CN109992385A - A kind of inside GPU energy consumption optimization method of task based access control balance dispatching - Google Patents
A kind of inside GPU energy consumption optimization method of task based access control balance dispatching Download PDFInfo
- Publication number
- CN109992385A CN109992385A CN201910205801.4A CN201910205801A CN109992385A CN 109992385 A CN109992385 A CN 109992385A CN 201910205801 A CN201910205801 A CN 201910205801A CN 109992385 A CN109992385 A CN 109992385A
- Authority
- CN
- China
- Prior art keywords
- task
- queue
- impact factor
- sequence
- energy consumption
- Prior art date
- Legal status (The legal status 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 status listed.)
- Granted
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
- G06F9/48—Program initiating; Program switching, e.g. by interrupt
- G06F9/4806—Task transfer initiation or dispatching
- G06F9/4843—Task transfer initiation or dispatching by program, e.g. task dispatcher, supervisor, operating system
-
- 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]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5011—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resources being hardware resources other than CPUs, Servers and Terminals
- G06F9/5022—Mechanisms to release resources
-
- 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]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
- G06F9/505—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals considering the load
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention belongs to equipment states and the task of task characteristic to balance dispatching technique field, disclose energy consumption optimization method inside a kind of GPU of task based access control balance dispatching, using balanced impact factor CB and SM the frequency of use information PHB handle of pending program, the CB-HRV task scheduling algorithm of balanced impact factor and SM frequency of use information is proposed；One group task is assigned in corresponding SM according to the balanced impact factor of task according to the time of task execution and executes by CB-HRV task scheduling algorithm；Strategies scheduler task balance is assigned in each SM；Meanwhile the service efficiency of device resource is improved using SM frequency of use information；The method for scheduling task of energy optimization is realized by the integration application to pending mission bit stream and resource information.The present invention realizes the scheduling of the dynamic equilibrium in multiple SM, has reached the energy optimization target of GPU.
Description
Technical field
The invention belongs to equipment states and the task of task characteristic balance dispatching technique field more particularly to one kind to be based on
Energy consumption optimization method inside the GPU of task balance scheduling.
Background technique
Currently, the prior art commonly used in the trade is such that graphics processor (English: Graphics Processing
Unit, abbreviation: GPU), also known as show core, vision processor, display chip, be one kind specially PC, work station,
The microprocessor that image operation works on game machine and some mobile devices (such as tablet computer, smart phone).Purposes be by
Display information required for computer system carries out conversion driving, and provides line scan signals to display, controls display
Correct display is one of the critical elements, and the important equipment of " human-computer dialogue " for connecting display and PC mainboard.
Video card undertakes the task of output display figure, for being engaged in professional graphic as an important component in host computer
Video card is extremely important for the people of design.
One, the energy optimization research of task based access control scheduling, carrying out energy optimization by task schedule in GPU can be from
Set about in terms of hardware and software two.Software-based energy optimization strategy is small because of hardware costs and implementation is relatively easy, because
This is considered as relatively effective energy optimization means.In general, it is divided by the task scheduling strategy of software towards program
The static method of design and the dynamic approach of oriented manipulation system and compiler.The static method of programming-oriented is straight
Connected the distribution of SM resource needed for programming carries out.Scholars propose static Task Assigned Policy to realize task
Scheduling, to reduce system energy consumption.The method of static compilation is not necessarily to carry out the modification of GPU operation system or hardware view,
Realize that cost is relatively small.But simultaneously as this method do not account for the resource situation and task execution of SM in GPU when
Between target, therefore energy optimization effect is limited.Especially when performing environment changes, the energy consumption of this method
Effect is just more difficult to be protected.If it is intended to obtaining the power dissipation obj ectives of specific environment, need before program scheduler to program
It is previously run to obtain parameter and Manual Override program codes.Second method is dynamic dispatching method, also
It is to carry out task balance using the time parameter of known GPU resource parameter and task execution during program executes
Scheduling.These parameters include the size for handling data, task to the degree of dependence of CACHE, utilization rate of equipment and installations of SM etc..Dynamically
Scheme balance dispatching is carried out by the analysis of analysis and task characteristic itself to the resource inside GPU, tend to obtain
Effect more better than static scheduling, weight analysis Dynamic Scheduling Strategy and propose the energy consumption of task based access control equilibrium dynamic dispatching a kind of
Optimization method.
Two, task based access control executes the scheduling strategy (Random allocation device) of time, GPU energy optimization
Research scene be the operative scenario based on more SM.It is tactful to each according to certain into the task of GPU from SM task dispatcher
SM distributes task.The status of each SM is equality in task implementation procedure, that is to say, that their task load ability is
Identical, task dispatcher is generally basede on equality strategy and carries out task distribution.But since task dispatcher is distributed in each SM
Task characteristic it is different, the service efficiency of each SM is not identical.
In current task schedule, it is often based upon the scheduling strategy of program execution time.Assuming that pending in system
Agenda is Pi (0≤i≤n), and the available processors resource in current system is SMi (0≤i≤n), then to allow
System executive energy consumption is minimum:
As shown from the above formula, system energy consumption is represented by all SM, CPU in system and the energy consumption of mainboard consumption
The sum of.Further it is represented by the product of respective power and time.The certain phase of program and quantity to be scheduled for one group
For SM, the sequence to scheduler task is different according to the difference of dispatching algorithm, but not function when change task execution
Consumption.The average power consumption to the sequence of scheduler task remains unchanged in other words.Therefore, system energy consumption can further indicate that
For average power consumptionWith the product of time T.In order to make system energy consumption minimum when executing agenda, it is necessary to make average function
ConsumptionIt is as small as possible with time T.And for different dispatching methods, average power consumptionIt is certain, therefore, for
System energy consumption minimum will to execute time T minimum.Many energy optimization schemes are all based on the scheme for executing the time,
That is making overall deadline of the task on SM minimum by a kind of dispatching algorithm.
Three, based on the task scheduling strategy (Random allocation device) for being randomly assigned equipment
It is executed in dynamic dispatching of the time as target to reduce average time, task is carried out based on equipment (SM) state
Scheduling, Random allocation device (RAD) is common energy optimization means, and equipment here refers in GPU
SM.A kind of method of low-power consumption distribution load.The thought of this method is directly equipment (SM) to be randomly assigned to times to be executed
It make sures with the state without considering equipment.That is, this method thinks that the SM equipment in GPU is all equality, without spy
Difference in property and in use state.The pseudocode that the algorithm of this strategy is realized has been shown below.
Random_Allocation_Device Approach(RAD)
This method is advantageous on time complexity, because there is no consider SM as Resource Properties for this method
Difference is randomly assigned that task to be executed is enabled to carry out SM positioning very fast, and executes at once.But as
Described in algorithm, for this method there is no difference of the SM as Resource Properties is considered, this also cannot be the characteristic of execution task
Correlation is carried out with SM to be executed to link.For example, working as some SM that some task is probabilistically assigned, it is more likely that the SM
Still it is occupied by other tasks, and needs just be released for quite a long time.Under this strategy being randomly assigned,
Task cannot be executed in time, and real-time is excessively poor.
Four, the task scheduling strategy (device history-based) based on device history state
If it is considered that the history use state of equipment (SM), carrying out task schedule based on equipment (SM) state is often to make
Method.DFB (device history-based) algorithm below such as considers the release conditions of SM.Function
IsSMfree (i) has inquired the state of equipment, if equipment is released, the SM resource of release will be placed to resource pool coequally
It is used for all tasks.
Device_Free_based Approach(DFB)
Due to consideration that the state of SM, the dispatching method are capable of providing more preferable than being randomly assigned in most cases
Performance.However, the dispatching method but cannot be guaranteed the consistency of task execution, because it has considered only current set
For whether occupied by previous task.That is this method does not consider the temporal correlation of program in task
With the influence of spatial coherence, this correlation is possible to will cause frequent switching of the task in each SM, and it is unnecessary to cause
Energy loss.Meanwhile the correlation of this task also will cause considerable task and migrate to hold to further result in SM in SM
The loss of line efficiency.
Five, it carries out further strategy according to the state of SM to attempt, this is the algorithm policy based on the performance of SM history
(Performance History Based), which has used the execution time ratio of GPU.This method is suitable in selection
SM can read the historical information for executing the time before executing task, calculate History Ratio Value by following formula
(PHB), the history resource frequency of use of referred to as some SM.
Ratio [i]=Execution on the SMi/Execution on the SMi+1；
The each SM of the parameter declaration when being executed between on frequency, which substantially reflects execution task in SM
On program time and spatial coherence；It is thus possible to improve the utilization rate of entire SM resource is excellent to carry out to energy consumption
Change.
Illustrate that the task dependencies of the SM are relatively preferable if the use of some SM is more frequent according to this strategy, hold
Line efficiency is also just relatively high.It is prioritized and is assigned to new task；The pseudo- generation of dispatching method based on frequency of use
Code.
Performance_history_Based Approach(PHB)
Although it is contemplated that in GPU each SM difference, however could not consider that present procedure executes in corresponding SM surplus
The remaining time, and then will lead to the excessive use of SM and owe to utilize.For example, if all pending programs are according to above-mentioned calculating
History resource frequency of use be all distributed on first SM, then first SM will excessive use and remaining SM will
It can become to owe utilization.
In conclusion problem of the existing technology is:
(1) loss caused by task migrates on each SM, such as the dispatching algorithm based on equipment release are not accounted for
(device history-based algorithm) considers preferentially using the SM being released, but there is no consideration task moving on SM
It moves；Probably due to being migrated repeatedly between SM for multiple tasks and generate a large amount of energy consumption, weaken the energy consumption and performance of GPU.
(2) device resource and the task of execution overall comprehensively consider；Algorithm based on the performance of SM history
Tactful (Performance History Based) although in the utilization rate of SM is considered, there is no combine to appoint
The migration characteristic of business cannot carry out balanced use to SM, it is inefficient to cause task execution.
Solve the difficulty of above-mentioned technical problem:
(1) current task scheduling algorithm is not high to the utilization rate of the computing unit (stream handle SM) of GPU, cannot be comprehensive
Close influence of the feature of the utilization efficiency and task that consider SM to energy consumption.
(2) transport phenomena in task schedule is common problem, and migration of the task in stream handle will cause
Energy loss, current scheduling strategy cannot effectively alleviate task immigration phenomenon.
(3) there is presently no effective energy optimization task scheduling algorithms to be able to solve two above problem, needs algorithm
Design.
Solve the meaning of above-mentioned technical problem:
(1) problem not high for computing unit utilization rate, the present invention analyze task schedule energy consumption under GPU environment and ask
The correlative factor of topic, and creatively taken out the history utilization for influencing the balanced impact factor CB and SM of task immigration
Rate HRV solves the problems, such as energy loss caused by the task immigration in scheduling using balance idea.By balance policy to SM
In task carry out rational management, the utilization rate of computing unit is improved, to reduce GPU energy loss.
(2) for energy loss caused by task immigration, the present invention comprehensively utilizes computing resource attribute (SM history resource
Utilization rate) with task characteristic (balanced impact factor), realize better energy optimization strategy by the way that the two is combined.It is logical
The utilization rate for crossing the sequence and SM to task balance sorts to optimize task scheduling strategy, and reducing in SM for task is moved
Phenomenon is moved, to realize the comprehensive energy consumption optimization of GPU.
(3) be directed to energy optimization task scheduling algorithm, the present invention construct task balance dispatching algorithm equilibrium influence because
Son-utilization rate of equipment and installations task scheduling strategy (CB-HRV), and carry out algorithm frame building and realize.
The present invention in order to effectively alleviate the above problem and enable GPU more widely application to adapt to program
Diversity, propose balance dispatching strategy, realize a low overhead, portable good task energy optimization model.This hair
Bright scheduling strategy has fully considered migration characteristic of the task on SM；And optimize GPU in conjunction with the equipment execution efficiency of SM
Inside energy consumption, referred to as based on task scheduling strategy (the CB-HRV Task of balanced impact factor-utilization rate of equipment and installations
Schedule Approach)。
Summary of the invention
In view of the problems of the existing technology, the present invention provides energy consumptions inside a kind of GPU of task based access control balance dispatching
Optimization method.
The invention is realized in this way energy consumption optimization method inside a kind of GPU of task based access control balance dispatching, the base
Energy consumption optimization method is used using the balanced impact factor CB and SM of pending program inside the GPU of task balance scheduling
Frequency information PHB handle proposes the CB-HRV task scheduling algorithm of balanced impact factor and SM frequency of use information；CB-HRV
One group task is assigned to corresponding SM according to the balanced impact factor of task according to the time of task execution by task scheduling algorithm
Middle execution；Strategies scheduler task balance is assigned in each SM；Meanwhile it being set using SM frequency of use information to improve
The service efficiency of standby resource.The task of energy optimization is realized by the integration application to pending mission bit stream and resource information
Dispatching method.
Further, the CB-HRV task scheduling algorithm specifically includes:
(1) the balance influence factor CB for calculating each task, calculates the utilization rate PHB of each SM resource；
(2) it resequences according to the sequence of the balance influence factor of task from big to small to each task, forms task team
Column；
(3) it resequences according to the sequence of the utilization rate PHB of SM resource from big to small to each SM；
(4) task queue of sequence is successively matched in the SM queue of sequence.
Further, the CB-HRV task scheduling algorithm forms two queues, and one is history frequency of use queue, separately
One is task balance impact factor；
The queue of task balance impact factor is indicated with TaskBalance [], and with SortTask (TaskBalance [i]
Function is to the queue descending sort；The queue of history frequency of use is indicated with SmRatio [], and with SortSm (SmRatio [i])
Function carries out descending sort to the queue；For the SM history frequency of use PHB after sequence, array SmRationIndex is used
[] storage；For the task balance impact factor sequence after sequence, stored using array TaskCBIndex [].
Further, the CB-HRV task scheduling algorithm further comprises:
(1) the balance influence factor CB of each task is calculated with CalculateBalance (P [i]) function first, and
It is stored in TaskBalance [i] array, the use of each SM resource is then calculated with CalculateRatio (SM [i])
Rate ratio value, and be stored in SmRatio [i] array；
(2) it is given using SortTask (TaskBalance [i]) according to the sequence of impact factor CB value from big to small each
Task rearrangement, and be stored in TaskCBIndex [] array, form task queue；With SortSm (SmRatio [i])
Function is resequenced according to the sequence of SM utilization rate from big to small to each SM, and result is stored in SmRationIndex
In [] array；
(3) task queue of sequence is successively matched in the SM queue of sequence.
Further, energy consumption optimization method inside the GPU of the task based access control balance dispatching further include:
The first step, analyzes the balanced impact factor information of pending task, and is ranked up from small to large；Comparing
Being easy to produce transport phenomena of the task distributes to the relatively high SM of utilization rate in GPU and not the receiving migration interference greatly of the task is divided
The relatively low SM of dispensing utilization rate；
Second, two queues are formed, first queue is the balanced descending sequence of impact factor CB according to task
Queue；The balanced impact factor of task is bigger, shows that the task is also bigger by the degree of migration interference；Second queue be
By the queue PHB of the descending sequence of the frequency of use of SM；Frequency of use PHB value is higher, illustrates that the frequency of use of the SM is got over
Height, execution efficiency are also higher；
The equilibrium maximum task of impact factor distributed to the highest SM of execution efficiency and executed by third step, is thus reduced
It is influenced as caused by task immigration；Successively task is distributed to according to balanced impact factor size the SM of descending sequence
It executes, realizes overall optimisation strategy；
4th step is being obtained by the way that the high task of balanced impact factor is configured in the higher SM of execution efficiency wait hold
In the case of the temporal information of row task, the impact factor information according to task is ranked up task, secondly obtains the history of SM
Frequency of use simultaneously sorts；The mission number distribution recorded according to the task queue of each SM is given in corresponding SM.
Another object of the present invention is to provide energy consumption inside a kind of GPU using the task based access control balance dispatching is excellent
The graphics processor of change method.
In conclusion advantages of the present invention and good effect are as follows: the present invention analyzes the correlation of task based access control scheduling balance
Theoretical and algorithm, and propose that bonding apparatus state balances scheduling strategy with the task of task characteristic: balanced impact factor-is set
Standby utilization rate task scheduling strategy (CB-HRV)；Strategy had both considered the Resource Properties of SM while having had also contemplated task in SM
Migration characteristic, dynamic workload balance dispatching is realized in GPU using algorithm.Resource of the present invention to each SM
Utilization rate is analyzed, and the balance dispatching of task is carried out using the characteristic of task, and the dynamic realized in multiple SM is flat
Weighing apparatus scheduling, has reached the energy optimization target of GPU.
The present invention is by reasonably distributing processing unit task disappearing to the system energy consumption during reducing task execution
Consumption.CB-HRV method for scheduling task obtains the time data of a group task first, and carries out according to balanced impact factor to them
Sequence, while the specific SM usage history utilization rate of the task of execution is ranked up, then successively using the thought of balance dispatching
Mission number after descending is assigned on the processing unit of descending sort, is realized based on balanced GPU energy optimization dispatching party
Method.
In experimental section, the present invention carries out Experimental comparison to four kinds of Dynamic Assignments respectively and analyzes superiority and inferiority；Experiment
The result shows that balanced impact factor-utilization rate of equipment and installations task scheduling strategy can be such that the energy consumption and performance of GPU improves
15.69%.The scheduling scheme of time is executed, the experimental results showed that, compared with the conventional method, CB-HRV scheme energy consumption average out to
10.38%；Compared with the conventional method, dispatching method of the invention is more effective, executes the scheduling scheme of time；Experimental result table
It is bright, compared with the conventional method, CB-HRV scheme energy consumption average out to 10.38% of the invention；Compared with the conventional method, from experiment
As a result it is seen on, CB-HRV method proposed by the present invention is effective, reasonable and feasible.
Detailed description of the invention
Fig. 1 is energy consumption optimization method flow chart inside the GPU of task based access control balance dispatching provided in an embodiment of the present invention.
Fig. 2 is the task immigration phenomenon schematic diagram in SM provided in an embodiment of the present invention；
In figure: (a) the long-living overload phenomenon of task；(b) it reschedules task and generates task immigration phenomenon.
Fig. 3 is the task completion time schematic diagram of different allocation plans in SM provided in an embodiment of the present invention.
Fig. 4 is the balanced impact factor schematic diagram of different benchmarks provided in an embodiment of the present invention.
Fig. 5 is Matrix Multiplication task execution timing diagram provided in an embodiment of the present invention.
Fig. 6 is Matrix Multiplicatio scheduling time distribution ratio example diagram provided in an embodiment of the present invention.
Fig. 7 is Histogram task execution timing diagram provided in an embodiment of the present invention.
Fig. 8 is Histogram scheduling time distribution ratio example diagram provided in an embodiment of the present invention.
Fig. 9 is Scalar products task execution timing diagram provided in an embodiment of the present invention.
Figure 10 is Scalar products scheduling time distribution ratio example diagram provided in an embodiment of the present invention.
Figure 11 is Scalar products task execution timing diagram provided in an embodiment of the present invention.
Figure 12 is Scalar products scheduling time distribution ratio example diagram provided in an embodiment of the present invention.
Figure 13 is average energy consumption (test one) schematic diagram of different scheduling schemes provided in an embodiment of the present invention.
Figure 14 is average energy consumption (test two) schematic diagram of different scheduling schemes provided in an embodiment of the present invention.
Figure 15 is average energy consumption (test three) schematic diagram of different scheduling schemes provided in an embodiment of the present invention.
Figure 16 is that the energy consumption of CB-HRV dispatching method provided in an embodiment of the present invention promotes contrast schematic diagram.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to embodiments, to this hair
It is bright to be further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, not
For limiting the present invention.
The present invention analyzes the correlative factor of task schedule energy consumption problem under GPU environment, and converts task for this problem
Scheduling problem.The balanced impact factor of task is creatively introduced, the task immigration solved in scheduling using balance idea is made
At energy loss problem, by balance policy in SM task carry out rational management reduce GPU energy loss.Synthesis makes
With the equalization characteristic of Resource Properties and task, better energy optimization strategy is realized by the way that the two is combined, and to equilibrium
Scheduling carries out.Optimize task scheduling strategy by the utilization rate sequence of sequence and SM to task balance, realizes
The comprehensive energy consumption of GPU optimizes.Construct task balance dispatching algorithm equilibrium impact factor-utilization rate of equipment and installations task scheduling strategy
(CB-HRV), and code construction has been carried out.
Application principle of the invention is described in detail with reference to the accompanying drawing.
As shown in Figure 1, energy consumption optimization method packet inside the GPU of task based access control balance dispatching provided in an embodiment of the present invention
Include following steps:
S101: the balanced impact factor information of pending task is analyzed, and is ranked up from small to large；Hold comparing
Being also easy to produce transport phenomena of the task distributes to the relatively high SM of utilization rate in GPU and not the receiving migration interference greatly of the task is distributed
Give utilization rate relatively low SM；
S102: two queues are formed, first queue is the balanced descending sequence of impact factor CB according to task
Queue；The balanced impact factor of task is bigger, shows that the task is also bigger by the degree of migration interference；Second queue be by
The queue PHB of the descending sequence of the frequency of use of SM；Frequency of use PHB value is higher, illustrates that the frequency of use of the SM is higher,
Execution efficiency is also higher；
The equilibrium maximum task of impact factor: being distributed to the highest SM of execution efficiency and executed by S103, thus come reduce by
It is influenced caused by task immigration.Next, task is successively distributed to descending sequence according to balanced impact factor size
SM execute, realize overall optimisation strategy；
S104: pending obtaining by the way that the high task of balanced impact factor is configured in the higher SM of execution efficiency
In the case of the temporal information of task, the impact factor information according to task is ranked up task, and the history for secondly obtaining SM makes
With frequency and sort；The mission number distribution recorded according to the task queue of each SM is given in corresponding SM.
Application principle of the invention is further described with reference to the accompanying drawing.
1, the transport phenomena in SM task schedule
In task schedule in GPU, some task can not be completed only in a fixed SM, and task is in SM
Migration be inevitable phenomenon.There are two task in Fig. 2, the scheduler in SM dynamically monitors its scheduling queue.Scheming
Transition process of the task in two SM in 2.It (a) is the process for being not carried out task immigration, two tasks are divided in two SM
It does not execute.As seen from Figure 2, task 1 has been dispensed on SM1 and SM2 and has executed.Second task (is used orange table by scheduler
Show) it is allocated to SM2 execution.But a part and task 2 of task 1 execute on SM2 simultaneously, and SM2 is caused to overload.At this moment,
In order to avoid SM is crashed due to overload, scheduler can dispatch task 2 and execute on loading lesser SM1, as shown in Figure 2.
Task dispatcher can redistribute task by the above strategy to avoid the exception of GPU.But at the same time, by
The redundant overheads of GPU are caused in task immigration, task needs switch between SM.If there is n task enters SM group
Part causes the increase of energy consumption since the expense that task immigration generates continues to increase.It should in the scheduling strategy of task
Migration of the task in each SM is avoided as far as possible.
2, schedule equalization tasks analysis of strategies
One task executes on multiple SM if there is multiple warp, warp needs.If considering holding for some task
Line efficiency considers task in each SM comprehensive deadline executed.Fig. 3 holds some task at two SM (SM1, SM2)
Row is illustrated.
In Fig. 3, the difference that the same task takes different Task Assigned Policies to generate in two SM is presented.
The task needs 100 clock cycle to complete, and in the first scenario, Task Assigned Policy is assigned with 20 periods on SM1
Task, and the task of subsequent 80 clock cycle give SM2 complete.Task allocation plan needs 80 clocks in total
Period completes.The task of 40 clock cycle is then allocated to SM1 by second of allocation plan, 60 clock cycle in addition
Task then distributes to SM2 completion.Entire task was completed to execute within 60 periods under this allocation plan.See, latter side
Case (60 clock cycle) is completed faster.Second of workload allocation plan more effectively uses the SM inside GPU
Resource, this scheme are more balanced allocation strategies.
In order to illustrate two kinds of tactful differences, balanced impact factor CB (coefficient of has been taken out
Balance):
In formula: the impact factor of CB expression task.Wherein σ is standard deviation, and λ is the average value of SM deadline；Respectively
The CB value for calculating CB the first allocation plan just now in scheme is 72.8%, and the CB value of second scheme is 45.1%.In order to
The rule for finding each task balance impact factor tests their CB value using one group of CUDA benchmark program.Experiment is chosen
4 typical CUDA benchmark programs, respectively Matrix Multiplication, histogram, scalar
Products and BlackScholes.Following table illustrates four used benchmark program parameters of experiment in detail.
1 benchmark of table
This 4 groups of benchmark programs are tested using simulator, and calculate their CB value, as a result the institute in Fig. 4
Show, the CB value that the CB value of MM is about 40%, HG has been more than 60%.The CB value of SP and BS is relatively small, and SP is about 35%, BS's
CB value is less than 5%.It is uneven that task dispatcher will monitor such workload during application program executes, and will not complete
Warp adjacent SM is re-assigned to from the SM of overload, more efficiently use SM.This illustrates that different task characteristic has not
Same balanced impact factor, the migration that the balanced impact factor gone out on missions and task are summarized from experiment have proportional relation, also
It is to say that a possibility that balanced impact factor of some task is bigger, it is migrated between SM is higher.Influence task is completed
Efficiency and be finally reflected in terms of energy consumption, observation inspires the present invention that all tasks are carried out overall row in some way
Then sequence is allocated task in SM by sequence, so as to improve the utilization rate of each SM, reaches whole energy optimization
Target.
3, the scheduling strategy of task based access control equilibrium impact factor and computing resource utilization rate
It in order to avoid the migration of task, and realizes the Task Assigned Policy of balance dispatching, finds a kind of scheduling of equilibrium
Strategy.The strategy can overcome the shortcomings of in original scheduling strategy.Realize that (1) avoids task immigration in unnecessary SM as far as possible
The phenomenon that, to reduce the energy loss generated due to task immigration.(2) comprehensive to use Resource Properties and task characteristic.Current
Scheme generally uses one of SM state or task characteristic as the foundation of scheduling strategy, and the two is combined realization
Better energy optimization strategy.(3) comprehensive energy consumption that rational management realizes GPU is carried out to the task in SM by balance policy
Optimization；This needs to fully consider the harmony of task schedule, more effectively utilizes SM computing resource.
In heterogeneous system, the pending task appropriate to SM distribution is that one of decision systems execution time is important
Factor.Therefore, pending is appointed by comprehensively considering balanced impact factor CB and SM history frequency of use information PHB
Business, which is reasonably assigned to the balance dispatching that can be achieved on task in appropriate SM and improves SM utilization rate, realizes energy optimization
Target.
According to the deficiency of Existing methods, the balanced impact factor CB and SM frequency of use of pending program are comprehensively utilized
Information PHB handle proposes the CB-HRV task scheduling strategy (CB-HRV of balanced impact factor and SM frequency of use information
Scheduling Scheme).The thought of CB-HRV task scheduling algorithm be according to task execution time by a group task by
It is assigned in corresponding SM and executes according to the balanced impact factor of task.It task balance is distributed according to this strategies scheduler
Into each SM, migration of the task in each SM is efficiently avoided, meanwhile, it is set using SM frequency of use information to improve
The service efficiency of standby resource.A kind of energy optimization is realized by the integration application to pending mission bit stream and resource information
Method for scheduling task.
Specific method is that analysis need to distribute the balanced impact factor of task and the history utilization rate of SM, comprehensive utilization
The two information carry out balanced efficient scheduling to task.The balanced impact factor information of pending task is analyzed first, and
It is ranked up from small to large.The smaller distribution for illustrating the task on SM of balanced impact factor is more uniform, is unlikely to by task
The interference of migration.On the contrary, illustrating that distribution of the task on SM is more uneven if balanced impact factor is bigger, taken office
The degree of business migration interference is bigger.Based on this analysis, being easier to generate transport phenomena of the task is distributed to sharp in GPU
Not the receiving migration interference greatly of the task is distributed to the relatively low SM of utilization rate with rate relatively high SM.Reduce on total
Transport phenomena of the task in SM, and the utilization rate of SM is improved simultaneously, improve the energy optimization strategy of entirety GPU.
In specific implementation, two queues are formed, as shown in Figure 5.First queue be according to task it is balanced influence because
The queue of the sub- descending sequence of CB.The balanced impact factor of task is bigger, shows that the task is also got over by the degree of migration interference
Greatly；Second queue is the queue PHB by the descending sequence of frequency of use of SM.Frequency of use PHB value is higher, illustrates this
The frequency of use of SM is higher, and execution efficiency is also higher.Then, the equilibrium maximum task of impact factor is distributed to execution efficiency
Highest SM is executed, and thus reduces the influence as caused by task immigration.According to balanced impact factor size successively task point
The SM of the descending sequence of dispensing is executed, and realizes overall optimisation strategy.
By the way that the high task of balanced impact factor (the relatively high task of mobilance) is configured to the higher SM of execution efficiency
In (PHB value is higher), better energy consumption effect can be obtained.It comprises the concrete steps that, in the temporal information feelings for obtaining pending task
Under condition, the impact factor information for being first depending on task is ranked up task, secondly obtains the history frequency of use of SM side by side
Sequence.Then the mission number distribution recorded according to the task queue of each SM is given in corresponding SM.
4, the dispatching algorithm of CB-HRV task scheduling strategy
The specific implementation steps of CB-HRV task scheduling algorithm are as follows:
(1) the balance influence factor CB for calculating each task, calculates the utilization rate PHB of each SM resource.
(2) it resequences according to the sequence of the balance influence factor of task from big to small to each task, forms task team
Column.
(3) it resequences according to the sequence of the utilization rate PHB of SM resource from big to small to each SM.
(4) task queue of sequence is successively matched in the SM queue of sequence.
CB-HRV algorithm needs to form two queues, and one is history frequency of use queue, the other is task balance shadow
Ring the factor.And it is sorted respectively to them.The queue of task balance impact factor is indicated with TaskBalance [], is used in combination
(TaskBalance [i] function is to the queue descending sort by SortTask.History frequency of use team is indicated with SmRatio []
Column, and descending sort is carried out to the queue with SortSm (SmRatio [i]) function.For the SM history frequency of use after sequence
PHB uses array SmRationIndex [] Lai Cunfang.For the task balance impact factor sequence after sequence, array is used
TaskCBIndex [] is stored.The pseudocode of table 2CB-HRV dispatching algorithm.
Balanced impact factor-utilization rate of equipment and installations the task scheduling algorithm of table 2
CB-HRV Task Scheduling Approach
Algorithm is described as follows:
(1) 3-5 row calculates the balance influence of each task with CalculateBalance (P [i]) function first
Factor CB, and be stored in TaskBalance [i] array, then 6-8 row is calculated with CalculateRatio (SM [i])
The utilization rate ratio value of each SM resource, and be stored in SmRatio [i] array.
(2) the 9th enforcements are given with SortTask (TaskBalance [i]) according to the sequence of impact factor CB value from big to small
Each task rearrangement, and be stored in TaskCBIndex [] array, form task queue.10th row SortSm
(SmRatio [i]) function is resequenced according to the sequence of SM utilization rate from big to small to each SM, and result is stored in
In SmRationIndex [] array.
(3) 11-31 rows are successively matched to the task queue of sequence in the SM queue of sequence.According to the quantity of task
Number (P) is divided into two kinds of situations.The first situation is the quantity that task quantity is not more than SM, and 11-18 row handles the first
Situation.Directly on the task orientation after sequence to correspondingly SM.Distribution is participated in function isSMfree (i) judgement first
Whether SM can be with.If SM can participate in distribution, task is assigned to correspondingly using function mappingSM [i] (P [i])
On SM.Second situation is the quantity that task quantity is greater than SM, and 19-28 row handles second situation.Such case then appoint
Business is divided into several groups (every group of quantity is NumSM), and task is still assigned to phase using function mappingSM [j] (P [i])
It answers on ground SM.29-30 row discharges two array of indexes, to use in next sub-distribution.
Application effect of the invention is explained in detail below with reference to experiment.
In order to compare RAD, DFB, PHB and CB-HRV dispatching method, performance pair has been carried out to four kinds of dispatching methods
Than analysis.Experiment is divided into 3 parts, and first part is experimental design and experimental situation parameter declaration；The second part is
The analysis of CB-HRV algorithm harmony, the present invention compared tetra- kinds of dispatching algorithms of RAD, DFB, PHB and CB-HRV appointing on SM
The equilibrium characteristic of business distribution, and analyzed.Third part is CB-HRV algorithm energy consumption analysis, comparative analysis of the present invention
The energy consumption and performance of RAD, DFB, PHB and CB-HRV dispatching algorithm.
1, experimental design and experiment parameter explanation
(1) implementation steps tested are as follows:
(a) time of the program on each GPU is obtained.It is successively run with RAD, DFB, PHB and CB-HRV dispatching method
20 test programs, and the execution time diagram on 4 SM is obtained, it thus calculates and executes shared time scale on each SM.
(b) the above test program is successively run with RAD, DFB, PHB and CB-HRV dispatching method, and utilizes power meter
The power consumption of entire device is measured by measurement electric current and its input voltage, to obtain the power consumption of four kinds of different dispatching methods
Data.
(2) experimental situation
The software and hardware experimental situation of this experiment is as shown in table 3.The CSM that the hardware experiments environment of heterogeneous system uses is i5-
7500 processors and 4 pieces of 1060 video cards of NVIDIA GeForce GTX.Installed System Memory in hardware experiments environment is 8GB, GPU
Inside save as 6GB.GPU use framework be Pascal, it possess 10 SM (Streaming Multiprocessors, each
SM contains 128 SMDA cores, in total 1280 SMDA core, it is possible to provide 4.4TFLOPS floating-point operation ability.Software experimentation ring
Border is windows 10, VS2015 and SMDA9.2.Following table is detailed description.
Used hardware environment in the experiment of table 3
2, CB-HRV dispatching algorithm harmony is analyzed
The present invention is separately operable 20 Matrix Multiplication, histogram, scalar on GPU
Products and BlackScholes test program, and every case is obtained in the execution of the execution 1 second on 4 SM
Between scheme.
(1) Matrix Multiplication test program is analyzed
The present invention is shown the distribution condition of 4 SM task execution times.Fig. 5 illustrates 20 Matrix
Task execution timing diagram of 4 SM in one second in the case of Multiplication and its corresponding input range.
Time proportion, this hair are executed in 4 SM in order to more intuitively study Matrix Multiplication
It is bright to be shown with pie chart.
The SM maximum execution time it can be seen from Fig. 6 in RAD method accounts for the entire figure for executing the time
38.71%, the SM the smallest execution time account for the 14.66% of the entire execution time, the execution time of the two difference 24.05%.
34.79%, the SM the smallest execution time that the SM maximum execution time in DFB method accounts for the entire execution time, which accounts for, entirely holds
The 14.66% of row time, the execution time of the two difference 17.64%.The SM maximum execution time in PHB method accounts for entirely
34.79%, the SM the smallest execution time for executing the time accounts for the 18.26% of the entire execution time, the two difference 16.53%
Execute the time.The maximum execution time in CB-HRV method accounts for the 27.67% of entire execution time, the smallest execution time
Account for the 23.06% of entire execution time, the execution time of the two difference 4.61%.It bright can find out to dispatch based on CB-HRV and calculate
Than other the three kinds scheduling of task distribution of the method on 4 SM are more balanced.
(2) Histogram test program is analyzed
The present invention is shown the distribution condition of 4 SM task execution times.Fig. 6 illustrate 20 Histogram and
Task execution timing diagram of 4 SM in one second in the case of input range corresponding to it.In order to more intuitively study
Histogram executes time proportion in 4 SM, and the present invention is shown with pie chart.
The SM maximum execution time it can be seen from Fig. 7 in RAD method accounts for the entire figure for executing the time
32.88%, the SM the smallest execution time account for the 16.78% of the entire execution time, the execution time of the two difference 16.10%.
31.78%, the SM the smallest execution time that the SM maximum execution time in DFB method accounts for the entire execution time, which accounts for, entirely holds
The 18.93% of row time, the execution time of the two difference 12.85%.The SM maximum execution time in PHB method accounts for entirely
30.76%, the SM the smallest execution time for executing the time accounts for the 20.64% of the entire execution time, and the two difference 9.11% is held
The row time.The maximum execution time in CB-HRV method accounts for the 25.12% of entire execution time, and the smallest execution time accounts for
Entire to execute the 24.85% of the time, the two difference is only 0.27% execution time.Can significantly it find out based on CB-HRV
Than other the three kinds scheduling of task distribution of the dispatching algorithm on 4 SM are more balanced.
(3) Scalar products test program is analyzed
The present invention is shown the distribution condition of 4 SM task execution times.Fig. 6 illustrates 20 Scalar
Task execution timing diagram of 4 SM in one second in the case of product and its corresponding input range.In order to more intuitive
Research test program time proportion is executed in 4 SM, the present invention is shown with pie chart.
The SM maximum execution time it can be seen from Fig. 8 in RAD method accounts for the entire figure for executing the time
38.50%, the SM the smallest execution time account for the 15.29% of the entire execution time, the execution time of the two difference 23.21%.
35.57%, the SM the smallest execution time that the SM maximum execution time in DFB method accounts for the entire execution time, which accounts for, entirely holds
The 19.35% of row time, the execution time of the two difference 16.22%.The SM maximum execution time in PHB method accounts for entirely
34.58%, the SM the smallest execution time for executing the time accounts for the 19.32% of the entire execution time, the two difference 15.26%
Execute the time.The maximum execution time in CB-HRV method accounts for the 29.23% of entire execution time, the smallest execution time
Account for the 21.43% of entire execution time, the execution time of the two difference 7.80%.It can be seen that being based on CB-HRV dispatching algorithm
Than other three kinds scheduling of task distribution on 4 SM are more balanced.
(4) BlackScholes test program is analyzed
The present invention is shown the distribution condition of 4 SM task execution times.Fig. 6 illustrates 20
Task execution timing diagram of 4 SM in one second in the case of BlackScholes and its corresponding input range, such as Fig. 9.
Time proportion is executed in 4 SM in order to more intuitively study test program, the present invention is carried out with pie chart
It shows.
The SM maximum execution time it can be seen from Figure 10 in RAD method accounts for the entire figure for executing the time
38.54%, the SM the smallest execution time account for the 13.82% of the entire execution time, the execution time of the two difference 24.72%.
34.61%, the SM the smallest execution time that the SM maximum execution time in DFB method accounts for the entire execution time, which accounts for, entirely holds
The 16.92% of row time, the execution time of the two difference 17.69%.The SM maximum execution time in PHB method accounts for entirely
34.42%, the SM the smallest execution time for executing the time accounts for the 17.42% of the entire execution time, the two difference 17.00%
Execute the time.The maximum execution time in CB-HRV method accounts for the 32.32% of entire execution time, the smallest execution time
Account for the 18.36% of entire execution time, the execution time of the two difference 13.96%.It can be seen that being based on CB-HRV dispatching algorithm
Than other three kinds scheduling of task distribution on 4 SM are more balanced.
It, can be from obtaining CB-HRV dispatching method in every kind of test in data from the harmonious analysis of 4 kinds of test programs
Harmony compared to other three kinds of scheduling schemes is all best.The sequence of its improvement is successively histogram,
Matrix Multiplication, scalar products and BlackScholes, the task of this and four kinds of test programs
Migration characteristic is identical.It is 4 tests that the best reason of Histogram improvement, which is due to histogram test program,
Task immigration phenomenon is the most apparent in program.
From the ratio for executing the time, CB-HRV scheduling is dispatched compared with RAD, DFB and PHB to the evenly distributed of task
And rationally.This demonstrates the equalization characteristic of CB-HRV dispatching method, and this harmony can effectively reduce task moving in SM
Phenomenon is moved, and more effectively utilizes the computing capability of all SM.Finally, this equalization characteristic will effectively reduce the energy of GPU
Consumption.
3, CB-HRV dispatching algorithm energy consumption analysis
In order to verify the energy consumption characteristics of CB-HRV dispatching algorithm, the present invention measures different scheduling strategies and different inputs
Under the conditions of four benchmark programs average energy consumption.To RAD, the energy consumption characteristics of DFB and PHB scheduling strategy compare analysis.
Table 4 illustrates four used benchmark program parameters of experiment in detail.
Table 4
Present invention uses 18 kinds of situations to emulate to each execution sequence by the present invention.Due to this four programs for
Energy consumption and performance have strong influence to scheduler program, therefore the present invention repeatedly tests all sequences in four programs.
Then, average numerical value has been used to be compared each case.Measurement for energy consumption, in this experiment, the present invention use
Power meter is as system energy consumption measuring tool.The energy consumption that it is measured is to be input to the energy consumption of heterogeneous system.
In an experiment, the present invention has selected typical dispatching method to be compared.These methods be respectively RAD, DFB and
PHB.In an experiment, since the performance of RAD and DFB scheduling mode and system energy consumption are larger by the execution sequence of pending program.
It is obtained for this purpose, using the method for being performed a plurality of times and averaging.Since the experimental situation of this experiment is GPU all in system
All be it is identical, therefore, the sequence that program executes equally produces influences on PHB method, same using more in an experiment
The secondary method averaged that executes obtains.In an experiment, the CB-HRV implementation method of this experiment is as follows, first to this experiment institute
The method of proposition is realized in VS2015；After obtaining CB-HRV output result, the program sequence obtained after output is utilized
Column, which reprogram, to be run.
In order to verify the energy consumption characteristics of CB-HRV dispatching algorithm, the present invention is four benchmark programs in different tune
The measurement of average energy consumption is carried out under degree strategy and different input conditions.The method compared is RAD, DFB and PHB scheduling
Strategy, input condition used in scheduling strategy consider the typical input condition of typical number of tasks and each task.Under
Table 5 is the energy consumption of RAD, DFB, PHB and CB-HRV scheduling in three tests.
Table 5
It is compared in order to clearer, the present invention is respectively shown the data tested three times with histogram, in figure
Abscissa indicate different dispatching methods, ordinate indicates the energy consumption of consumption.
The incoming task number of test 1 is 24, energy consumption measurement the result is shown in Figure 13.It can be seen that the method for RAD is disappeared in figure
The average energy consumption of consumption is most, and average energy consumption consumed by DFB and PHB is suitable, the display of energy consumption minimal graph 13 consumed by CB-HRV
The energy consumptions of various methods.In general, to the RAD method phase based on 24 incoming tasks and its corresponding input range
Than CB-HRV method realizes average energy saving 13.53%；Compare the average energy conservation 5.16% of DFB；Compare PHB method, CB-HRV
Method averagely saves 5.60% energy.
The incoming task number of test 2 is that 40, Figure 14 shows that RAD method consumes most energy.DFB and PHB disappears in energy
Consumption is taken second place, and compared with RAD method, CB-HRV method realizes average energy saving 11.41%；Compare the average energy conservation of DFB
5.14%；PHB method is compared, CB-HRV method averagely saves 3.93% energy.
The energy of the energy consumption measurement the result is shown in Figure 15 of test 3, the consumption of RAD method is most, and DFB and PHB are having the same flat
Equal energy consumption, and the energy of CB-HRV consumption is minimum.Compared with RAD method, CB-HRV method realizes average energy conservation
7.97%；Compare the average energy conservation 4.95% of DFB；PHB method is compared, CB-HRV method averagely saves 4.48% energy.
In order to which the energy consumption of clearer display CB-HRV method promotes effect, the present invention is with enhancing rate comparison diagram to above
Three experiment energy consumption effects indicated, as shown in figure 16.
In conclusion under the experiment of different number of tasks and different input range situations, the CB-HRV of this experiment proposition
Method can evenly distribute task compared with RAD, DFB and PHB.From energy consumption effect, average compared with RAD method energy saving 10.97%,
Compared with the average energy conservation 5.09% of DFB, compared with the average energy conservation 4.67% of PHB.
Why the energy consumption of mean consumption is at most to only considered the quantity of task RAD method, does not consider the state of SM
And the status consideration of task, it causes and only considers number of tasks and do not consider the state of SM and execute the time and then consume more
Energy consumption.Average energy consumption consumed by DFB and PHB is almost suitable, the reason is that causing in the environment of identical GPU in PHB method
It is DFB method that PHB dispatching method, which is degenerated,.The use state that the SM in system is considered due to DFB and PHB, in task
It is better than RAD method in distribution, the consumption of the energy consumption of DFB method and PHB method is less than RAD method in figure.
It can be seen from the figure that CB-HRV method is more preferable than the energy-saving effect of other three kinds of methods, this is attributed to the fact that CB-
The harmony of HRV algorithm.Distribution due to task in each SM is more rationally effective, and it is existing to reduce migration of the task between SM
As so that new algorithm achieves better energy optimization effect.Experimental result illustrates the CB-HRV method that this experiment proposes
With more task balance characteristic, energy consumption and performance is also more preferable.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (6)
1. energy consumption optimization method inside a kind of GPU of task based access control balance dispatching, which is characterized in that the task based access control equilibrium tune
Energy consumption optimization method utilizes balanced impact factor CB and SM the frequency of use information PHB handle of pending program inside the GPU of degree,
It is proposed the CB-HRV task scheduling algorithm of balanced impact factor and SM frequency of use information；CB-HRV task scheduling algorithm foundation
One group task is assigned in corresponding SM according to the balanced impact factor of task and executes by the time of task execution；Strategies scheduler
It task balance is assigned in each SM；Meanwhile the service efficiency of device resource is improved using SM frequency of use information；It is logical
Cross the method for scheduling task that energy optimization is realized to the integration application of pending mission bit stream and resource information.
2. energy consumption optimization method inside the GPU of task based access control balance dispatching as described in claim 1, which is characterized in that described
CB-HRV task scheduling algorithm specifically includes:
(1) the balance influence factor CB for calculating each task, calculates the utilization rate PHB of each SM resource；
(2) it resequences according to the sequence of the balance influence factor of task from big to small to each task, forms task queue；
(3) it resequences according to the sequence of the utilization rate PHB of SM resource from big to small to each SM；
(4) task queue of sequence is successively matched in the SM queue of sequence.
3. energy consumption optimization method inside the GPU of task based access control balance dispatching as claimed in claim 2, which is characterized in that described
CB-HRV task scheduling algorithm formed two queues, one is history frequency of use queue, the other is task balance influence because
Son；
The queue of task balance impact factor is indicated with TaskBalance [], and with SortTask (TaskBalance [i] function pair
The queue descending sort；The queue of history frequency of use is indicated with SmRatio [], and with SortSm (SmRatio [i]) function pair
The queue carries out descending sort；For the SM history frequency of use PHB after sequence, deposited using array SmRationIndex []
It puts；For the task balance impact factor sequence after sequence, stored using array TaskCBIndex [].
4. energy consumption optimization method inside the GPU of task based access control balance dispatching as claimed in claim 2, which is characterized in that described
CB-HRV task scheduling algorithm further comprises:
(1) the balance influence factor CB of each task is calculated with CalculateBalance (P [i]) function first, and stored
In TaskBalance [i] array, the utilization rate of each SM resource is then calculated with CalculateRatio (SM [i])
Ratio value, and be stored in SmRatio [i] array；
(2) each task weight is given according to the sequence of impact factor CB value from big to small using SortTask (TaskBalance [i])
New sort, and be stored in TaskCBIndex [] array, form task queue；With SortSm (SmRatio [i]) function according to
The sequence of SM utilization rate from big to small is resequenced to each SM, and result is stored in SmRationIndex [] array；
(3) task queue of sequence is successively matched in the SM queue of sequence.
5. energy consumption optimization method inside the GPU of task based access control balance dispatching as described in claim 1, which is characterized in that described
Energy consumption optimization method inside the GPU of task based access control balance dispatching further include:
The first step, analyzes the balanced impact factor information of pending task, and is ranked up from small to large；Being easier to produce
The task of raw transport phenomena distributes to the relatively high SM of utilization rate in GPU and not the receiving migration interference greatly of the task is distributed to utilization
The relatively low SM of rate；
Second, two queues are formed, first queue is the team according to the balanced descending sequence of impact factor CB of task
Column；The balanced impact factor of task is bigger, shows that the task is also bigger by the degree of migration interference；Second queue is by SM
The queue PHB of the descending sequence of frequency of use；Frequency of use PHB value is higher, illustrates that the frequency of use of the SM is higher, executes effect
Rate is also higher；
The equilibrium maximum task of impact factor is distributed to the highest SM of execution efficiency and executed by third step, thus come reduce due to
It is influenced caused by task immigration；Successively the SM that task distributes to descending sequence is executed according to balanced impact factor size,
Realize overall optimisation strategy；
4th step is obtaining pending by the way that the high task of balanced impact factor is configured in the higher SM of execution efficiency
In the case of the temporal information of business, the impact factor information according to task is ranked up task, and the history for secondly obtaining SM uses
Frequency simultaneously sorts；The mission number distribution recorded according to the task queue of each SM is given in corresponding SM.
6. energy consumption optimization method inside a kind of GPU using task based access control balance dispatching described in Claims 1 to 5 any one
Graphics processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910205801.4A CN109992385B (en) | 2019-03-19 | 2019-03-19 | GPU internal energy consumption optimization method based on task balance scheduling |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910205801.4A CN109992385B (en) | 2019-03-19 | 2019-03-19 | GPU internal energy consumption optimization method based on task balance scheduling |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109992385A true CN109992385A (en) | 2019-07-09 |
| CN109992385B CN109992385B (en) | 2021-05-14 |
Family
ID=67130518
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910205801.4A Expired - Fee Related CN109992385B (en) | 2019-03-19 | 2019-03-19 | GPU internal energy consumption optimization method based on task balance scheduling |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109992385B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110532071A (en) * | 2019-07-12 | 2019-12-03 | 上海大学 | A kind of more application schedules system and method based on GPU |
| CN110991732A (en) * | 2019-11-28 | 2020-04-10 | 武汉理工大学 | Building material equipment manufacturing process optimization scheduling method based on energy consumption clustering |
| CN111240461A (en) * | 2020-01-09 | 2020-06-05 | 黔南民族师范学院 | Task scheduling-based heterogeneous computing system low-power consumption method |
| CN111309475A (en) * | 2020-01-21 | 2020-06-19 | 上海悦易网络信息技术有限公司 | Detection task execution method and device |
| CN113157407A (en) * | 2021-03-18 | 2021-07-23 | 浙大宁波理工学院 | Dynamic task migration scheduling method for parallel processing of video compression in GPU |
| CN113448705A (en) * | 2021-06-25 | 2021-09-28 | 皖西学院 | Unbalanced job scheduling algorithm |
| CN115237605A (en) * | 2022-09-19 | 2022-10-25 | 四川大学 | Data transmission method between CPU and GPU and computer equipment |
| CN117573321A (en) * | 2023-11-24 | 2024-02-20 | 数翊科技(北京)有限公司 | Periodic scheduling task dynamic arrangement method and system based on database |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101901042A (en) * | 2010-08-27 | 2010-12-01 | 上海交通大学 | Method for reducing power consumption based on dynamic task migrating technology in multi-GPU (Graphic Processing Unit) system |
| US9268601B2 (en) * | 2010-04-05 | 2016-02-23 | Nvidia Corporation | API for launching work on a processor |
| US9286114B2 (en) * | 2012-12-13 | 2016-03-15 | Nvidia Corporation | System and method for launching data parallel and task parallel application threads and graphics processing unit incorporating the same |
| CN106502632A (en) * | 2016-10-28 | 2017-03-15 | 武汉大学 | A kind of GPU parallel particle swarm optimization methods based on self-adaptive thread beam |
| CN108121433A (en) * | 2017-12-06 | 2018-06-05 | 中国航空工业集团公司西安航空计算技术研究所 | A kind of method for scheduling task of GPU multicomputer systems |
| CN109445565A (en) * | 2018-11-08 | 2019-03-08 | 北京航空航天大学 | A kind of GPU QoS guarantee method exclusive and reserved based on stream multiple processor cores |
-
2019
- 2019-03-19 CN CN201910205801.4A patent/CN109992385B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9268601B2 (en) * | 2010-04-05 | 2016-02-23 | Nvidia Corporation | API for launching work on a processor |
| CN101901042A (en) * | 2010-08-27 | 2010-12-01 | 上海交通大学 | Method for reducing power consumption based on dynamic task migrating technology in multi-GPU (Graphic Processing Unit) system |
| US9286114B2 (en) * | 2012-12-13 | 2016-03-15 | Nvidia Corporation | System and method for launching data parallel and task parallel application threads and graphics processing unit incorporating the same |
| CN106502632A (en) * | 2016-10-28 | 2017-03-15 | 武汉大学 | A kind of GPU parallel particle swarm optimization methods based on self-adaptive thread beam |
| CN108121433A (en) * | 2017-12-06 | 2018-06-05 | 中国航空工业集团公司西安航空计算技术研究所 | A kind of method for scheduling task of GPU multicomputer systems |
| CN109445565A (en) * | 2018-11-08 | 2019-03-08 | 北京航空航天大学 | A kind of GPU QoS guarantee method exclusive and reserved based on stream multiple processor cores |
Non-Patent Citations (2)
| Title |
|---|
| XULONG TANG,ASHUTOSH PATTNAIK,HUAIPAN JIANG,ONUR KAYIRAN: "Controlled Kernel Launch for Dynamic Parallelism in GPUs", 《2017 IEEE INTERNATIONAL SYMPOSIUM ON HIGH PERFORMANCE COMPUTER ARCHITECTURE 》 * |
| 李君科，郭兵，沈艳，李德光，黄彦辉，戚正伟: "GPU内置传感器的功耗数据矫正方法研究", 《电子科技大学学报》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110532071B (en) * | 2019-07-12 | 2023-06-09 | 上海大学 | Multi-application scheduling system and method based on GPU |
| CN110532071A (en) * | 2019-07-12 | 2019-12-03 | 上海大学 | A kind of more application schedules system and method based on GPU |
| CN110991732A (en) * | 2019-11-28 | 2020-04-10 | 武汉理工大学 | Building material equipment manufacturing process optimization scheduling method based on energy consumption clustering |
| CN110991732B (en) * | 2019-11-28 | 2023-06-16 | 武汉理工大学 | Building material equipment manufacturing process optimization scheduling method based on energy consumption clustering |
| CN111240461A (en) * | 2020-01-09 | 2020-06-05 | 黔南民族师范学院 | Task scheduling-based heterogeneous computing system low-power consumption method |
| CN111309475A (en) * | 2020-01-21 | 2020-06-19 | 上海悦易网络信息技术有限公司 | Detection task execution method and device |
| CN111309475B (en) * | 2020-01-21 | 2022-12-02 | 上海万物新生环保科技集团有限公司 | Detection task execution method and equipment |
| CN113157407A (en) * | 2021-03-18 | 2021-07-23 | 浙大宁波理工学院 | Dynamic task migration scheduling method for parallel processing of video compression in GPU |
| CN113157407B (en) * | 2021-03-18 | 2024-03-01 | 浙大宁波理工学院 | Dynamic task migration scheduling method for parallel processing video compression in GPU |
| CN113448705B (en) * | 2021-06-25 | 2023-03-28 | 皖西学院 | Unbalanced job scheduling algorithm |
| CN113448705A (en) * | 2021-06-25 | 2021-09-28 | 皖西学院 | Unbalanced job scheduling algorithm |
| CN115237605A (en) * | 2022-09-19 | 2022-10-25 | 四川大学 | Data transmission method between CPU and GPU and computer equipment |
| CN117573321A (en) * | 2023-11-24 | 2024-02-20 | 数翊科技(北京)有限公司 | Periodic scheduling task dynamic arrangement method and system based on database |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109992385B (en) | 2021-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109992385A (en) | A kind of inside GPU energy consumption optimization method of task based access control balance dispatching | |
| Kim et al. | Racing and pacing to idle: Theoretical and empirical analysis of energy optimization heuristics | |
| Selvarani et al. | Improved cost-based algorithm for task scheduling in cloud computing | |
| Wen et al. | Energy and cost aware scheduling with batch processing for instance-intensive IoT workflows in clouds | |
| Quarati et al. | Hybrid clouds brokering: business opportunities, QoS and energy-saving issues | |
| Shieh et al. | Energy and transition-aware runtime task scheduling for multicore processors | |
| Kandhalu et al. | Energy-aware partitioned fixed-priority scheduling for chip multi-processors | |
| Liu et al. | A waterfall model to achieve energy efficient tasks mapping for large scale GPU clusters | |
| Sartori et al. | Distributed peak power management for many-core architectures | |
| Hao et al. | An adaptive algorithm for scheduling parallel jobs in meteorological Cloud | |
| Hao et al. | Adaptive energy-aware scheduling method in a meteorological cloud | |
| CN111240461B (en) | Heterogeneous computing system low-power consumption method based on task scheduling | |
| Liu et al. | Energy-efficient scheduling of real-time tasks on heterogeneous multicores using task splitting | |
| Sheikh et al. | Energy-efficient cache-aware scheduling on heterogeneous multicore systems | |
| CN107589993A (en) | A kind of dynamic priority scheduling algorithm based on linux real time operating systems | |
| Guyon et al. | How much energy can green HPC cloud users save? | |
| Ahmed et al. | An energy efficient demand-response model for high performance computing systems | |
| CN105426247A (en) | HLA federate planning and scheduling method | |
| Hu et al. | Energy-minimized scheduling of intermittent real-time tasks in a CPU-GPU cloud computing platform | |
| Huang et al. | GPU Energy optimization based on task balance scheduling | |
| Guyon et al. | Involving users in energy conservation: A case study in scientific clouds | |
| Xu et al. | Actor-critic with transformer for cloud computing resource three stage job scheduling | |
| Eitschberger et al. | Energy-efficient task scheduling in manycore processors with frequency scaling overhead | |
| Eyerman et al. | Revisiting symbiotic job scheduling | |
| Qouneh et al. | Optimization of resource allocation and energy efficiency in heterogeneous cloud data centers |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210514 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |