WO2012093498A1 - Système de gestion de ressources efficace en énergie et procédé pour des processeurs multicœur hétérogènes - Google Patents

Système de gestion de ressources efficace en énergie et procédé pour des processeurs multicœur hétérogènes Download PDF

Info

Publication number
WO2012093498A1
WO2012093498A1 PCT/JP2011/050604 JP2011050604W WO2012093498A1 WO 2012093498 A1 WO2012093498 A1 WO 2012093498A1 JP 2011050604 W JP2011050604 W JP 2011050604W WO 2012093498 A1 WO2012093498 A1 WO 2012093498A1
Authority
WO
WIPO (PCT)
Prior art keywords
task
energy consumption
heterogeneity
energy
mapping
Prior art date
Application number
PCT/JP2011/050604
Other languages
English (en)
Inventor
Wei Sun
Original Assignee
Nec Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nec Corporation filed Critical Nec Corporation
Priority to PCT/JP2011/050604 priority Critical patent/WO2012093498A1/fr
Publication of WO2012093498A1 publication Critical patent/WO2012093498A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5094—Allocation of resources, e.g. of the central processing unit [CPU] where the allocation takes into account power or heat criteria
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the present invention relates to an energy-efficient resource management system of heterogeneous multicore processors, and more specifically to a design of resource management software for heterogeneous multicore processors and some methods of energy-efficiently managing hardware and software resources.
  • heterogeneous multicore architectures are widely used in the various fields of computation because heterogeneous multicore architectures promise greater energy/area efficiency than their homogeneous counterparts. This efficiency can only be realized, however, when the operating system (OS) of the heterogeneous multicore system is able to well make use of the heterogeneities of hardware and software. In other words, the heterogeneous multicore system needs resource management.
  • the core component of any resource management is its scheduler which implements a function of task assignments, mapping tasks to computing resources, to realize some management target(s).
  • a heterogeneous multicore processor is usually modeled as a set of different computing units with shared memory. Each core is a unit which can work in different speeds under different voltage values. A core is also referred to as a processor core. The value of core voltage scales within a scope and the related technology refers to Dynamic Voltage Scaling (DVS).
  • the software running on the processor is multithread programs and each thread is usually modeled as a task.
  • Resource management is a software system which efficiently decides the mapping between tasks and cores. With different knowledge of physical features, the complexity of decision problem is greatly changed. For example, the decision problem can be solved in polynomial time if the execution time and the energy cost can be computed in terms of simple functions derived from physics.
  • Min-min simulations for the minimum completion time scheduling problem.
  • a heuristic named Min-min is shown to outperform the others.
  • MEMC Minimum Energy Minimum Completion time
  • MEME Minimum Energy Minimum Completion time
  • CRME Contention Resolved Minimum Energy
  • Fig. 1 illustrates the procedure of MEME.
  • step 1 11 a fitness value for energy or power consumption is first defined, and, in step 1 12, the minimum fitness value of each task is calculated based on a current partial schedule.
  • step 1 13 the pair ⁇ task, core> with the overall minimum fitness value is selected from the calculated minimum fitness values, and, in step 1 14, the selected pair is mapped.
  • step 1 15 it is determined whether there is an unprocessed task or not. If there is an unprocessed task, the process returns to step 112 otherwise the process of MEME terminates.
  • Fig. 2 illustrates the procedure of MEMC.
  • load balance ratio / and upper threshold L of the load balance ratio are defined.
  • Load balance ratio / is a measure of load balance with a certain sensitiveness.
  • load balance ratio / is actually measured in step 122, and the measured load balance ratio / is compared with the upper threshold L in step 123. If 1>L, a task is mapped to a core with minimum completion time (MCT) in step 124 otherwise the task is mapped to a core with minimum energy (ME) in step 125. After step 124 or 125, it is determined whether there is an unprocessed task or not, in step 126. If there is an unprocessed task, the process returns to step 122 otherwise the process of MEMC terminates.
  • MCT minimum completion time
  • ME minimum energy
  • Fig. 3 illustrates the procedure of CRME.
  • step 131 a fitness value for energy or power consumption is first defined, and, in step 132, the mapping with the minimum fitness value is found for each task.
  • step 133 it is determined whether contention occurs or not. Contention means a situation in which at least two task-core pairs have the minimum fitness value. If there is contention, a task with the fastest fitness degradation is selected from the contended pairs in step 134 and the process goes to step 135. If there is no contention, the process directly goes to step 135. In step 135, the selected task is mapped to the relevant core, and then, it is determined whether there is an unprocessed task or not, in step 136. If there is an unprocessed task, the process returns to step 132 otherwise the process of CRME terminates.
  • a resource management aims at improving availability, throughput, response time, or reducing energy consumption.
  • a resource management of high throughput and low energy consumption for heterogeneous multicore processors A multicore system with n tasks, m cores and q core voltage stages is represented by two three-dimensional matrices of data, i.e., two data cubes, each with three axes
  • the data in one matrix i.e., a time cube
  • the other matrix i.e., an energy cube
  • the energy efficient task mapping problem is to find n data from the two cubes such that the overall completion time and the overall energy consumption of all tasks are comparatively small.
  • ETC ⁇ t i>J k ⁇ nxmxg denote the time cube and EEC ⁇ e i i i c ⁇ xmxg the energy cube.
  • Element t iJ:k in matrix ETC represents a completion time of task when this task i is executed on core j with voltage level k.
  • element e ⁇ k in matrix EEC represents an energy consumption of task when this task i is executed on core j with voltage level k.
  • a task mapping M is n tuples ⁇ e [l,ri],j € [l, w],& e [l,q] ⁇ and each number in the tuples is a natural number.
  • the mapping problem can be formalized as follows:
  • a task mapping is shown in Fig. 4, in which each task is an energy block with execution time on "Time” axis and run-time voltage on “Voltage” axis. Cores are aligned along "Core” axis. The capacity or volume of each block is the energy consumed by the corresponding task. The depth is left unused in this figure for further improvement of core sharing. Two cores and nine tasks are shown in Fig. 4. Task execution times and energy consumptions change along with different cores and voltage values. The goal of task mapping is to find such task assignments with minimum summarized capacity and minimum overall task completion time, hereinafter also named "makespan.”
  • An object of the present invention is to provide a resource management system which can achieve minimum makespan and minimum energy consumption in heterogeneous multicore processors.
  • Another object of the present invention is to provide a resource management method which can achieve minimum makespan and minimum energy consumption in heterogeneous multicore processors.
  • a resource management method which explores task mapping with minimized total energy consumption and maximized throughput for heterogeneous multicore processor systems, comprises: preparing a plurality of task mapping methods; testing a first three-dimensional matrix and a second three- dimensional matrix to know a type of heterogeneity, the first three-dimensional matrix representing completion time of each combination of task, core and voltage setting, and the second three-dimensional matrix representing energy consumption of each combination of task, core and voltage setting; selecting one of the plurality of task mapping methods based on the type of heterogeneity; mapping tasks to cores of a target processor by using the selected task mapping method.
  • a resource management method which explores task mapping with minimized total energy consumption and maximized throughput for heterogeneous multicore processor systems, comprises: initially filing a first three-dimensional matrix and a second three-dimensional matrix based on characteristics of tasks and cores of a target system, the first three-dimensional matrix representing completion time of each combination of task, core and voltage setting, and the second three-dimensional matrix representing energy consumption of each combination of task, core and voltage setting; executing a plurality of task mapping methods in parallel using values stored in the first and second three-dimensional matrices; selecting one of results of the plurality of task mapping methods based on one of a specific task mapping goal from a user and a system design target to map the tasks to the cores; monitoring the cores to measure completion time and energy consumption of each task; modifying the values in the first and second three-dimensional matrices based on the measured completion time and energy consumption.
  • a management system which explores task mapping with minimized total energy consumption and maximized throughput for heterogeneous multicore processor systems, comprises: a storage holding a first three-dimensional matrix and a second three-dimensional matrix, the first three- dimensional matrix representing completion time of each combination of task, core and voltage setting, and the second three-dimensional matrix representing energy consumption of each combination of task, core and voltage setting; a recognition and detecting unit testing the first and second three-dimensional matrices to know a type of heterogeneity; a scheduler including a plurality of subcomponents each executing different task mapping method; and a selector selecting one of the subcomponents based on the type of heterogeneity and sending values of the first and second third-dimensional matrices to the selected subcomponent.
  • a resource management system which explores task mapping with minimized total energy consumption and maximized throughput for heterogeneous multicore processor systems, comprises: a storage holding a first three-dimensional matrix and a second three-dimensional matrix, the first three- dimensional matrix representing completion time of each combination of task, core and voltage setting, and the second three-dimensional matrix representing energy consumption of each combination of task, core and voltage setting; an estimating unit initially filing the first and second three-dimensional matrices based on characteristics of tasks and cores of a target processor; a scheduler including a plurality of subcomponents executing different task mapping methods in parallel based on values stored in the first and second three-dimensional matrixes; a selector selecting one of outputs of the subcomponents based on one of a specific task mapping goal from a user and a system design target to map the tasks to the cores; and a correction unit monitoring the cores to measure completion time and energy consumption of each task to modify the values in the first and second three
  • Fig. 1 is a flowchart illustrating procedure of MEME (Minimum Energy Minimum Energy);
  • Fig. 2 is a flowchart illustrating procedure of MEMC (Minimum Energy Minimum Completion time);
  • Fig. 3 is a flowchart illustrating procedure of CRME (Contention Resolved Minimum Energy);
  • Fig. 4 is a schematic view illustrating a task mapping
  • Fig. 5 is a flowchart illustrating procedure of MMER (Min-Min with Energy Relaxing);
  • Fig. 6 is a view illustrating an example of algorithm of MMER;
  • Fig. 7 is a flowchart illustrating procedure of KTEM (Min-Min with Key Task Energy Minimizing);
  • Fig. 8 is a view illustrating an example of algorithm of KTEM
  • Fig. 9 is a flowchart illustrating procedure of S M ;
  • Fig. 10 is a view illustrating an example of algorithm of S M ;
  • Fig. 11 is a view illustrating a system architecture according to one exemplary embodiment of the present invention.
  • Fig. 12 is a block diagram illustrating an example of a system embodying the architecture shown in Fig. 11 ;
  • Fig. 13 is a flowchart illustrating procedure of dynamically correcting estimated ETC and EEC
  • Fig. 14 is a view illustrating a system architecture according to another exemplary embodiment of the present invention.
  • Fig. 15 is a block diagram illustrating an example of a system embodying the architecture shown in Fig. 14.
  • Each exemplary embodiment supports the architecture with known ETC and EEC matrices or unknown matrices.
  • the two matrices can be built through static software test or dynamic run-time data estimation.
  • each of the exemplary embodiments make uses of multiple task mapping methods.
  • the multiple task mapping methods include three new methods, which are newly proposed in this description, to adapt to different practical applications.
  • the system of resource management can choose the most suitable task mapping method according to its function of heterogeneity recognition or application's requests, or use all mapping methods and then choose the most desired result.
  • the architecture has four implementations in reality.
  • the first method is named Min-Min with Energy Relaxing (MMER), which maps a task to a core in which the completion time of the task is minimal if the makespan will increase by the task with any voltage value. If there is a voltage value with which the makespan will not increase, the task will choose a core where its energy consumption is reduced most greatly.
  • MMER Min-Min with Energy Relaxing
  • Fig. 5 illustrates the procedure of MMER.
  • the task mapping with minimum completion time is selected, and, in step 212, it is determined whether makespan is increased or not. If makespan is increased, the process goes to step 215 otherwise it is determined whether voltage with less energy cost and without increasing makespan exists or not in step 213. If the voltage with less energy cost exists, then the task mapping is updated and the process goes to step 215, otherwise the process directly goes to step 215. In step 215, the task is mapped to the corresponding core. It is determined whether there is an unprocessed task or not, in step 216. If there is an unprocessed task, the process returns to step 212 otherwise the process of MMER terminates.
  • the algorithm of MMER is illustrated in Fig. 6 by means of pseudocode. As shown in Fig. 5 and 6, the task mapping method MMER tries to reduce the energy consumption of a task without increasing the makespan.
  • Fig. 7 illustrates the procedure of KTEM.
  • step 221 TEH is calculated for each task in accordance with eq. (1) and, in step 222, tasks are sorted by non-increasing TEH into E. Then, the task with overall minimum completion time is selected in step 223.
  • step 224 it is determined whether makespan is increased or not. If the makespan is increased, then the first task is get out from E in step 225 and this task is mapped to a core with the minimum energy (ME) in step 226. If the makespan is not increased in step 224, then the task with the minimum completion time is mapped in step 227.
  • ME minimum energy
  • step 226 or 227 the mapped task is deleted from E in step 228, and it is determined whether there is an unprocessed task or not, in step 229. If there is an unprocessed task, the process returns to step 223 otherwise the process of KTEM terminates.
  • KTEM The algorithm of KTEM is illustrated in Fig. 8 by means of pseudocode.
  • the task mapping method KTEM uses Task Energy Heterogeneity to manage the energy elasticity of each task. The task with the greatest elasticity will be mapped to a core where its energy consumption is minimized, if the makespan must increase whichever task is mapped.
  • the third method is named S 3 M 2 , which sorts tasks three times to obtain such a task list in which the tasks in one end have the greater energy elasticity and the tasks in the other end have the greater time elasticity.
  • this method defines Task Time Heterogeneity (TTH) as follows:
  • Tasks are sorted by non-increasing TEH into a list E and also sorted by non-increasing TTH into a list T.
  • This method defines the distance ii, for each task / as follows:
  • the tasks are sorted by non-decreasing d into another list ET.
  • the task with the greatest ET elasticity, i.e., the first one in ET, will be mapped to a core where its energy consumption is minimized, if the makespan must increase whichever task is mapped.
  • Fig. 9 illustrates the procedure of S M .
  • TEH is calculated for each task in accordance with eq. (1) in step 231 and TTH is calculated for each task in accordance with eq. (2) in step 232.
  • tasks are sorted by non-increasingly TEH into E in step 233 and sorted by non- increasingly TTH into Tin step 234.
  • distance dj is calculated for each task in accordance with eq. (3).
  • the tasks are sorted by non-decreasing d into E in step 236, and the last task is taken from ET in step 237.
  • step 238 it is determined whether this task will increase makespan or not.
  • the first task in ET is mapped to a core with minimum energy (ME) in step 239 otherwise the last task in ET is mapped to a core with minimum completion time (MCT) in step 240. Then, the mapped task is deleted from ET in step 241, and it is determined whether there is an unprocessed task or not, in step 242. If there is an unprocessed task, the process returns to step 237 otherwise the process of S M terminates.
  • ME minimum energy
  • MCT minimum completion time
  • the algorithm of S M is illustrated in Fig. 10 by means of pseudocode.
  • the task mapping method S M uses both Task Energy Heterogeneity (TEH) and Task Time Heterogeneity (TTH) to manage the energy elasticity and time elasticity of each task.
  • Task Energy Heterogeneity and Task Time Heterogeneity are normalized and aligned through ET distances. The task with the greatest ET elasticity will be mapped to a core where its energy consumption is minimized, if the makespan must increase whichever task is mapped.
  • Fig. 11 illustrates the architecture of the resource management system of the present embodiment.
  • the architecture shown in Fig. 11 includes: two data matrices, EEC and ETC, each of which has n x m x q items of data; a heterogeneity recognition and detection (HRD) unit which can test the matrices to know the type of heterogeneity; a mapping method selection (MMS) unit; and a pool of task mapping methods.
  • the type of heterogeneity includes, for example, High/Low(H/L) task heterogeneity, H/L core heterogeneity, and H/L energy heterogeneity.
  • the HRD unit recognizes type of heterogeneity of the data matrices.
  • the MMS unit automatically select the most suitable task mapping method from the pool according to the recognition result from the HRD unit.
  • the MMS unit receives the specific mapping goal from user request, and possibly comprehensively select the most suitable task mapping method from the pool in terms of both the recognition result of the HRD unit and the specific task mapping goal.
  • the mapping goal includes, for example, only minimizing energy without optimizing throughput or completion time, or both being required.
  • the pool stores various task mapping methods.
  • the stored methods include, for example, methods based on simple mapping policies, methods based on scheduling algorithms, and possibly heuristics.
  • Examples of the methods stored in the pool are, for example, CRME, MEME, MEMC, MMER, KTEM, S 3 M 2 and so on. Each method is implemented as a functional subcomponent. It should be noted that MMER, KTEM and S 3 M 2 are newly proposed by this description as described above.
  • the task mapping is then carried out using the subcomponent of the method selected by the MMS unit and the schedule is thus created.
  • the EEC and ETC matrices are filled and controlled in different ways based on nature of applications.
  • one of this kind of applications consists of a set of programs periodically running.
  • a static program test which executes each task in every core and measures the execution time and energy consumption. The test may take a period of time but it is worth for static applications because the system and the applications are only tested one time. After all data are collected, the data are filled into EEC and ETC matrices.
  • a data estimation (DE) unit is necessary and estimates the execution time and energy consumption.
  • the estimation can be done through code analysis and some physical models. Then the estimated data are filled into the matrices.
  • a core monitor and data correction (MC) unit actually measures the execution time and energy consumption of a running task on the target processor and then corrects the corresponding data in the EEC and ETC. According to the errors, the DE unit may modify the uncorrected remaining data.
  • the overall accuracy of the system can be achieved if the applications run repeated for enough time, but the accuracy absolutely depends on the accuracy of DE unit if such an application only runs one time.
  • Fig. 12 illustrates a resource assignment system according to the architecture shown in Fig. 11.
  • the resource assignment system is configured to map tasks to cores of target processor 10 which is a heterogeneous multicore processor system.
  • the system includes: matrix storage 21 storing the ETC and EEC matrices; MMS unit 22; HRD unit 23; scheduler 24 mapping tasks to cores of target processor 10 to creating schedule; DE unit 26; and MC unit 27.
  • Subcomponents of the task mapping methods such as CRME, MEME, MEMC, MMER, KTEM, S 3 M 2 and so on are disposed inside pool unit 25 which is arranged in scheduler 24.
  • data of ETC and EEC are directly supplied to matrix storage 21 as static data when the applications are static. If the applications are not static, data of ETC and EEC are arranged by DE unit 26 and MC unit 27 based on dynamically correcting process. Details of the dynamically correcting process will be explained later.
  • MMS unit 22 and HRD unit 23 work in the same way as described above, and the data of ETC and EEC are supplied to one of the subcomponents in pool unit 25 from MMS unit 22 in accordance with the selection result of MMS unit 22. Then the subcomponent to which the data of ETC and EEC are supplied, i.e., the subcomponent of the selected task mapping method, maps the tasks to the cores of target processor 10 and thus scheduler 24 creates the schedule. The schedule is applied to target processor 10 and target processor 10 operates based on the schedule.
  • DE unit 26 initially estimates data in ETC and EEC base on task data and core data.
  • the task data describes characteristics of each task and the core data describes characteristics of each core of target processor 10.
  • the tasks are mapped based on the estimated ETC and EEC data and executed by target processor 10.
  • MC unit 27 monitors the cores of target processor 10 and record the completion time and energy
  • step 253 MC unit 27 compares data in ETC and EEC with the measured data and determines whether both data coincide with each other or not. If both data coincide, the process goes to step 257 otherwise MC unit 27 replaces the old data in ETC and EEC with the measured data in step 254.
  • step 255 it is determined whether all data of a task have been corrected or not. If there are remaining data, then DE unit 26 modifying the remaining data in EEC and ETC based on the measured data, in step 256, and the process goes to step 257, otherwise the process directly goes to step 257.
  • step 257 it is determined whether there is an uncorrected task or not. If all tasks have been corrected then the process of DE-MC terminates otherwise the process goes to step 252.
  • Fig. 13 illustrates architecture of a resource management system according to another exemplary embodiment.
  • the architecture shown in Fig. 13 is different from one shown Fig. 11 in that the MMS unit and the HRD unit are removed and an output method selection (OMS) unit is added. Since the MMS unit is not provided, data of EEC and ETC matrices are supplied to the multiple task mapping methods in the pool in parallel. The subcomponents in the pool execute the task mapping methods in parallel and mapping results are then supplied to the OMS unit in parallel.
  • the OMS unit selects the most desired output from the parallel outputs of the pool according to the specific task mapping goals from user requests or the system design target. The selected output, i.e., schedule, is then supplied to the target processor.
  • Fig. 14 illustrates a resource assignment system according to the architecture shown in Fig. 13.
  • the resource assignment system is configured to map tasks to cores of target processor 10 which is a heterogeneous multicore processor system.
  • the system includes: matrix storage 21 storing the ETC and EEC matrices; scheduler 24 mapping tasks to cores of target processor 10 to creating schedule; DE unit 26; MC unit 27; and OMS unit 28.
  • Scheduler 24 includes pool unit 25 and data of ETC and EEC matrices are supplied to the multiple subcomponents in pool unit 25 in parallel. Each subcomponent maps the tasks to cores in accordance to corresponding task mapping method.
  • OMS unit 28 is arranged on the output side of scheduler 24 and selects one of the outputs of scheduler 24. The selected output, i.e., schedule, is applied to target processor 10.
  • Matrix storage 21, DE unit 16 and MC unit 27 work in the same manner as the case shown in Fig. 11.
  • the above exemplary embodiments try to achieve minimum makespan and minimum energy consumption in heterogeneous multicore processors.
  • the existing methods in related arts cannot well exploit the heterogeneity of tasks and cores.
  • the task mapping in related arts is rigid without well making use of DVS, or conservative in searching the solution space, or optimistic to assume that physical features models are well known. Although some advanced techniques were proposed, these techniques are time-costly because of the complexity of the problem.
  • the methods according to the exemplary embodiments can effectively search the solutions with low worst-case time costs.
  • the three methods, MMER, KTEM and S 3 M 2 i.e., are designed for different applications with certain heterogeneities of tasks and cores and avoid that only energy consumption is optimized or only makespan, practically implying throughput, is optimized.
  • the resource management according to the exemplary embodiments is able to manage multiple task mapping methods and support four different implementations to adapt to various environments.
  • Variant 1 The whole or part of the exemplary embodiments disclosed above have further variants as described below.
  • Variant 1 The whole or part of the exemplary embodiments disclosed above have further variants as described below.
  • the resource management system minimizing the total energy consumption and maximizing the throughput for heterogeneous multicore processor systems is provided.
  • the resource management system is characterized as:
  • mapping Method Selection (c) a system architecture of selecting the most suitable mapping method, the system architecture including the Heterogeneity Recognition and Detection (HRD) unit and the Mapping Method Selection (MMS) unit;
  • HRD Heterogeneity Recognition and Detection
  • MMS Mapping Method Selection
  • the metrics of managing the heterogeneities includes at least one of (a) Task Energy Heterogeneity (TEH), (b) Task Time Heterogeneity (TTH), and (c) Distance.
  • TEH is defined by eq. (1) and represents the task energy consumption elasticity.
  • TTH is defined by eq. (2) and represents the task execution time elasticity.
  • Distance is defined by eq. (3) and represents the alignment of TEH and TTH.
  • Term "elasticity" means the variety of the values in heterogeneous multicore.
  • the methods of managing the heterogeneities of energy, time and their alignment are provided. Especially the method of creating energy-time alignment includes sorting TEH, TTH and Distance to align energy-time heterogeneity.
  • the HRD unit includes functions of recognition and detection.
  • the MMS unit chooses a most suitable task mapping method according to the output of the HRD unit and user requests.
  • the OMS unit operates a set of methods to create different task mapping schedules and then selects the most desired one as the final decision in terms of user requests or system definitions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Power Sources (AREA)

Abstract

L'invention porte sur un procédé de gestion de ressources explorant un mappage de tâche avec une consommation d'énergie totale rendue minimale et un débit rendu maximal pour des systèmes de processeurs multicœur hétérogènes. Le procédé comprend : la préparation d'une pluralité de procédés de mappage de tâche ; le test d'une première matrice tridimensionnelle et d'une seconde matrice tridimensionnelle pour savoir un type d'hétérogénéité, la première matrice tridimensionnelle représentant un temps d'achèvement de chaque combinaison de configuration de tâche, de cœur et de tension, et la seconde matrice tridimensionnelle représentant une consommation d'énergie de chaque combinaison de configuration de tâche, de cœur et de tension ; la sélection de l'un de la pluralité de procédés de mappage de tâche sur la base du type d'hétérogénéité ; le mappage de tâches à des cœurs par utilisation du procédé de mappage de tâche sélectionné.
PCT/JP2011/050604 2011-01-07 2011-01-07 Système de gestion de ressources efficace en énergie et procédé pour des processeurs multicœur hétérogènes WO2012093498A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/JP2011/050604 WO2012093498A1 (fr) 2011-01-07 2011-01-07 Système de gestion de ressources efficace en énergie et procédé pour des processeurs multicœur hétérogènes

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2011/050604 WO2012093498A1 (fr) 2011-01-07 2011-01-07 Système de gestion de ressources efficace en énergie et procédé pour des processeurs multicœur hétérogènes

Publications (1)

Publication Number Publication Date
WO2012093498A1 true WO2012093498A1 (fr) 2012-07-12

Family

ID=46457363

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2011/050604 WO2012093498A1 (fr) 2011-01-07 2011-01-07 Système de gestion de ressources efficace en énergie et procédé pour des processeurs multicœur hétérogènes

Country Status (1)

Country Link
WO (1) WO2012093498A1 (fr)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2016082320A1 (fr) * 2014-11-26 2016-06-02 浪潮(北京)电子信息产业有限公司 Système hôte hétérogène hybride, procédé de configuration de ressource et procédé d'ordonnancement de tâches
US10114613B2 (en) 2016-09-07 2018-10-30 International Business Machines Corporation Mixed-precision memcomputing system
CN111090505A (zh) * 2017-11-02 2020-05-01 联发科技股份有限公司 多处理器系统中任务调度的切换策略
CN111679897A (zh) * 2020-06-05 2020-09-18 重庆邮电大学 异构多核片上系统任务分配方法和装置

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004252900A (ja) * 2003-02-21 2004-09-09 Sharp Corp 非対称マルチプロセッサシステム、それを備えた画像処理装置および画像形成装置
JP2006293768A (ja) * 2005-04-12 2006-10-26 Univ Waseda マルチプロセッサシステム及びマルチグレイン並列化コンパイラ
JP2011501268A (ja) * 2007-10-11 2011-01-06 マイクロソフト コーポレーション 階層的予約資源スケジューリング・インフラストラクチャ

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004252900A (ja) * 2003-02-21 2004-09-09 Sharp Corp 非対称マルチプロセッサシステム、それを備えた画像処理装置および画像形成装置
JP2006293768A (ja) * 2005-04-12 2006-10-26 Univ Waseda マルチプロセッサシステム及びマルチグレイン並列化コンパイラ
JP2011501268A (ja) * 2007-10-11 2011-01-06 マイクロソフト コーポレーション 階層的予約資源スケジューリング・インフラストラクチャ

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2016082320A1 (fr) * 2014-11-26 2016-06-02 浪潮(北京)电子信息产业有限公司 Système hôte hétérogène hybride, procédé de configuration de ressource et procédé d'ordonnancement de tâches
US9904577B2 (en) 2014-11-26 2018-02-27 Inspur (Beijing) Electronic Information Industry Co., Ltd Hybrid heterogeneous host system, resource configuration method and task scheduling method
US10114613B2 (en) 2016-09-07 2018-10-30 International Business Machines Corporation Mixed-precision memcomputing system
CN111090505A (zh) * 2017-11-02 2020-05-01 联发科技股份有限公司 多处理器系统中任务调度的切换策略
CN111090505B (zh) * 2017-11-02 2024-01-26 联发科技股份有限公司 一种多处理器系统中任务调度的方法和系统
CN111679897A (zh) * 2020-06-05 2020-09-18 重庆邮电大学 异构多核片上系统任务分配方法和装置

Similar Documents

Publication Publication Date Title
Nesmachnow et al. Energy-aware scheduling on multicore heterogeneous grid computing systems
US9715408B2 (en) Data-aware workload scheduling and execution in heterogeneous environments
EP2417534B1 (fr) Mécanisme optimisé de migration de machine virtuelle
Shmueli et al. Backfilling with lookahead to optimize the packing of parallel jobs
Chen et al. Samr: A self-adaptive mapreduce scheduling algorithm in heterogeneous environment
Liu et al. Multi-objective scheduling of scientific workflows in multisite clouds
US20110022870A1 (en) Component power monitoring and workload optimization
US8230249B2 (en) Dynamic selection of server states for servers in a cluster of servers
CN108701060A (zh) 用于计算系统自动调整的方法
US10642652B2 (en) Best trade-off point on an elbow curve for optimal resource provisioning and performance efficiency
Chen et al. HAT: history-based auto-tuning MapReduce in heterogeneous environments
Li et al. Opportunistic scheduling in clouds partially powered by green energy
Oxley et al. Makespan and energy robust stochastic static resource allocation of a bag-of-tasks to a heterogeneous computing system
Pakize A comprehensive view of Hadoop MapReduce scheduling algorithms
US9424212B2 (en) Operating system-managed interrupt steering in multiprocessor systems
CN112783616B (zh) 一种并发冲突处理方法、装置及计算机存储介质
Huang et al. Novel heuristic speculative execution strategies in heterogeneous distributed environments
Deng et al. A data and task co-scheduling algorithm for scientific cloud workflows
Marcus et al. Releasing Cloud Databases for the Chains of Performance Prediction Models.
Makrani et al. Energy-aware and machine learning-based resource provisioning of in-memory analytics on cloud
CN109388486A (zh) 一种针对异构内存与多类型应用混合部署场景的数据放置与迁移方法
Choi et al. Task Classification Based Energy‐Aware Consolidation in Clouds
Hu et al. Improved heuristic job scheduling method to enhance throughput for big data analytics
WO2012093498A1 (fr) Système de gestion de ressources efficace en énergie et procédé pour des processeurs multicœur hétérogènes
Maroulis et al. A holistic energy-efficient real-time scheduler for mixed stream and batch processing workloads

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 11855114

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 11855114

Country of ref document: EP

Kind code of ref document: A1