Link to the University of Pittsburgh Homepage
Link to the University Library System Homepage Link to the Contact Us Form


Zhou, Xiuyi (2012) DYNAMIC THERMAL MANAGEMENT FOR MICROPROCESSORS THROUGH TASK SCHEDULING. Doctoral Dissertation, University of Pittsburgh. (Unpublished)

Updated Version

Download (13MB) | Preview


With continuous IC(Integrated Circuit) technology size scaling, more and more transistors are integrated in a tiny area of the processor. Microprocessors experience unprecedented high power and high temperatures on chip, which can easily violate the thermal constraint. High temperature on the chip, if not controlled, can damage or even burn the chip. There are also emerging technologies which can exacerbate the thermal condition on modern processors. For example, 3D stacking is an IC technology that stacks several die layers together, in order to shorten the communication path between the dies to improve the chip performance. This technology unfortunately increases the power density per unit volumn, and the heat from each layer needs to dissipate vertically through the same heat sink. Another example is chip multi-processor. A chip multi-processor(CMP) integrates two or more independent actual processors (called “cores”), onto a single integrated circuit die. As IC technology nodes continually scale down to 45nm and below, there is significant within-die process variation(PV) in the current and near-future CMPs. Process variation makes the cores in the chip differ in their maximum operable frequency, and the amount of leakage power they consume. This can result in the immense spatial variation of the temperatures of the cores on the same chip, which means the temperatures of some cores can be much higher than other cores.

One of the most commonly used methods to constrain a CPU from overheating is hardware dynamic thermal management(HW DTM), due to the high cost and inefficiency of current mechanical cooling techniques. Dynamic voltage/frequency scaling(DVFS) is such a broad-spectrum dynamic thermal management technique that can be applied to all types of
processors, so we adopt DVFS as the HW DTM method in this thesis to simplify problem discussion. DVFS lowers the CPU power consumption by reducing CPU frequency or voltage
when temperature overshoots, which constrains the temperature at the price of performance loss, in terms of reduced CPU throughput, or longer execution time of the programs. This thesis mainly addresses this problem, with the goal of eliminating unnecessary hardware-level DVFS and improving chip performance.

The methodology of the experiments in this thesis are based on the accurate estimation of power and temperature on the processor. The CPU power usage of different benchmarks
are estimated by reading the performance counters on a real P4 chip, and measuring the activities of different CPU functional units. The jobs are then categorized into powerintensive(hot) ones and power non-intensive(cool) ones. Many combinations of the jobs with mixed power(thermal) characteristics are used to evaluate the effectiveness of the algorithms we propose. When the experiments are conducted on a single-core processor, a compact dynamic thermal model embedded in Linux kernel is used to calculate the CPU temperature. When the experiments are conducted on the CMP with 3D stacked dies, or the CMP affected by significant process variation, a thermal simulation tool well recognized in academia is used.

The contribution of the thesis is that it proposes new software-level task scheduling algorithms to avoid unnecessary hardware-level DVFS. New task scheduling algorithms are proposed not only for the single-core processor, but aslo for the CMP with 3D stacked dies,
and the CMP under process variation. Compared with the state-of-the-art algorithms proposed by other researchers, the new algorithms we propose all show significant performance improvement.

To improve the performance of the single-core processors, which is harmed by the thermal overshoots and the HW DTMs, we propose a heuristic algorithm named ThreshHot, which judiciously schedules hot jobs before cool jobs, to make the future temperature lower. Furthermore, it always makes the temperature stay as close to the threshold as possible while not overshooting.

In the CMPs with 3D stacked dies, three heuristics are proposed and combined as one algorithm. First, the vertically stacked cores are treated as a core stack. The power of jobs is balanced among the core stacks instead of the individual cores. Second, the hot jobs are moved close to the heat sink to expedite heat dissipation. Third, when the thermal emergencies happen, the most power-intensive job in a core stack is penalized in order to lower the temperature quickly.

When CMPs are under significant process variation, each core on the CMP has distinct maximum frequency and leakage power. Maximizing the overall CPU throughput on all the cores is in conflict with satisfying on-chip thermal constraints imposed on each core. A maximum bipartite matching algorithm is used to solve this dilemma, to exploit the maximum performance of the chip.


Social Networking:
Share |


Item Type: University of Pittsburgh ETD
Status: Unpublished
CreatorsEmailPitt UsernameORCID
Zhou, Xiuyixiz44@pitt.eduXIZ44
ETD Committee:
TitleMemberEmail AddressPitt UsernameORCID
Committee ChairYang, Junjuy9@pitt.eduJUY9
Committee MemberJones, Alexakjones@pitt.eduAKJONES
Committee MemberLevitan, Stevenlevitan@pitt.eduLEVITAN
Committee MemberMelhem, Ramimelhem@cs.pitt.eduMELHEM
Committee MemberLi, Guangyonggul6@pitt.eduGUL6
Thesis AdvisorYang, Junjuy9@pitt.eduJUY9
Date: 2 February 2012
Date Type: Publication
Defense Date: 28 September 2010
Approval Date: 2 February 2012
Submission Date: 29 November 2011
Access Restriction: No restriction; Release the ETD for access worldwide immediately.
Number of Pages: 133
Institution: University of Pittsburgh
Schools and Programs: Swanson School of Engineering > Electrical Engineering
Degree: PhD - Doctor of Philosophy
Thesis Type: Doctoral Dissertation
Refereed: Yes
Uncontrolled Keywords: Microprocessor, Thermal management, Task scheduling algorithms, 3D die stacking, Chip Multiprocessor, Process variation
Date Deposited: 02 Feb 2012 16:34
Last Modified: 15 Nov 2016 13:55


Monthly Views for the past 3 years

Plum Analytics

Actions (login required)

View Item View Item