (such as multimedia and network processing) that are mapped to multiprocessor architectures. The paper discusses several design techniques that integrate dynamic voltage scaling (DVS) along with loop transformations (pipelining and unrolling), and apply dynamic power management (DPM) as the final design step. The paper presents an optimal mixed-integer linear programming (MILP) formulation along with three modifications that trade-off solution quality for reduced run times. The paper also presents a heuristic technique along with deterministic (LPPWUdetLPPWUdet) and simulated annealing based (LPPWUsaLPPWUsa) optimization strategies for solving the system-level low power design problem. The proposed techniques are evaluated by extensive experimentation with multimedia applications (MPEG-1 decoder, JPEG decoder, MP3 encoder), and synthetic taskgraphs (with 10–40 nodes). The proposed techniques are compared with two existing strategies that apply (i) loop pipelining (FP ) with DPM but no DVS, and (ii) DVS and DPM but no loop transformations (LPS [Luo, Jha, Power conscious joint scheduling of periodic task graphs and aperiodic tasks in distributed real-time embedded systems, Proceedings of the International Conference on Computer Aided Design, November 2000]), respectively. The optimal MILP formulation, LPPWUdetLPPWUdet and LPPWUsaLPPWUsa give an average power reduction of 50.68%50.68%, 48.57%48.57% and 49.23%49.23%, respectively, for multimedia applications when compared against FP. While all our techniques are able to satisfy the performance constraints for JPEG and MPEG-1 decoding applications, the LPS technique fails in many cases. Further, the results produced by our deterministic and simulated annealing based techniques for multimedia benchmarks are on an average within 8.04% and 2.75%, respectively, of the optimum solution produced by the MILP based approach. The experimentation with large synthetic taskgraphs demonstrate that the run times of the heuristic techniques scale very well.
Embedded system applications in multimedia and network processing domains have witnessed an increase in complexity and performance requirements. These two factors coupled with the need for shorter design turn around times and ease of future upgrades have led to the advent of programmable multiprocessor architectures for design of such applications. Examples of commercial multiprocessor architectures aimed at these embedded applications include System-on-Chip (SoC) designs such as Intel IXP series (IXP1200, IXP2400, IXP2800) processors , TI TMS320C8x , Motorola C-port C-5 , and board level implementations such as Sun SX2500 board , Alacorn FastImage 1500 , Synergy microsystem's MantaQX . These architectures are deployed in portable devices (such as DVD players, digital cameras), set top boxes (HDTV), edge and backbone routers. All these implementations have low power consumption as a key design requirement. The portable devices are constrained by battery lifetime. The set top boxes, edge and back bone routers have thermal budgets which in turn translate into power consumption constraints. Consequently, innovative system level low power design techniques are required for the implementation of these embedded applications on multiprocessor architectures. System-level low power optimization is enabled by utilization of the dynamic power management (DPM)  and dynamic voltage scaling (DVS)  (also known as dynamic voltage frequency scaling) capabilities of a processor. Both, DPM and DVS were developed to address the challenge of increased power consumption in CMOS devices. DPM exploits the idle times in application behavior and turns off the power supply to several sub-systems of the processing element. Therefore, DPM reduces the stand-by power or leakage power consumption of the application. For example, the Intel SA1100 StrongARM processor  has two low power DPM modes namely idle and sleep in addition to the normal run mode. In the idle mode the CPU clock is switched off and in the sleep mode the power supply to majority of the chip is turned off. DVS trades-off the active power consumption with performance while executing the non-timing critical portions of the application. In CMOS technologies the active power consumption of a chip can be specified as P=Cswitch×V2×fP=Cswitch×V2×f where CswitchCswitch is the average capacitance switched per clock cycle, V is the operating voltage and f is the frequency of operation. The propagation delay through a CMOS inverter can be approximated as tp=CL/2V(1/kp+1/kn)tp=CL/2V(1/kp+1/kn)  where CLCL is the load capacitance, V is the operating voltage, and kpkp and knkn are the process gain factors of the p- and n-type devices, respectively. The frequency of the device can then be calculated as f=1/tp=2V/CL(kpkn/(kp+kn))=KVf=1/tp=2V/CL(kpkn/(kp+kn))=KV, where K=2/CL(kpkn/(kp+kn))K=2/CL(kpkn/(kp+kn)). Thus, the frequency of a processor in CMOS technology is linearly dependent on the operating voltage. Therefore, reduction in the supply voltage results in a cubic reduction in power consumption at the expense of a linear slow down in the processor speed. DVS exploits this relationship to provide variable operating voltages and corresponding frequencies for the processor. For example, the Intel SA1100 StrongARM processor  supports supply voltages that range from 0.8 to 1.5 V with corresponding operating frequencies ranging from 59 to 206 MHz. DPM and DVS can be currently applied to multiprocessor board level architectures. In SoC based architectures, the emerging globally synchronous locally asynchronous design methodology  with multiple voltage and clock islands would also include DPM and DVS techniques. Embedded system applications in the multimedia and network processing domains demonstrate periodic behavior. Real life implementations of these applications allow a greater time to process each data block than the period (period=1/throughput)(period=1/throughput). Time to process each data block is typically denoted by latency or deadline. For example, the MPEG-1 video decoder is required to support a stream of upto 1.5 Mbps. If we consider that the decoder in each execution generates a macroblock of 8×88×8 (64 values) of 8 bits each, then the throughput requirement can be specified as 2929 macroblocks/s (approximately) or a period of View the MathML source345μs/block. If the stream is intended for human viewing, an initial latency of View the MathML source100μs will not result in detectable degradation. In other words, an MPEG-1 decoder that generates a steady stream of 2929 View the MathML sourcemacroblocks/s(period=345μs/block) with a latency of 445 View the MathML source(100+345)μs/block will satisfy the performance requirements. Performance constraints with deadline greater than period coupled with multiprocessor target architecture enable the application of system level loop transformations such as pipelined scheduling and unrolling. Pipelined scheduling and loop unrolling are two powerful loop transformations for throughput maximization of applications. Pipelined scheduling constructs a steady state that overlaps instances of tasks belonging to different iterations of the original specification. In the steady state there is only one instance of every task in the application. Unrolling as the name suggests transforms the original specification by replicating successive iterations. Thus, the transformed specification includes more than one instance of each task in the original loop. The two transformations can also be applied in an integrated manner where the specification is first unrolled and then scheduled into a pipeline. This paper presents system level design techniques that integrate loop transformations with DPM and DVS to minimize the power consumption of applications mapped to multiprocessor architectures. In the following two sections we present a motivating example for the techniques discussed in this paper.
نتیجه گیری انگلیسی
In this paper, we addressed the problem of system-level low power design of an embedded multiprocessor architecture executing periodic applications such as multimedia and network traffic processing. We presented an MILP formulation that integrated loop transformation techniques namely, pipelining and unrolling with system-level low power techniques, namely, DVS and DPM to minimize the power consumption of the application, subject to performance constraint. We presented several linearization schemes that can be employed to linearize seemingly non-linear equations in the MILP formulation. Although the MILP is of tremendous value due to its ability to generate optimal solutions, its solution time grows exponentially with the number of inputs. Therefore, we also proposed several techniques to counter this problem. We proposed three MILP based techniques that relax one or more constraints to arrive at the solution at a faster rate. We also presented a heuristic technique that can easily be adapted to perform deterministic or simulated annealing based optimization to solve the same problem in polynomial time. We performed extensive experimentation with several multimedia applications, as well as large synthetic task graphs. We compared our formulations with existing techniques such as FP and LPS . The integration of pipelining and loop unrolling gave large reductions in power consumption in comparison to existing techniques. All our techniques gave better results than the existing strategies for system level low power design. In particular our simulated annealing based optimization technique gave the best trade-off between result quality and solution generation time.