201137776 六、發明說明： 【發明所屬之技術領域】 本發明大體上係關於綠色計算，且更特定言之，本發明 旨在在一計算中心中提供一動態工作量分配，其最佳化總 能量消耗。 【先前技術】 資料中心活動愈來愈多地自排程批次作業之處理工作量 移至遵循商業要求之動態「按需」（on demand)作業執行。 此動態工作量比排程工作量更不可預測，因此，在對資料 中心之後端基礎結構進行尺寸設定以保證預定義效能目標 方面存在一問題。通常，舉例而言，在與被提供作業執行 月艮務之消費者的服務等級協議（Service Level Agreement， SLA)中定義效能目標。若後端尺寸過大，則其可能會導致 計算能力浪費且因此導致能量浪費。若後端太小，則可能 會違反SLA，此情形暗示商業機會丟失。 資料中心中之作業執行的能量消耗可視諸如機器被使用 之持續時間、所使用機器^類型及空氣冷卻需求的因素而 定。機器被負載得愈多，則給定公司基礎結構上能量消耗 增加得愈多。 為了找到如何才能最佳化工作量以控制電力消耗的辦 法，吾人可參考2009年1月公開的來自Microsoft之Albert Greenberg、James Hamilton ' David A. Maltz、Parveen Patel的名為「The Cost of a Cloud: Research Problems in Data Center Networks」之論文之段落2.3「Power」，該論 151217.doc 201137776 文係在 ACM STGCOMM Computer Communication Review 之第1號第39卷中，且可在以下Web位址得到：201137776 VI. INSTRUCTIONS OF THE INVENTION: TECHNICAL FIELD OF THE INVENTION The present invention relates generally to green computing, and more particularly, to providing a dynamic workload distribution in a computing center that optimizes total energy Consumption. [Prior Art] Data center activities are increasingly shifting the processing workload of self-scheduled batch operations to dynamic "on demand" operations that follow commercial requirements. This dynamic workload is more unpredictable than the scheduled workload, so there is a problem in sizing the back-end infrastructure of the data center to ensure a predefined performance goal. Typically, for example, a performance goal is defined in a Service Level Agreement (SLA) of a consumer who is performing a job execution. If the back end size is too large, it may result in wasted computing power and thus waste of energy. If the backend is too small, it may violate the SLA, which implies a loss of business opportunity. The energy consumption of the job performed in the data center can be determined by factors such as the duration of use of the machine, the type of machine used, and the air cooling requirements. The more the machine is loaded, the more energy consumption increases for a given company infrastructure. In order to find out how to optimize the workload to control the power consumption, we can refer to the publication of "The Cost of a Cloud" from Microsoft's Albert Greenberg, James Hamilton 'David A. Maltz and Parveen Patel in January 2009. : Research Problems in Data Center Networks, Paragraph 2.3, "Power", 151217.doc 201137776 is available in ACM STGCOMM Computer Communication Review No. 1, Volume 39, and is available at the following Web address:
Http://ccr.sigcomm.org/online/files/p68-v39nl 〇-greenberg.pdf 此先前技術之論文含有典型資料中心中成本之分析，且 提供關於如何使用雲端架構來減少成本的一些提示。該論 文對成本進行分類，攤銷成本被分攤成45%用於伺服器 (CPU、記憶體、儲存系統）、25%用於基礎結構（配電及冷 卻）、15%用於功耗（power draw)(電力公用事業成本 (electrical utility cost))及15%用於網路（鏈路、轉接設備） 成本。若吾人想要最佳化為實體或虛擬之現有公司資料中 心基礎結構的使用，則需要減少用於功耗之成本（其表示 成本之顯著部分（即使不為最大部分））。該論文之2.3段落 建議，使用由Green Grid Association在以下Web位址定義 為比率（總設施電力）/(IT設備電力）之電力使用效率（PUE) 量度來追縱電力成本：Http://www.thegreengrid.org。亦提 議，減少機器上之PUE可對基礎結構成本有影響。降低每 一伺服器上之功耗會減少資料中心中之較熱點且減少冷卻 成本。 因此，需要最佳化在公司之資料中心上所分派之工作量 的分配，以便節省電力，同時維持針對消費者之良好服 務。 【發明内容】 因此，本發明之一目標係提供一種用以藉由控制供執行 151217.doc 201137776 作業之每一機器上之電力消耗在—資料中心中分 作量的方法及系統。 菜工 根據技術方案i，藉由一種指派—處理任務至一電 組中之-電腦的方法來達成該目標，該方法 驟： v 量測該電腦群組中之每-電腦之瞬時溫度，且選擇該溫 度不超過一給定溫度臨限值之電腦； z恤 判定該群組中之每一選定電腦之瞬時能量消耗； 判定每一該電腦之瞬時處理吞吐量； 計异每-該電腦之—瞬時效率指數，其表示考量針對节 電腦所判定之該瞬時能量消耗及該處理吞吐量的該機器Z 效率； ’-·工判疋成目前具有最高瞬 指派該任務至該等選定電腦中 時效率指數之任一者。 根據技術方案2,亦藉由任一技術方案k方法來達成該 目標，其中該效率指數對於該群組内之不同電腦子集係不 同的，且其中該計算步驟進—步考量適用於每—各別電腦 所屬之該子集的該瞬時能量消耗及該瞬時處理吞吐量；且 指派該任務之該步驟應用於該電腦子集申具有最高效率 指數之任-者’且在該子集中應用於該電腦中具有最低溫 度之任一者。 根據技術方案3，亦藉由姑讲 力稭由技術方案1或2之方法來達成該 目標，該方法進一步包冬这路亡呼欲一 + 乂 .右所有該等輯時溫度超過該給 定溫度臨限值，則請求葬山太兮f _被么 月木猎由在孩電腦群組之位置外部的一 151217.doc -6 - 201137776Http://ccr.sigcomm.org/online/files/p68-v39nl 〇-greenberg.pdf This prior art paper contains an analysis of costs in a typical data center and provides some tips on how to use the cloud architecture to reduce costs. The paper classifies costs, amortization costs are divided into 45% for servers (CPU, memory, storage systems), 25% for infrastructure (distribution and cooling), 15% for power consumption (power draw ) (electrical utility cost) and 15% for network (link, transit equipment) costs. If we want to optimize the use of an existing corporate data center infrastructure that is physical or virtual, we need to reduce the cost of power consumption (which represents a significant portion of the cost, if not the largest part). The 2.3 paragraph of the paper suggests using the Power Grid Efficiency (PUE) metric defined by the Green Grid Association as the ratio (total facility power) / (IT equipment power) to track the cost of electricity: Http://www .thegreengrid.org. It is also proposed that reducing PUE on the machine can have an impact on infrastructure costs. Reducing power consumption on each server reduces hotspots in the data center and reduces cooling costs. Therefore, there is a need to optimize the distribution of workloads distributed across the company's data centers to conserve power while maintaining good service to consumers. SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide a method and system for controlling the amount of power consumed in each of the machines for performing the operations of 151217.doc 201137776 in the data center. According to the technical solution i, the worker achieves the goal by assigning a processing task to a computer in a power group, the method: v measuring the instantaneous temperature of each computer in the computer group, and Selecting a computer whose temperature does not exceed a given temperature threshold; the z-shirt determines the instantaneous energy consumption of each selected computer in the group; determines the instantaneous processing throughput of each of the computers; - an instantaneous efficiency index, which represents the Z-efficiency of the machine for the instantaneous energy consumption determined by the computer and the processing throughput; '-·Working judgments currently have the highest instantaneous assignment of the task to the selected computers Any of the efficiency indices. According to the technical solution 2, the target is also achieved by any technical solution k method, wherein the efficiency index is different for different computer subsets in the group, and wherein the calculation step is further applied to each of the calculation steps. The instantaneous energy consumption of the subset to which the respective computer belongs and the instantaneous processing throughput; and the step of assigning the task is applied to the subset of the computer that has the highest efficiency index and is applied in the subset The computer has any of the lowest temperatures. According to the technical solution 3, the target is also achieved by the method of the technical solution 1 or 2, and the method further includes the sorrow of the road, and the temperature exceeds the given temperature. Temperature threshold, then request the funeral mountain too f _ by the moon wood hunting by the location outside the child computer group 151217.doc -6 - 201137776
處理服務來處理該任務D 根據技術方案4，亦藉由技術方案1至3中任一項之方法 來達成該目標，其中若所計算之該最高效率指數小於一給 定效率臨限值，則請求藉由在該電腦群組之位置外部的一 處理服務來處理該任務。 根據技術方案5，亦藉由技術方案丨至4中任一項之方法 來達成該目標，其中將每一電腦丨之該效率指數GIJ計算 GI_i cp_i/(((CP_i/CP_Tot)*(FP_T〇t-ITP_T〇t))+PS_i) 其t，cp_i為電腦i之該處理吞吐量，cP_T〇t為該電腦 群組之該處理吞吐量，FP_TGt為該電腦群組之該位置的總 瞬時能量消耗，ITP_Tot為該電腦群組之該瞬時能量消 耗，且PS_1為—個電腦之該瞬時能量消耗。 根據技術方案6，亦藉由技術方案1至5中任一項之方法 來達成該目標’該方法進一步包含：計算每一電腦之表示 該機器上之節省的一瞬時節省指數，該節省指數為每一電 腦之該效率指數對總能量成本之比率；且 執行指派任務之該步驟至料選定電腦巾、_定成目前 具有最尚節省指數之任一者。 一項之方法 數SIJ定義 根據技術方案7，亦藉由技術方案丨至6年 來達成該目標，其中將每一電腦I之該節^ 為： SI」GIU/Ec)，其中，GI—i為電腦】之該效率指數 且Ec為系統所處之位置中的該總能量成本。 151217.doc 5 201137776 根據技術方案8，卡拉丄 亦猎由一種電腦程式產品來 標，該電腦程式產口力人个逆取透曰 . 碼指令，料程式設計 57φ,^ ^ 上執仃該程式時執行根據技術方案1 至7中任一項之方法之步驟。 根據技術方案9,亦藉 稭由用於私派一處理任務至一電腦 群組中之一電腦及用 1目#，Μ ^ 通信構件的系統來達成 5玄目標，該系統包含： _ 一分配管理器，政絲 h ^ . /、、，·工調適以自一排程器接收該待指派任 務之一先決條件執行環境； -熱/電力管理器，其用认丨r + 用於收集瞬時溫度及能量消耗且用於 傳輸此貧訊至該分配管理器； -該分配管理器經調適 ,,.„ iS 選擇e亥群組中具有該任務先決條 件執仃ί哀境且該溫声 又超k 一、疋溫度臨限值之該等電 腦； -5亥分配官理器經調適 率才a數’其考量針對該 處理呑吐量； 以计算表示每一電腦之效率的一效 電腦所判定之電腦瞬時能量消耗及 -該分配管理器經調適 等選定電腦中經判定成 者。 以使用通信構件來指派該任務至該 目月ϋ具有最两瞬時效率指數之任— 根據技術方案10，亦获出姑，片士在η 丌糟由技術方案9之系統來達成Processing the service to process the task D. According to the technical solution 4, the method is also achieved by the method of any one of the technical solutions 1 to 3, wherein if the calculated highest efficiency index is less than a given efficiency threshold, then The request is processed by a processing service external to the location of the computer group. According to the technical solution 5, the method is also achieved by the method of any one of the technical solutions to 4, wherein the efficiency index GIJ of each computer is calculated as GI_i cp_i/((CP_i/CP_Tot)*(FP_T〇 t-ITP_T〇t))+PS_i) t, cp_i is the processing throughput of computer i, cP_T〇t is the processing throughput of the computer group, and FP_TGt is the total instantaneous energy of the location of the computer group Consumption, ITP_Tot is the instantaneous energy consumption of the computer group, and PS_1 is the instantaneous energy consumption of the computer. According to the sixth aspect, the method is also achieved by the method of any one of the technical solutions 1 to 5. The method further comprises: calculating an instantaneous saving index of each computer representing a saving on the machine, the saving index is The ratio of the efficiency index of each computer to the total energy cost; and the step of performing the assignment task to the selected computer towel, _ is currently the one with the most significant savings index. The method number of the SIJ is defined according to the technical solution 7, and the technical solution is also reached to 6 years to achieve the goal, wherein the section of each computer I is: SI "GIU/Ec), wherein GI-i is The efficiency index of the computer and Ec is the total energy cost in the location of the system. 151217.doc 5 201137776 According to technical solution 8, Karaoke is also hunted by a computer program product. The computer program produces a counter-product. The code command, program design 57φ, ^ ^ on the program The steps of the method according to any one of claims 1 to 7 are performed. According to the technical solution 9, the system also uses a system for processing a task to a computer group and a system of 1 mesh #, Μ ^ communication component to achieve 5 玄 targets, the system includes: _ an allocation Manager, political wire h ^ . /,,, · work adaptation to receive one of the tasks to be assigned from a scheduler preconditions execution environment; - thermal / power manager, which uses the reference r + for collecting transients Temperature and energy consumption and used to transmit this information to the distribution manager; - the allocation manager is adapted, .... iS selects the e-hay group with the task prerequisites 仃 哀 且 and the warm sound again Super k, 疋 temperature threshold of these computers; -5 hai allocation of the official processor after the adaptation rate is a number 'the consideration for the processing vomiting; to calculate the efficiency of each computer to determine the efficiency of each computer Instantaneous energy consumption of the computer and - the distribution manager is determined by the selected computer, etc.. The communication component is used to assign the task to the month with the highest instantaneous efficiency index - according to the technical solution 10 Get aunt, the film is in the η System 9 to achieve
/ 一，，小/ S尔观*不逆取琢E 私二系統包3清求構件，該請求構件係用於：若所有駕 腦之該溫度皆超過'哈定溫度臨限值，或若該最高計算免 率指數小於-給定效率臨限值，則經由通信構件來請求襄 151217.doc 201137776 由在該電腦群组之位置外部的—處理服務來處理該任務。 該解決方案為_種動態工作量分配系統，其最佳化在公 司基，結構内工作量之分配。該最佳化藉由首先在具有較 低能量消耗之機！I上分配工作量且接著藉由使資料中心内 之溫度保持均—來最小化總能量需求，因為避免熱點會降 低空氣冷卻需求。只要處理伺服器之能量消耗或溫度變得 太高（此情形意謂吾人正接近計算能力極限），便使用在公 司邊界外之網料接自自地投送另外卫作量至外部計算^ 供者。此等計算服務之計費通常&「計次付費」 ㈣)，因&，若本端資料中心能夠跟上工作量，則將無額 外成本。 該解決方案提供-齡要㈣最有效率之電腦且使用較 低效率之電腦以僅處置峰值條件的機構。或許，該解決方 案亦可考慮使具有最差〇[之電腦斷電或休眠，且僅在最有 效率之電腦將超載時向具有最差GI之電腦供電。 PUE經計算為（總設施電力）/(IT設備電力）。首先，計算 整個資料中心之此指數，而非單一電腦之此指數。本解決 方案考慮到電腦之計算能力，因為_ 、 貝枓中心可具有良好 PUEU、謂多數設施電力係用以向電腦供電)，但該等電腦 可月“提供局計算能力。又’ 一其他資料中心可具有較差 PUE，但來自此資料中心之—些電腦可提供較高計算能 在使用「綠色指數」的情況下，結果H資料中 心可比前一資料中心好。 提供關於⑽最大電腦溫度之臨限值作為參數會允許使 151217.doc 201137776 其隨著時間而變化。金也丨二& 舉例而言，使用者可決定計書彳使 據其能夠在給定時間範圍 x 用其自有太騎能發電廠所產 生之電力進行動態調適的臨限值。 【實施方式】 淋圖1展示本發明之較佳實施例的系統環境。如同先前技 術之解決方案，服務請求者（130)提交其作業。在文件中， 吾人考慮批次作業之提交，然而，本發明之相同解決方案 可應用於其他類型之作業（諸如’互動式任務’或可排程 =行之任何作業）的工作量。對於排程互動式任務，需要 遠端存取機構（諸如，Te細、遠端桌上型電腦，等等）以 允許服務請求者與供執行作業之電腦互動。 所有服務請求者不具有相同數目之作業—些服務 者可具有針對基本服務之請求，其他服務請求者可在峰值 :間期間且以最高優先權提交大量作業。通常，前端伺服 裔（115)與服務請求者建立介面連接，但係取決於如下排程 〇〇 °玄排耘器安裝於負責系統管理功能以在分散式系統上 散佈軟體或收集監視資訊之中央伺服器上，一個功能係藉 由排程電腦集區（125)上之作業以用於其執行的排程器處 置。如同在先前技術中，系統管理係藉由用戶端·伺服器 應用％式執行，伺服器應用程式在中央伺服器上執行且 用戶端應用程式（代理程式12〇)在電腦集區中之每一電腦上 執行。對於各種系統管理功能，伺服器與代理程式一起通 信。排程器使用系統管理伺服器之通信服務來控制分散式 電腦上之作業執行。在圖丨中，其他系統管理功能係在本 1512l7.doc 201137776 發明之範缚外，在較佳實施例之中央祠服器⑽謂的_ 中心管理器（SCM)(115))上僅表示排程器。作為先前技術 之排程器(_的補充，一新組件(動態工作量執行服務分 配管理器⑴0))安裝於咖上。此額外組件能夠考量來自 新獨立計算設備之資料，-組件(熱/電力管理器⑽))能 夠監視該新獨立計算設備上之熱及電力且在請求分配管理 器後隨即計算變數，且將此等結果發送回至分配管理写。 在較佳實施例中’熱/電力管理器設備為需要計算資源及 包括電纜線及探針之特定硬體的獨立設備。 分配管理器與熱/電力管理器通信以獲得藉由熱/電力管 理器收集之熱/電力資料，且將根據以節省電力及減少具 有過熱電腦之資料中^中之不平衡熱為目標的演算法在所 有電腦當中選擇-電腦q在集區中不存在分配管理写可 選擇之任何電腦’則分配管理器向外部計算服務提供者 ⑽)請求新機器。在處理該請求且向該新機器供應所需軟 體之後，分配管理器在該新機器上提交作業。 應注意，作業提交者可為人類服務請求者（130)，或在 ，1中未表不之其他計算***。仍根據較佳實施例，工作 量=管理員⑽）（其為在先前技術之解決方案的情況下 負貝提供作業工作量分派策略及資訊至排程器的管理員） 时本發月之較佳實施例中提供環境相關策略及資訊至排程 器’該排程器將其儲存於與分配管理器（1H))共用之排程資 ；： 庫中。熱/電力官理器將歷史資料以及來自管理員 〇〇5)之資簡存於熱/電力資料存放庫（M5)中。在較佳實/ a,, small / S er view * not reverse 琢 E private two system package 3 clear component, the request component is used: if all the brain temperature exceeds the 'hading temperature threshold, or if The highest calculated exemption index is less than - given the efficiency threshold, the request is processed via the communication component 襄 151217.doc 201137776 is processed by the processing service outside the location of the computer group. The solution is a dynamic workload allocation system that optimizes the allocation of work within the company's base and structure. This optimization is first achieved with a machine with lower energy consumption! The workload is allocated on I and then the total energy demand is minimized by keeping the temperature in the data center uniform—because avoiding hot spots reduces air cooling requirements. As long as the energy consumption or temperature of the processing server becomes too high (in this case, the person is approaching the limit of computing power), the net material outside the company's boundary is used to send another amount of self-delivery to the external computing. By. The billing for these computing services is usually & "pay-as-you-go" (4)), due to &, if the local data center can keep up with the workload, there will be no additional costs. The solution provides an organization that treats peak conditions with the most efficient computers and uses less efficient computers. Perhaps the solution could also consider having the worst computer power outage or hibernation, and powering the computer with the worst GI only when the most efficient computer is overloaded. The PUE is calculated as (total facility power) / (IT equipment power). First, calculate this index for the entire data center, not the index for a single computer. This solution takes into account the computing power of the computer, because _, Bessie Center can have good PUEU, that is, most facilities power is used to supply power to the computer), but these computers can provide "office computing power." The center may have a poor PUE, but some computers from this data center can provide higher calculations in the case of using the "green index". As a result, the H data center can be better than the previous data center. Providing a threshold for (10) maximum computer temperature as a parameter would allow 151217.doc 201137776 to change over time. For example, the user may decide to calculate the threshold for dynamic adaptation of the power generated by his own Taiji Power Plant for a given time frame x. [Embodiment] Figure 1 shows a system environment of a preferred embodiment of the present invention. As with the prior art solution, the service requester (130) submits his or her job. In the document, we consider the submission of batch jobs, however, the same solution of the present invention can be applied to the workload of other types of jobs (such as 'interactive tasks' or any jobs that can be scheduled = line). For scheduled interactive tasks, a remote access mechanism (such as a Te, a remote desktop, etc.) is required to allow the service requester to interact with the computer for performing the job. All service requesters do not have the same number of jobs - some may have requests for basic services, and other service requesters may submit a large number of jobs during peak: and with the highest priority. Usually, the front-end server (115) establishes an interface connection with the service requester, but it depends on the following schedule. The system is installed in the central part of the system management function to distribute software or collect monitoring information on the distributed system. On the server, a function is handled by the scheduler for scheduling operations by scheduling jobs on the computer pool (125). As in the prior art, system management is performed by the client-server application %, the server application is executed on the central server and the client application (agent 12〇) is in each of the computer pools. Executed on the computer. For various system management functions, the server communicates with the agent. The scheduler uses the communication services of the system management server to control job execution on the distributed computer. In the figure, other system management functions are outside the scope of the invention of the 1512l7.doc 201137776, and only the row is represented on the central server (10) of the preferred embodiment (the central manager (SCM) (115)). Program. As a prior art scheduler (with the addition of _, a new component (Dynamic Workload Execution Service Assignment Manager (1) 0)) is installed on the coffee. This additional component can take into account data from the new stand-alone computing device, the component (thermal/power manager (10)) can monitor the heat and power on the new stand-alone computing device and calculate the variables immediately after requesting the allocation manager, and The results are sent back to the allocation management write. In the preferred embodiment, the thermal/power manager device is a stand-alone device that requires computing resources and specific hardware including cables and probes. The distribution manager communicates with the thermal/electricity manager to obtain thermal/electrical data collected by the thermal/power manager and will be based on calculations to save power and reduce unbalanced heat in the overheated computer The method is selected among all computers - the computer q does not have any computer that allocates management writes in the pool, and the distribution manager requests the new machine from the external computing service provider (10). After processing the request and supplying the required software to the new machine, the allocation manager submits the job on the new machine. It should be noted that the job submitter may be a human service requester (130), or other computing system not shown in . Still according to the preferred embodiment, the workload = administrator (10)) (which is the administrator of the workload distribution strategy and information to the scheduler in the case of the prior art solution) In the preferred embodiment, environment-related policies and information are provided to the scheduler 'the scheduler stores it in the scheduler shared with the allocation manager (1H); The heat/electricity officer stores the historical data and the funds from the administrator 〇〇5) in the heat/electricity data repository (M5). In the better
S 151217.doc 201137776 /Ή二理问一工作量環境管理員直接且單獨地控制排程器 刀配官理器及用於熱/電力管理服務之設備。在―皇他可 能實施例中，可經由SCM來控制排程器/分配管理器及用 於熱/電力管理服務之設備。 " 、分配管理器及熱/電力管理器在較佳實施例中 杈佳地貫施為電腦程式’而非實施為硬體形式（例如，硬 體微編碼邏輯）°提供組態參數（臨限值）及輸入（描述先決 條件之作業定義）至此項技術中已知之此等程式。 圖2為根據較佳實施例的用於排程工作流之本發明之她 流程圖。如同先前技術，排㈣將接收對提交給^作業之 請求。排程器委派電腦之選擇至新獨立組件(分配管理 器）。該程序在排程器調用分配管理器（細）時開始以選擇 電腦用於執行具有特定作業^義之作業。作業Μ包含作 業識別符及用於其執行之先決條件。先決條件包括作業系 統、其修補程式等級，及需要事先安裳用於作業執行之直 他軟體。排程器提供作為輸人之作^義至分配管理器，。、 分配管理器首先藉由選擇資料中心中對應於作業定義之要 求且因此具有經安裝用於作業執行之合適先決條件之本端 電腦之子集來執行現有電腦（實體及虛擬）之筛選（210)。如 同先前技術，可藉由與安裝於每-電腦上之代理程式通作 來檢查電腦安裝。 ★可發生如下情況：先前技術之排程器應用基於使用者所 簽訂之合約在適於作業執行之電腦當中選擇「最快」電腦 或「最有效率」電腦的策略。舉例而言，提交作業至此資 J2 151217.doc 201137776 才:中心之消費者可多付費以使執行速度最大化。在必須考 量與某一消費者所簽訂之此合約的狀況下，首先藉由在適 2作業執行之電腦當中識別最快電腦之分配管理器應用此 策略’在此選擇之後藉由分配管理器執行綠色最佳化。若 已選擇之最快電腦因為其超載而過熱或電力過強，則不執 行作業至此電腦。 夂』也熱/電力官理器設備全局地收集⑽)用於資料 中心及資料中心之每一電腦的熱及電力資訊。週期地及/ =月长時發送（215)此貧訊至分配管理器。分配管理器總 ::用自熱/電力管理器所接收之最新熱及電力資訊。用 订作業之每一電腦之分配管理器驗證溫度不超過 ㈣Ό (或針對其所屬之機器類型)所定義之臨 Μ ^臨限值為用於分配管理器組件之組態參數，其 臨：=:例中為程式。分配管理器僅選擇(220)溫度低於 =二之電腦。接著在溫度低於臨限值且最具綠 ㈣在文件中關於圖3之描述來描述如 驟:所使用之演算法:在已經基於節省電 貝科中心中之不平衡埶 (225) 〇 L所選擇的電腦當中選擇本端電腦 至本端電腦時，且當因為任何電腦 未選擇该等電腦(對測試2 (戰而 至外部計算雲端⑽）。發送㈣^」）時’卸載該作業 務提供者。在處理該於灰曰盗之请求至外部計算服 之後，八西…”向該新機器供應所需先決條件 外錢盗上提供作業。藉由SCM處 1512l7.doc -13- S. 201137776 置:端作業執行之追縱以用於對消費者之未來計費。 +右已選擇目標系統電腦（對測試230之回答「是」），則 藉由刀配官理器發送作業以用於在選定電腦上執行（240)。 ^〜田SCM凊求新機器時，其將有可能指定一租用 週d使得一旦内部環境復原至正常狀態，隨即不釋放所 ’、w之機D„而貫情為，使所供應之機器保持被分配一時 間週期。如前文所述，系統管理在外部系統上所需要之資 源之供應，而且追蹤委派作業之計費。 圖3為根據較佳實施例的用於基於溫度及電力節省來選 擇(v驟225)目;^電腦之本發明之詳細流程圖。分配管理器 讀取藉由熱/電力管理器收集之最新溫度及電力資訊。在 車又佳貝施例中，使用藉由熱/電力管理器傳輸之最新資 訊。在其他實施例中，藉由分配管理器在運作中向熱/電 力官理器請求該資訊。應注意，在電腦上所量測之溫度隨 著電力而增加，且因此隨著電腦工作量而增加。使用溫度 監視以在中心内部均一地***工作量，以便避免熱點：藉 由使所有電腦具有幾乎相同之溫度，溫度在資料中心内部 達到平衡，且不過度使用冷卻系統。此外，若超過一電腦 之溫度臨限值’則不發送工作量至該電腦（22〇)。 分配管理器計算已經選擇之電腦中之每一者的綠色指數 (GI)。待選擇之電腦必須具足夠綠色以執行作業。此情形 意謂每一電腦不能超載且不能在資料中心中消耗太多電 力。如稍後在文件中關於圖4之描述所解釋，GI總體上依 據考量電腦及資料中心所消耗之電力之電力消耗以及電腦 151217.doc • 14· 201137776 及貧料中心之計算能力來量測電腦之良好行為。使用藉由 "、、電力S理器收集之電力資訊的分配管理器計算每一電 =GI ’且若最高GI小於特^臨限值（對測試310之回答 疋」）則不忐選擇資料中心之任何電腦用於作業執行 ) 該程序返回至總流程圖中之步驟230以用於請求 在外部雲端計算服務中之電腦。肖㈣範圍群組（例如， 群組A用於自⑴之⑺，且群組6用於自3至7之叫對電腦 行刀頒（320)。選擇（33〇)具有最高GI之電腦群組，且在 j群組中選擇⑽)具有最低溫度之電腦作為目標電腦。接 著°亥私序返回至總流程圖之步驟230。最終將在步驟34〇 中所選擇之目標電腦上執行作業。 如前文所述，熱臨限值及以範圍較佳地係由管理員 (1 〇5)疋義，在較佳實施例中，熱臨限值及α範圍係作為 組癌參數被給予分配管理器程式。 項其他較簡單之實施例將係替換基於目標電腦之指數 值來選擇目標電腦之步驟32〇、33〇、34〇以選擇具有最高 綠色指數之電腦。 圖4展示根據較佳實施例的藉由分配管理器計算之綠色 才曰數之公式。電腦之綠色指數定義此電腦之「綠色程 度」。電腦之指數愈Α ’則在此電腦上之作業執行愈便 且。右針對一個電腦所計算之綠色指數低於一臨限值，則 不選擇該電腦用於下一作業執行。綠色指數之臨限值係作 為組態參數被給予分配管理器組件。在下文中看出，當計 异考量一電腦之電力消耗及其計算能力的綠色指數（GI) 151217.doc 15 201137776 時’該電腦之GI反映該電腦在資料中心中的效率如何。 根據先則技術之建議來監視資料中心中之電力。集中於 電力公用事業成本（功耗）的綠色指數為電力及計算能力變 數之函數，而非量測資料中心之使用效率（pUE=(總設施電 力）/(ιτ設備電力））。在圖4中看出，電腦綠色指數（GI)公式 (400)使用每一電腦之電力、所有電腦之總電力及資料中心 之所有設施之電力作為變數。GI亦考量該電腦之計算能力 及所有電腦之計算能力。 在較佳實施例中，（例如）以每小時千瓦特為單位來量剩 電腦電力，且（例如）以每秒百萬兆次浮點運算為單位來量 測計算能力❶若電腦组態不改變，則電腦之計算能力不= 變化：FLOPS(每秒浮點運算）為藉由電腦在__秒内執行之 洋點運算的數目。此單位考量電腦之硬體及作業系統能 力。週期地量測電腦電力，其隨著電腦之工作量而變化。 電腦所具有之作業執行愈多，則電腦電力增加得愈多。 在圖4之綠色指數公式（4〇〇)的情況下吾人看出， 叫_)為隨著電腦上之計算能力cp」而增：的比率： 又，若電腦之電力PS」降低，則指數較大。綠色指數考量 電腦之計算能力制於在電腦外之其餘資料中心設施之= 力（FP—T。而_Tot)(例如’用於冷卻***之電幻的影继。 該影響係…之IP能力對電腦之總…：率 (CP_i/CP_Tot)成比例。 干 151217.doc -16· 201137776 間而改變：舉例而言，在夜晚 仪充期間之FP—Tot因為溫度較低 而小於在白天期間之FP Tot，且1 Λ 一 且其而要較少能量用於冷卻 系統。 在本發明之一其他實施例中 個電腦上之節省表示為綠色指 的比率： ’使用以下節省指數以將一 數對資料中心之總能量成本 SI_i=GI_i*(i/Ec) 其中： 兆次浮點運算/費用） 兆次浮點運算/每小時千 SI-i=系統1之節省指數（每秒百萬 GI —1-系統！之綠色指數（每秒百 瓦特） ’、4所處之位置中的能量成本（費用/每小時千瓦特） 广指數亦隨著時間而變化，因為能量成本通常隨著時 間而改變。 太t總結綠色指數或節省指數之計算時，應注意，不能在 非撼執I㈣，同時保持與消費者之相同服務等級’除 ^貝料中心之公司決定花費更多電力或改變硬體。 佳只施例及任何其他可能之其他實施例的方法及 官理抵次作紫夕τ Α θ ’、 作置’或任何其他類型之處理任務（諸 可排程執行之互動式任務）。 【圖式簡單說明】 展示本發明之較佳實施例的系統環境； :為根據較佳貫施例的用於排程工作流之本發明之總 151217.doc -17- § 201137776 圖3為根據較佳實施例的用於基於溫度及電力節省來選 擇目標電腦之本發明之詳細流程圖； 圖4展示根據較佳實施例的藉由分配管理器計算之綠色 指數之公式。 【主要元件符號說明】 100 排程器 105 工作量環境管理員 110 動態工作量執行服務分配管理器 115 前端伺服器/服務中心管理器（SCM) 120 代理程式 125 電腦集區 130 服務請求者 140 熱/電力管理器 145 熱/電力資料存放庫 150 外部計算服務提供者 400 電腦綠色指數（GI)公式 151217.doc •18·S 151217.doc 201137776 /Ή二理理1 A workload environment administrator directly and separately controls the scheduler Knife with the official processor and equipment for thermal / power management services. In the "Almighty Possible" embodiment, the scheduler/distribution manager and the equipment for the heat/power management service can be controlled via the SCM. " The Distribution Manager and the Thermal/Power Manager are preferably implemented as computer programs in the preferred embodiment' rather than being implemented in hardware form (eg, hardware micro-coded logic). Limits) and inputs (description of the job describing the prerequisites) to such programs as are known in the art. 2 is a flow chart of the present invention for scheduling workflows in accordance with a preferred embodiment. As in the prior art, row (iv) will receive a request to submit to the ^ job. The scheduler delegates the selection of the computer to the new standalone component (Assignment Manager). The program begins when the scheduler calls the allocation manager (thin) to select the computer to execute the job with a specific job. The job Μ contains the job identifier and the prerequisites for its execution. Prerequisites include the operating system, its patch level, and the software that needs to be used in advance for job execution. The scheduler is provided as an input to the distribution manager. The distribution manager first performs the screening of the existing computer (physical and virtual) by selecting a subset of the data center corresponding to the job definition and thus having the appropriate prerequisites installed for the job execution (210). ). As with the prior art, the computer installation can be checked by working with the agent installed on each computer. ★ It can happen that the prior art scheduler application is based on a contract signed by the user to select the "fastest" computer or "most efficient" computer among the computers suitable for the job execution. For example, submitting the assignment to this account J2 151217.doc 201137776 Only: Consumers in the center can pay more to maximize execution speed. In the case of having to consider the contract with a certain consumer, first apply this strategy by identifying the fastest computer's allocation manager among the computers that are executed in the 2 jobs. 'After this selection, execute by the distribution manager. Green optimization. If the fastest computer selected has been overheated or overpowered due to its overload, the operation will not be performed on this computer.夂 也 also thermal / power official equipment to collect (10) globally for the thermal and electrical information of each computer in the data center and data center. Periodically and / = month long send (215) this poor message to the distribution manager. Distribution Manager Total: The latest heat and power information received by the Thermal/Power Manager. Use the distribution manager of each computer in the subscription to verify that the temperature does not exceed (4) Ό (or for the type of machine it belongs to). The threshold is the configuration parameter used to assign the manager component. : In the example, it is a program. The Distribution Manager only selects (220) computers with temperatures below =2. Then the temperature is below the threshold and the most green (four) is described in the document with respect to Figure 3: The algorithm used: in the imbalance based on the saved electric Becco Center (225) 〇L When the local computer is selected to the local computer, and if the computer is not selected by any computer (for test 2 (war to external computing cloud (10)). Send (four) ^") 'unload the service provider. After processing the request for the pirate pirate to the external computing service, the eight s... to supply the new machine with the required preconditions to provide the job. The SCM is located at 1512l7.doc -13-S. 201137776: The tracking of the execution of the end job is used to charge the future of the consumer. + The right target computer is selected (the answer to test 230 is "Yes"), and the job is sent by the knife to be selected for use in the selection. Execute on the computer (240). ^~ Tian SCM begged for a new machine, it will be possible to specify a rental week d so that once the internal environment is restored to a normal state, then the machine D will not be released, and the supplied machine will remain The system is managed for a period of time. As described above, the system manages the supply of resources needed on the external system and tracks the billing of the delegated operations. Figure 3 is a diagram for selecting based on temperature and power savings, in accordance with a preferred embodiment. (v. 225); a detailed flow chart of the invention of the computer. The distribution manager reads the latest temperature and power information collected by the heat/power manager. In the case of the car, the heat is used. / Latest information on power manager transmission. In other embodiments, the information is requested by the distribution manager to the thermal/electrical power processor during operation. It should be noted that the temperature measured on the computer increases with power. And therefore increase with computer workload. Use temperature monitoring to evenly split the workload within the center to avoid hot spots: by having all computers have nearly the same temperature, the temperature is in the data center The balance is reached and the cooling system is not overused. In addition, if the temperature threshold exceeds one computer, the workload is not sent to the computer (22〇). The distribution manager calculates each of the selected computers. Green Index (GI). The computer to be selected must be green enough to perform the job. This situation means that each computer cannot be overloaded and cannot consume too much power in the data center. As described later in the document with respect to Figure 4 Explain that the GI generally measures the good behavior of the computer based on the power consumption of the electricity consumed by the computer and the data center and the computing power of the computer 151217.doc • 14· 201137776 and the poor material center. Use by ", electricity The distribution manager of the power information collected by the S processor calculates each power = GI 'and if the highest GI is less than the special threshold (the answer to the test 310), then no computer of the data center is selected for the job execution. The program returns to step 230 in the general flow diagram for requesting the computer in the external cloud computing service. Xiao (4) range group (for example, group A is used for (7) (7), and group 6 is used for calling computer from 3 to 7 (320). Select (33〇) computer group with the highest GI Group, and select (10) the computer with the lowest temperature as the target computer in the j group. Next, return to the general flow chart step 230 in the private order. The job will eventually be executed on the target computer selected in step 34〇. As mentioned above, the thermal threshold and the range are preferably defined by the administrator (1 〇 5). In the preferred embodiment, the thermal threshold and the alpha range are assigned to the distribution management as group cancer parameters. Program. Other simpler embodiments would replace steps 32, 33, and 34 of the target computer based on the target computer's index value to select the computer with the highest green index. Figure 4 shows an equation for the number of green talents calculated by the allocation manager in accordance with a preferred embodiment. The green index of the computer defines the "greenness" of this computer. The more the index of the computer is, the easier it is to perform homework on this computer. If the green index calculated for one computer is below a threshold, the computer is not selected for the next job execution. The threshold of the green index is given to the allocation manager component as a configuration parameter. As seen below, when considering the power consumption of a computer and the green index (GI) of its computing power (GI) 151217.doc 15 201137776, the GI of the computer reflects how efficient the computer is in the data center. Monitor the power in the data center according to the recommendations of the prior art. The green index, which focuses on the cost of electricity utilities (power consumption), is a function of the power and computing power variables, rather than the efficiency of the measurement data center (pUE = (total facility power) / (ιτ equipment power)). As seen in Figure 4, the Computer Green Index (GI) formula (400) uses the power of each computer, the total power of all computers, and the power of all facilities in the data center as variables. GI also considers the computing power of the computer and the computing power of all computers. In a preferred embodiment, the computer power is measured, for example, in kilowatts per hour, and the computing power is measured, for example, in units of millions of floating point operations per second. Change, then the computing power of the computer is not = change: FLOPS (floating point per second) is the number of oceanic operations performed by the computer in __ seconds. This unit considers the hardware and operating system capabilities of the computer. The computer power is measured periodically, which varies with the workload of the computer. The more the computer has more homework, the more computer power is increased. In the case of the green index formula (4〇〇) in Figure 4, I see that _) is the ratio that increases with the computing power cp on the computer: Also, if the computer's power PS is reduced, the index Larger. The Green Index considers the computer's computing power in the rest of the data center facilities outside the computer = FP-T. and _Tot (for example, 'The illusion of the illusion used for the cooling system. The impact is the IP capability of... It is proportional to the total...: rate (CP_i/CP_Tot) of the computer. Change 151217.doc -16· 201137776: For example, the FP-Tot during the night charge is lower than during the day because of the lower temperature FP Tot, and 1 其 and less energy for the cooling system. In other embodiments of the invention, the savings on a computer are expressed as the ratio of green fingers: 'Use the following savings index to compare pairs Total energy cost of the data center SI_i=GI_i*(i/Ec) where: Mega floating point operation/cost) Mega floating point operation / hourly SI-i = system 1 savings index (million GI per second - 1-System! Green Index (100 watts per second) ', Energy cost in the location of 4 (cost / kilowatts per hour) The wide index also changes with time, because energy costs usually change over time Too summarize the green index or save the index Attention should be paid to the fact that it is not possible to maintain the same level of service as the consumer, but the company that has the same service level as the 'because of the bedding center decides to spend more power or change the hardware. Good example and any other possible implementations. The method of the example and the principle of the process are as follows: purple τ Α θ θ θ ', set or any other type of processing tasks (interactive tasks that can be scheduled to execute). [Simplified illustration] Show the preferred embodiment of the present invention System Environment of an Embodiment; : A total of 151 217.doc -17- § 201137776 of the present invention for scheduling workflow according to a preferred embodiment. FIG. 3 is for temperature and power saving according to a preferred embodiment. A detailed flow chart of the present invention for selecting a target computer; Figure 4 shows a formula for calculating a green index by an allocation manager according to a preferred embodiment. [Main Element Symbol Description] 100 Scheduler 105 Workload Environment Administrator 110 Dynamic Workload Execution Service Allocation Manager 115 Front End Server/Service Center Manager (SCM) 120 Agent 125 Computer Rack 130 Service Requestor 140 Heat/Power Manager 145 Hot/Electric Data repository 150 external computing service provider computer 400 Green index (GI) formula 151217.doc • 18 ·