None.
None.
None.
The invention disclosed broadly relates to the field of power and heat management, and more particularly relates to the field of using an energy limit in a sliding window for power and heat management of computing systems while executing various workloads.
Recent technologies are rapidly increasing the power density at all levels. At the computer chip level, for example, there are multi-core chips from major processor manufacturers. At the rack level, blade servers are dominating the market. In data centers, usage of the above technologies rapidly increases the power density of data centers.
Since power consumption is defined as the instantaneous rate at which energy is consumed, using a computing system (or device) of higher power density leads to higher energy consumption. Higher energy leads to higher heat generated by the computing systems and thus leads to higher temperatures to be endured by the systems. High temperature is problematic for computing systems because higher temperature leads to shorter MTBF (mean time between failures) and thus makes systems unreliable. To maintain the appropriate temperature would require stronger cooling systems, which in turn also require higher power consumption.
Thus, power and heat management are critical concerns for large computing systems. Through power and heat management, one goal is to control power consumption to avoid high temperature.
There are multiple technologies to decrease the power consumption of a computing system and thus decrease the generated heat. One such technology is known as clock throttling, where processors are switched to standby modes or deep sleep modes that consume less power. Another technology is Dynamic Voltage and Frequency Scaling (DVFS or DVS). By lowering voltage and frequency (processing speed) together, power consumption is decreased more effectively than other technologies that change frequency only. The current technologies that decrease power consumption cannot avoid decreasing frequency, which will lower the performance.
The current solutions to avoid high temperature in power and heat management fall into two categories: reactive methods and proactive methods. For reactive methods, heat sensors are installed for computing systems and signals are emitted when the temperature is higher than a threshold. Such signals may trigger a power reduction that leads to a reduction of the system's performance. For proactive methods, a power limit, which in general is less than the peak (full capacity) power, is applied such that power consumption at any time is held below the limit.
The drawback of the first method is that the power reduction, which causes a reduction in processing speed, often occurs when the processors are very busy (heavily loaded). Thus, a power cut will inevitably degrade the performance of the workload when the workload demand is high. This is like closing lanes of a highway at rush hour. The drawback of the power limit method is that the method can be too conservative, and will not use the full capacity of a computing system all the time. When the workload is bursty, a workload characteristic exhibited by many web sites and applications, power limit approach degrades the performance during a burst even when the workload is light before and after the burst. Instead of limiting the power consumption all the time, it is natural to think that we can use low power consumption when the workload is light and use high power consumption during the burst to improve the performance of the whole system while we still limit the sum of the power consumption during any period of time.
Briefly, according to an embodiment of the invention, we describe a method for specifying a temporal window of a certain size that slides along time, specifying energy consumption as the accumulated power consumption in a temporal window, and applying a threshold that limits the energy consumption in a sliding window, while the power consumption is able to fluctuate within the window.
According to another embodiment of the present invention, we discuss an electronic device or system that includes: a device that includes: a) an energy limit algorithm that receives as input an energy limit and a power consumption history of the last sliding window ended at the current time and returns power management decisions that can have a plurality of optimization goals with the energy limit as a constraint; and b) a power consumption history that provides the power consumption history of the last sliding window.
The electronic system further includes a process that includes: a) a parameter setter that receives user input and sets energy limit parameters; b) a power controller that adjusts the power level of the processor and/or other hardware; and c) a decision maker that receives an energy limit from the parameter setter, invokes a process from the energy limit algorithm to get power management decisions, and adjusts the power level of the hardware through power controller based on the decisions.
The method can be either implemented as machine executable instructions executed by a programmable information processing system or as hard coded logic in a specialized computing apparatus such as an application-specific integrated circuit (ASIC) or field programming gate array (FPGA).
To describe the foregoing and other exemplary purposes, aspects, and advantages, we use the following detailed description of an exemplary embodiment of the invention with reference to the drawings, in which:
While the invention as claimed can be modified into alternative forms, specific embodiments thereof are shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that the drawings and detailed description thereto are not intended to limit the invention to the particular form disclosed, but on the contrary, the intention is to cover modifications, equivalents and alternatives falling within the scope of the present invention.
We describe a method for specifying a temporal window of a certain size that slides along time, specifying energy consumption as the accumulated power consumption of computing systems in a temporal window, and applying a threshold that limits the energy consumption in a sliding window, while the power consumption is able to fluctuate within the window. Since the energy limit does not impose a limit on instantaneous power consumption, the system processing speed may on average be thereby improving the performance of workloads with bursty patterns of system processing demand.
Referring now to
The system 100 includes inter alia a processor device 102 that in turn comprises a decision maker 110, a workload predictor 160, a power controller 140, and a parameter setter 120. The system also includes a memory device 104 that stores an energy limit algorithm 230 and a power consumption history 150.
The system 100 may receive a Temperature Event 180 such as an over-heating event or temperature measurement from some external temperature monitors. The over-heating event may trigger an emergency program, which may ask Power Controller 140 to shut down the Processor 102.
The system 100 also receives input from users. The input can be user-defined energy limit, which includes an energy limit value and a sliding window size, the information used by parameter setting 120 to derive the energy limit, the workload information used by workload predictor 160 to predict workload, and the workload performance requirements that may assist energy limit algorithm 130 to make power decisions.
The decision maker 110 accepts energy limit settings from parameter setter 120 and temperature event 180 from external sources. The decision maker 110 invokes energy limit algorithm 130 to get power management decisions that satisfy the energy limit and it invokes power controller 140 to adjust the power usage level based on the power management decisions. The power management decision may include the power usage level at present or in the future.
The parameter setter 120 uses input 170 to define the energy limit parameters, including an energy limit value and a sliding window size. If Input 170 already includes the energy limit parameters, the parameter setter 120 passes them to decision maker 110. In other cases, input 170 may include information of the hardware such as power function, which is the power consumption as a function of the operating frequency (speed), the information of the cooling system, and the temperature cap. Based on the information, parameter setter 120 can calculate the energy limit parameters that should avoid over-heating situation for the system.
The energy limit algorithm 130 requires as input an energy limit from the decision maker 110 and the power consumption of the last sliding window ended at the current time from the power consumption history 150. It may also use workload predictor 160 to gain knowledge of the workload and make better decisions for a specific workload. The energy limit algorithm 130 returns power management decisions that can achieve different optimization goals with energy limit as a constraint. In the embodiment in
The power controller 140 uses power management technologies to control power usage; it also measures and saves the power usage level in power consumption history 150. Exemplary power management technologies include DVFS (dynamic voltage and frequency scaling) and clock throttling.
The workload predictor 160 predicts the demand for processing capacity of the incoming workload based on the workload information from input 170 and the workload history kept by it. An exemplary input information for a periodic real-time workload would include the period, the approximate execution time at a specific processing speed, and deadlines for completion. This workload information has well-defined processing requirements such that the workload predictor 160 can predict the demand of processing capacity for the specific workload in the future. An exemplary workload history may include the execution information for the last 10 runs of the specific workload such that the predictor can refine its estimate of the time that would be required to execute the workload at the present processing speed, or at a different processing speed.
Referring to
The decision maker 110 also accepts temperature event 180. If the system 100 is over-heating, then in step 260 the decision maker 110 invokes an emergency process. An exemplary emergency process includes shutting down the processor immediately. Otherwise, in step 230 the decision maker 110 invokes the energy limit algorithm 130. In step 240 the decision maker 110 receives the power management decision that satisfies the given energy limit. Then, in step 250 the decision maker 110 invokes the power controller 140 to adjust the power usage level following the power management decision, if necessary, to enforce the energy limit. In step 270, the decision maker 110 can adjust energy limit if necessary. For example, if over-heating happens with the current energy limit, the decision maker 110 may adjust the energy limit to avoid over-heating in the future.
An embodiment of the present invention is directed toward a method to modulate power consumption of software executing on power-aware systems. The method includes steps of: specifying a temporal window of a fixed size that slides along time; specifying energy consumption as the accumulated power consumption in a temporal window; and applying a threshold that limits the energy consumption in a sliding window, while allowing the instantaneous power consumption to fluctuate within that sliding window. The step of applying a threshold that limits the energy consumption includes choosing a possibly time-dependent control (such as processor speed) such that the energy limit constraint will be satisfied, and enforcing the decision using appropriate power management technologies such as DVFS or clock throttling.
In the method, we set a limit for energy consumption during a sliding window of time. Given a window size w and a limit value Eth, we require that the energy consumption during any temporal window of size w is limited by the value Eth. That is, for any time s, we have ∫ss+wP(t) dt≦Eth, where P(t) is the power consumption at time t. For current computer chip technologies, the power consumption is a function of the frequency (speed). Let power function p(s) define the power consumption at a speed s, we have P(t)=p(s(t)).
Compared with known methods, a method according to an embodiment of the invention is capable of achieving a desired trade-off between power and performance. First, this method is proactive and is able to avoid over-heating situation that requires an immediate power cut, which results in a speed reduction. We can avoid a speed reduction when the processing demand is currently high but was low in previous temporal window. Second, the method as described herein has a better performance compared with the power limit method when the workload is bursty.
Referring to
From
The method provides a new constraint to use in power and heat management problem. Compared with the power limit method, the energy limit method is more flexible since it limits the sum of power consumption in a sliding temporal window instead of limiting the power consumption at any instant.
The energy limit can be used in workload scheduling, workload shaping and workload admission control. In workload scheduling, it can be used in both static and dynamic scheduling. Here, we give an example of a dynamic scheduling algorithm. We assume that no pre-knowledge of the workload, which is represented as multiple tasks, is known beforehand. However, when a new task arrives, its own workload is known.
At a scheduling point, we derive a future speed function that specifies the speed at which the processor will run as a function of time during a given time window. The speed function is designed to achieve other scheduling goals while satisfying the energy limit. The scheduling point can be a calling request from other programs, can be triggered by timer events, or can be a time point when the system workload is changed by task arrival or task termination.
As an example, we study a specific goal to maximizing performance while satisfying an energy limit. That is, the derived speed function is one that finishes the current system workload at the earliest time. In this deliberately simple example, we suppose that the processor can only run at two different speeds: smax and 0. Though we use this specific setting to demonstrate an application of the energy limit method, note that our method can be extended to other scheduling settings such as static scheduling, other optimal goals such as throughput, and other technologies such as DVFS.
To further simplify our discussion, we assume that the power consumption at speed 0 is 0. Let the energy limit be Eth in a sliding window of size w. We can prove that the maximal workload that can be finished in a sliding window is smax t*, where t*=Eth/p(smax). Define the energy budget δ(t) at time t as the unused energy in the sliding window ending at time t. That is, δ(t)=Eth−∫(t−w)+εtP(s)ds. We can use all the energy δ(t) between time t and time t+ε without violating the energy limit (ε is defined to be an infinitesimally small time interval).
After the speed function is derived in step 304, we get the current system workload in step 306. If the system workload is known and can be finished at a time t before the end of the current sliding window, then we do not need to provide processing capacity after t and therefore we set the speed after t to 0 in step 310. Otherwise in step 312, we set a timer to invoke the method in
We give an example in
In the above description, numerous specific details are set forth by way of exemplary embodiments in order to provide a more thorough description of the present invention. It will be apparent, however, to one skilled in the art, that the present invention may be practiced without these specific details. In other instances, well-known features have not been described in detail so as not to obscure the invention. The preferred embodiments of the inventions are described herein in the Detailed Description, Figures and Claims. Unless specifically noted, it is intended that the words and phrases in the specification and claims be given the ordinary and accustomed meaning as understood by those of skill in the applicable art. If any other meaning is intended, the specification will specifically state that a special meaning is being applied to a word or phrase.
Therefore, while there has been described what is presently considered to be the preferred embodiment, it will be understood by those skilled in the art that other modifications can be made within the spirit of the invention. The above descriptions of embodiments are not intended to be exhaustive or limiting in scope. The embodiments, as described, were chosen in order to explain the principles of the invention, show its practical application, and enable those with ordinary skill in the art to understand how to make and use the invention. It should be understood that the invention is not limited to the embodiments described above, but rather should be interpreted within the full meaning and scope of the appended claims.
Number | Name | Date | Kind |
---|---|---|---|
4463432 | Carter, II | Jul 1984 | A |
4620283 | Butt et al. | Oct 1986 | A |
5560022 | Dunstan | Sep 1996 | A |
6021401 | Oravetz et al. | Feb 2000 | A |
6067482 | Shapiro | May 2000 | A |
6453198 | Torgerson | Sep 2002 | B1 |
6813719 | Athas | Nov 2004 | B2 |
6885115 | Hatori et al. | Apr 2005 | B2 |
7155912 | Enis et al. | Jan 2007 | B2 |
7209838 | Wright et al. | Apr 2007 | B1 |
7373222 | Wright et al. | May 2008 | B1 |
7444526 | Felter et al. | Oct 2008 | B2 |
7640760 | Bash et al. | Jan 2010 | B2 |
7715950 | Lasseter et al. | May 2010 | B2 |
7783903 | Piazza | Aug 2010 | B2 |
7840821 | Krishnan et al. | Nov 2010 | B2 |
7920942 | Lasseter et al. | Apr 2011 | B2 |
20030061523 | Stanley | Mar 2003 | A1 |
20030084358 | Bresniker et al. | May 2003 | A1 |
20030097603 | Athas | May 2003 | A1 |
20070050650 | Conroy et al. | Mar 2007 | A1 |
20080288123 | Krishnan et al. | Nov 2008 | A1 |
20080307240 | Dahan et al. | Dec 2008 | A1 |
20090044027 | Piazza | Feb 2009 | A1 |
20090119523 | Totten | May 2009 | A1 |
20100037078 | Gross et al. | Feb 2010 | A1 |
20100315223 | Gross et al. | Dec 2010 | A1 |
20100318828 | Elting et al. | Dec 2010 | A1 |
Number | Date | Country | |
---|---|---|---|
20110046804 A1 | Feb 2011 | US |