WO2014015725A1 - 基于应用效果即时反馈的显卡虚拟化下资源调度系统、方法 - Google Patents

基于应用效果即时反馈的显卡虚拟化下资源调度系统、方法 Download PDF

Info

Publication number
WO2014015725A1
WO2014015725A1 PCT/CN2013/077457 CN2013077457W WO2014015725A1 WO 2014015725 A1 WO2014015725 A1 WO 2014015725A1 CN 2013077457 W CN2013077457 W CN 2013077457W WO 2014015725 A1 WO2014015725 A1 WO 2014015725A1
Authority
WO
WIPO (PCT)
Prior art keywords
graphics card
scheduling
application
physical
executor
Prior art date
Application number
PCT/CN2013/077457
Other languages
English (en)
French (fr)
Inventor
于淼
戚正伟
管海兵
王胤
Original Assignee
上海交通大学
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 上海交通大学 filed Critical 上海交通大学
Priority to US15/563,951 priority Critical patent/US10922140B2/en
Publication of WO2014015725A1 publication Critical patent/WO2014015725A1/zh

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]
    • G06F9/5005—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
    • 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]
    • G06F9/5061—Partitioning or combining of resources
    • G06F9/5077—Logical partitioning of resources; Management or configuration of virtualized 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
    • G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38—Concurrent instruction execution, e.g. pipeline or look ahead
    • G06F9/3836—Instruction issuing, e.g. dynamic instruction scheduling or out of order instruction execution
    • 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
    • G06F9/44—Arrangements for executing specific programs
    • G06F9/455—Emulation; Interpretation; Software simulation, e.g. virtualisation or emulation of application or operating system execution engines
    • G06F9/45533—Hypervisors; Virtual machine monitors
    • G06F9/45558—Hypervisor-specific management and integration aspects
    • 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
    • G06F9/4881—Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T1/00—General purpose image data processing
    • G06T1/20—Processor architectures; Processor configuration, e.g. pipelining
    • 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
    • G06F9/44—Arrangements for executing specific programs
    • G06F9/455—Emulation; Interpretation; Software simulation, e.g. virtualisation or emulation of application or operating system execution engines
    • G06F9/45533—Hypervisors; Virtual machine monitors
    • G06F9/45558—Hypervisor-specific management and integration aspects
    • G06F2009/45575—Starting, stopping, suspending or resuming virtual machine instances
    • 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
    • G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/30003—Arrangements for executing specific machine instructions
    • G06F9/30072—Arrangements for executing specific machine instructions to perform conditional operations, e.g. using predicates or guards
    • G—PHYSICS
    • G09—EDUCATION; CRYPTOGRAPHY; DISPLAY; ADVERTISING; SEALS
    • G09G—ARRANGEMENTS OR CIRCUITS FOR CONTROL OF INDICATING DEVICES USING STATIC MEANS TO PRESENT VARIABLE INFORMATION
    • G09G5/00—Control arrangements or circuits for visual indicators common to cathode-ray tube indicators and other visual indicators
    • G09G5/36—Control arrangements or circuits for visual indicators common to cathode-ray tube indicators and other visual indicators characterised by the display of a graphic pattern, e.g. using an all-points-addressable [APA] memory
    • G09G5/363—Graphics controllers

Definitions

  • the present invention relates to a system and method applied to the field of computer application technology, in particular to a physical graphics card (GPU) scheduling system method between virtual machines based on application immediate effect feedback, specifically A resource scheduling system based on virtualization of graphics cards based on instant feedback of application effects.
  • GPU graphics card
  • GPU Virtualization is being widely used in data centers that perform GPU Computing, including but not limited to: Cloud Gaming, Video Rendering, and Generic Graphics Computing (General Purpose GPU Computing).
  • VGA Passthrough Graphics Graphics Array Passthrough assigns each available physical graphics card to each virtual machine running.
  • VGA Passthrough assigns each available physical graphics card to each virtual machine running.
  • the disadvantages of this method are as follows: First, the general commercial motherboard only supports two to three graphics cards, so a special motherboard is required to simultaneously run multiple virtual machines that require graphics card support; second, for each virtual machine, inside During its operation, it usually cannot use up the physical graphics resources it owns. In this technology, the remaining graphics resources cannot be given to other virtual machines, thus causing waste of physical graphics resources.
  • GPU Virtualization on VM ware's Hosted I/O Architecture was published in Volume 43 Issue 3 of the SIGOPS Operating Systems Review in 2009. Later, in 2008, Multimedia Computing and Networking ⁇ Bautin ⁇ et al. proposed a scheduling strategy for the division of physical graphics resources among multiple applications in Graphic Engine Resource Management. Then in the Usenix ATC in 2011, Kato et al. in the Timegraph: GPU scheduling for real-time multi-tasking environments paper proposed to increase the physical graphics acceleration user by introducing the use of graphics resources and modifying the operating system graphics driver. The ability of the program.
  • the above two methods can maximize the utilization of available physical graphics resources while providing the ability to accelerate graphics for multiple virtual machines.
  • the disadvantages of these methods are: On the one hand, you need to modify the operating system or the graphics card driver. When you apply to a virtual machine, you even need to modify the hypervisor or the guest application in the virtual machine.
  • the existing method can not obtain the accelerated Guest Application running effect feedback data, the existing physical GPU resource scheduling system and method are blind, and the obtained resource scheduling effect is one - invention content
  • the present invention is directed to the above-mentioned deficiencies of the prior art, and provides a physical graphics resource scheduling system and method between virtual machines based on application immediate effect feedback.
  • the traditional GPU virtualization technology sends the graphics card command and data in the virtual machine to the host physical graphics application interface (Host GPU API) through the host physical graphics card instruction transmitter (GPU HostOps Dispatch).
  • the present invention provides The method inserts a dispatching executor (Agent) between the GPU HostOps Dispatch and the Host GPU API by means of a function hook, delays the transmission of instructions and data in the GPU HostOps Dispatch, and monitors the display performance related to the Guest Application and the usage of the physical graphics card resources.
  • Agent dispatching executor
  • the time- or timing-based graphics card resource scheduling algorithm refers to the beginning, end, and use of the graphics card resources, either partially or completely based on absolute or relative time.
  • the system described in the present invention immediately accepts the user's decision to enable or disable each agent through the Scheduling Controller, and selects and selects the scheduling method to be used, and changes the corresponding parameter settings of each agent according to the same.
  • the Scheduling Controller displays or records one or more contents of the current physical graphics card resource scheduling and usage, and the usage of the application graphics card resources in each virtual machine.
  • the present invention employs Advanced Prediction (Advanced Prediction) to delay the transmission of instructions and data in the GPU HostOps Dispatch to achieve precise control of Frame Latency.
  • This advanced prediction technique includes Frame Rendering Performance Prediction and Flush Single Queued Frame:.
  • the render/display command queue frame-by-frame advancement technique also includes a Mark Flush Frame and a Commit Flush Frame:.
  • the tag operation is optional and is used to mark a frame of the virtual machine in the render/display command queue (including but not limited to the previous frame or the first few frames) to indicate the frame that can be moved out of the graphics buffer (including But not limited to the frame forced display).
  • the commit operation forces a frame (which is the corresponding marked frame after the markup operation) to move out of the physical graphics card buffer, leaving the physical graphics buffer with free space.
  • the system and method described herein do not require modification of the host operating system, host graphics driver, Hyper V i SOT , virtual machine operating system, virtual machine graphics driver, or applications within the virtual machine. Moreover, the system and method described herein achieves a performance overhead of less than 5% at runtime, enabling or stopping without incurring significant virtual machine pause times (only millisecond pauses are required).
  • a resource scheduling system for video card virtualization based on instant feedback of application effects including a host physical graphics card instruction transmitter, a host physical graphics application
  • the interface also includes the following modules: a scheduling executor connected between the host physical graphics card instruction transmitter and the host physical graphics application interface; a scheduling controller connected to the scheduling executor,
  • the scheduling controller receives the user command and passes the user command to the scheduling executor; the scheduling executor receives the user command from the scheduling controller, monitors the running state of the application, and transmits the application graphics card status result to the scheduling controller At the same time, according to the scheduling algorithm specified by the scheduling controller, the periodic/event computing needs to meet the delay time required for the lowest application graphics card state, delaying the instruction and data transmission of the host physical graphics instruction transmitter to the host physical graphics application.
  • Program interface The dispatch controller receives and processes the display of scheduling results and scheduling status from the dispatch executor.
  • the scheduling controller receives the user command, parses the operation of each scheduling executor, the configuration of the scheduling algorithm, and corresponding parameters, and passes the user command to the scheduling executor module, and receives the status result from the scheduling executor module. Displayed to the user.
  • the scheduling controller comprises the following modules:
  • a console for receiving user commands, which input configuration and corresponding parameters about the scheduling algorithm, and obtain scheduling results from the scheduling communicator and display them to the user;
  • a scheduling communicator for communicating with the dispatch controller and one or more dispatch executors, for installing/unloading the dispatch executor, passing the user command to the dispatch executor, and receiving the application graphics status from the dispatch executor result.
  • the scheduling executor comprises a module, a scheduler, configured to receive a designation of a scheduling algorithm and a parameter configuration thereof in a user command, and is responsible for finding a location of a corresponding scheduling algorithm, configuring a scheduling algorithm, and running a corresponding scheduling.
  • Algorithm delaying the instruction and data sent by the host physical graphics instruction transmitter to the host physical graphics application interface on demand;
  • An application graphics card status monitor which is used to collect the status of the graphics card from the host physical graphics application interface, and thereby generate the application graphics status result, and simultaneously feed the application graphics status report to the scheduler and pass it to the scheduling control. Dispatched communicator in the device.
  • the application graphics state comprises: a graphics card physical state and/or a logic state metric associated with the application category.
  • Physical state measures such as the graphics card's load factor (GPU Load), temperature, voltage, and so on.
  • Logic state measure for computer 3D games, the corresponding graphics state measurement is Frames per second (FPS). For computer general GPU operations, the corresponding graphics state measurement is calculated per second (Ops), The application's graphics GPU Usage and so on.
  • a graphics card resource scheduling method for video card virtualization in a resource scheduling system is provided, and a physical hook instruction method transmitter and a host physical graphics card application are implemented by a function hook method. Inserting a scheduler (Agent) between interfaces, delaying the instructions and data sent by the host physical graphics card to the host
  • the physical graphics application interface which monitors the application-related display performance and physical graphics resource usage, and provides feedback to any time or timing-based graphics resource scheduling algorithm, without the need for virtual machine applications, host operating systems,
  • the virtual machine operating system, graphics driver, and virtual machine manager make any modifications and have low performance loss.
  • the specific process of the method is: after one or more virtual machines are started, when the client needs to install the resource scheduling system, the method executed by the application is searched by the scheduling controller or scheduled to execute according to a user-specified process.
  • the device is bound to the corresponding virtual machine; then the scheduling communicator in the scheduling controller establishes communication with each of the bound scheduling executors; when scheduling the graphics card resources, the client issues an instruction to select a scheduling algorithm (may be a third-party development) Scheduling algorithm) and provide corresponding parameters, the console distributes the user command to each dispatching executor by the dispatching communicator after receiving the client instruction; the scheduling executor runs the selected graphics card resource scheduling algorithm according to the user command, delaying the host physics
  • the instruction and data in the graphics card instruction transmitter are sent to the host physical graphics application interface; at the same time, the application graphics status monitor collects the status of the graphics card from the host physical graphics application interface, and thereby generates the application graphics status.
  • Subsequent periodicity/event will The program card status result is fed back to the scheduler and passed to the dispatch communicator in the dispatch controller; when the client needs to uninstall the resource dispatch system, the client issues an uninstall command through the dispatch controller, and the console receives the client command.
  • the user command is distributed by the dispatch communicator to each dispatch executor, and the dispatch executor stops its own operation after receiving the uninstall command.
  • the graphics card resource usage method uses an advanced prediction method in the graphics card resource scheduling method, and cooperates with the transmission of the instruction and data in the transmitter of the delayed host physical graphics card instruction to achieve the purpose of accurately controlling the interframe delay, and the advanced prediction method of the graphics card resource is used.
  • the rendering/display overhead prediction is to predict the current consumption time of the physical graphics card resource according to the physical graphics resource consumption time history corresponding to the host physical graphics application interface;
  • the rendering/display command queue advances the frame-by-frame operation including the mark operation and the submit operation, wherein the mark operation is an optional operation for using a frame of the virtual machine in the render/display command queue (including but not limited to the previous frame or before)
  • the mark operation is an optional operation for using a frame of the virtual machine in the render/display command queue (including but not limited to the previous frame or before)
  • a number of frames are marked to indicate the frame that can be removed from the graphics card buffer (including but not limited to the frame forcibly displayed), and the commit operation forces a frame (before the tagged operation is executed to correspond to the marked frame) to move out of the physical graphics buffer. , so that the physical graphics card buffer has free space.
  • the step of binding the scheduling executor to the corresponding virtual machine is specifically - step 1.1, searching for a specified virtual machine image rendering process according to the user specified information (depending on the virtual machine manager design, the processes may also be Virtual machine process), or select all relevant virtual machine image rendering processes, perform step 1.2 to step 1.6 for each of these virtual machine processes;
  • Step 1.2 create a new thread (Thread) in the process, and load a schedule executor therein;
  • Step 1.3 accessing the schedule executor entry, initializing the schedule executor
  • Step 1.4 Find a set of physical interface application interface addresses of the host loaded by the process, and modify each of the hosts.
  • Step 1.5 After the processing function returns the address of the old host physical graphics application interface address, the instruction is executed, and the contents of each register are restored, so that the original host physical graphics application interface can be correctly executed after the processing function is finished; Step 1.6 , the thread must not end.
  • the step of delaying the transmission of instructions and data from the host physical graphics instruction transmitter to the host physical graphics application interface is specifically -
  • Step 2.1a in the processing function specified by the resource scheduling algorithm, predicting the current consumption time of the physical graphics card resource according to the physical video card resource consumption time history corresponding to the physical graphics application interface of the host, and stopping the timing of the central processing unit ( Central Processing Unit, CPU)
  • CPU Central Processing Unit
  • Step 2.2a suspending the execution of the central processing unit for a period of time, wherein the length of the period is calculated by the scheduling algorithm according to the current consumption time of the central processing unit and the predicted consumption time of the physical graphics card resource;
  • Step 2.3a start timing the physical graphics card resource consumption time
  • Step 2.4a calling the original host physical graphics application interface
  • Step 2.5a stopping the timing of the physical graphics card resource consumption time, and updating the current physical graphics card resource consumption time to the physical graphics card resource consumption time history corresponding to the host physical graphics application interface;
  • Step 2.1b in the processing function specified by the resource scheduling algorithm, stop timing the current CPU consumption time, and start timing the physical graphics card resource consumption time;
  • Step 2.2b calling the original host physical graphics application interface
  • Step 2.3b stop timing the physical graphics card resource consumption time
  • Step 2.4b suspending the central processing unit for a period of time, the length of the period is calculated by the scheduling algorithm according to the current consumption time of the central processing unit and the current consumption time of the physical graphics card resource;
  • step 2.1c in the processing function specified by the resource scheduling algorithm, the rendering/display command queue commit operation of the frame-by-frame push operation is performed, which forcibly A frame (such as a tagged frame before the markup operation) is removed, so that the physical graphics card buffer has free space; stop timing the central processor consumes time;
  • step 2.2c the rendering/display overhead prediction is based on the physical graphics card corresponding to the host physical graphics application interface.
  • the source consumption time history predicts the current consumption time of the physical graphics card resource;
  • Step 2.3c suspending the execution of the central processing unit for a period of time, wherein the length of the period is calculated by the scheduling algorithm according to the current consumption time of the central processing unit and the predicted consumption time of the physical graphics card resource;
  • Step 2.4c start timing the physical graphics card resource consumption time
  • Step 2.5c calling the original host physical graphics application interface
  • Step 2.6c stop timing the physical graphics card resource consumption time
  • Step 2.7c start timing the next CPU consumption time, and optionally perform a marking operation of the rendering/display command queue frame-by-frame advance operation, which is a frame of the virtual machine in the rendering/display command queue (including but not limited to the previous frame) Or the first few frames) are marked to indicate the frame that can be moved out of the graphics card buffer (including but not limited to the frame forced display).
  • the application graphics card status monitor collects the status of the graphics card from the physical graphics application interface of the host, specifically - step 3.1, in the processing function specified by the resource scheduling algorithm, invokes the host physical graphics application Interfaces, operating system kernels, or graphics card drivers provide interfaces to collect graphics card status according to resource scheduling algorithm requirements and user commands, such as graphics card load rate (GPU Load), temperature, voltage, FPS, Ops, graphics card load rate of the application, etc.
  • GPU Load graphics card load rate
  • Step 3.2 In the processing function specified by the resource scheduling algorithm, invoke the original host physical graphics application interface; preferably, the step of generating the status of the application graphics card is specifically:
  • Step 4.1 The user specifies a status reporting frequency, and the status reporting frequency is obtained in the scheduling executor;
  • Step 4.2 When the status reporting time point arrives, the application graphics card status monitor in the scheduling executor transmits the accumulated status result to the scheduling communicator in the scheduling controller;
  • Step 4.3 the scheduling executor clears its own status result buffer
  • the step of the scheduling executor stopping the operation after receiving the uninstallation instruction is specifically - step 5.1, after each scheduling executor receives the uninstallation instruction, starting the uninstallation process from step 5.2 to step 5.3; Restore the host physical card application interface address set loaded by the process, modify the code at the address of each host physical graphics application interface to the content of the original application interface address, so that the process will use the host physical graphics card each time later.
  • the original application interface logic is run when the application interface is used;
  • Step 5.3 Binding the scheduling executor to the end of the thread inserted in the corresponding virtual machine, thereby uninstalling the scheduling executor;
  • the resource scheduling algorithm specifically includes the following steps - Step 6.1, for the virtual machine group VM1, VM2- To VMn, the scheduler in the scheduling executor of each virtual machine parses the user method configuration, and obtains the load rate of the minimum graphics card to be satisfied, and the minimum number of frames per second (the scope of application of the patent is not limited to Computer games, for other graphics applications, can be measured for different states), user-specified detection period T;
  • Step 6.2 during the operation, the processing function will be called multiple times.
  • the prediction method technique
  • step 2.1b to step 2.5a
  • step 2.4b use the prediction method (technique) to perform step 2.1b to Step 2.4b ;
  • Step 6.3 For each T cycle, if a virtual machine VMm does not satisfy the state measurement, find and reduce the minimum number of frames per second of the virtual machine having the largest minimum number of frames per second; reduce the number of frames per second Depending on the application GPU Load of the application of the last few frames, the number of frames per second and the load rate of the application graphics cards of the most recent frames are linear;
  • Step 6.4 For each T cycle, if the physical graphics usage rate does not meet the minimum graphics card loading rate, increase the minimum number of frames per second for all virtual machines; increase the number of frames per second depending on the application of the last few frames.
  • the graphics card load rate, the number of frames per second, and the graphics card load rate of the application of the most recent frames are linear;
  • Step 6.5, Step 6.2 to Step 6.4 remain valid until the user specifies the method to end or replace the method or uninstall the scheduler.
  • the GPU HostOps Dispatch corresponding to each virtual machine is installed in the present invention by a separately owned scheduling executor.
  • a globally unique dispatch controller is connected to one or more dispatch executors.
  • the present invention has the following advantages over the prior art: First, there is no need to make any modifications to the virtual machine application, the host operating system, the virtual machine operating system, the graphics card driver, and the virtual machine manager. Existing systems typically require extensive modification of one of the above to achieve similar scheduling capabilities, which can result in existing systems having to evolve to be compatible with the latest applications, operating systems, or graphics drivers.
  • the present invention does not require a pause in the operation of the machine during installation or unloading. This feature makes the system easy to deploy in commercial systems, especially for commercial servers requiring 7X 24 hour availability.
  • the present invention has extremely high performance when the performance of the graphics card resource scheduling capability between virtual machines is greatly improved, and the overall performance loss is less than 5%.
  • Figure 1 is a schematic view of the module of the present invention.
  • FIG. 2 is a schematic diagram of the architecture of the present invention.
  • a resource scheduling system for video card virtualization based on application effect instant feedback includes: a scheduling executor module and a scheduling controller module, wherein: a scheduling controller module and a scheduling executor module Connect, pass user commands to the dispatch executor and receive the status of the graphics card status it returns.
  • the dispatch executor is inserted into the host Between the graphics card sender GPU HostOps Dispatch and the host physical graphics application interface Host GPU API, the corresponding call and data are delayed.
  • the dispatch executor module is also responsible for collecting the physical state of the graphics card and/or the logic state measure using the Host GPU API of the host physical graphics application interface. This embodiment is directed to a computer game running in a virtual machine, and thus the collected physical state and logical state include: Application GPU Load and FPS.
  • the scheduling controller module includes: a console sub-module and a scheduling communicator sub-module, wherein: the control station sub-module is configured to receive user commands, where the commands input configuration and corresponding parameters about the scheduling algorithm.
  • the console sub-module periodically/events gets the scheduling results from the scheduling communicator sub-module and displays them to the user.
  • the dispatch communicator sub-module is responsible for scheduling the controller module to communicate with one or more dispatch executor modules, and is responsible for installing/uninstalling the dispatch executor, passing user commands to the dispatch executor sub-module, and the like.
  • the eventuality refers to the occurrence of a target event one or more times, but the time interval of occurrence is not constant, and the distribution of events in time can be expressed mathematically as a time series of non-periodic nature.
  • the scheduling executor module includes: a scheduler submodule, an application graphics card status monitor submodule, where: a scheduler submodule receives a designation of a scheduling algorithm and its parameter configuration in a user command, and responsible for running the corresponding scheduling algorithm according to the configuration, delaying the instruction and data in the GPU HostOps Dispatch to the Host GPU API as needed.
  • the application graphics status monitor sub-module is responsible for collecting the status of the graphics card from the Host GPU API, and thereby generating the application graphics status, and then periodically/evently feeding back the application graphics status results to the scheduler sub-module and passing it to the scheduler.
  • a dispatch communicator submodule in the controller module is responsible for collecting the status of the graphics card from the Host GPU API, and thereby generating the application graphics status, and then periodically/evently feeding back the application graphics status results to the scheduler sub-module and passing it to the scheduler.
  • the application graphics state refers to the physical state of the graphics card and/or the logical state measure associated with the type of application.
  • the collected physical state and logical state include: Application GPU Load and FPS.
  • This embodiment is directed to the VMWare Player 4.0 virtual machine manager system, thus specifying the virtual machine image rendering process, i.e., the virtual machine process.
  • the virtual machine image rendering process i.e., the virtual machine process.
  • only the user selects all relevant virtual machine image rendering processes.
  • Step 1 The user selects all related virtual machine processes, and performs step two to step six for each of the virtual machine processes:
  • step two a new thread (Thread) is created in the process, and the schedule executor module is loaded therein.
  • Thread a new thread
  • Step 3 Access the scheduling executor module entry, and initialize the scheduling executor module.
  • Step 4 Find the host physical card application interface address set loaded by the process, modify the code of each host physical graphics application interface address, point it to the entry of the corresponding processing function in the scheduling executor module, and save each Register contents. Causes the process to run the handler function each time the host physical graphics application interface is used (Handlers).
  • Step 5 After the processing function returns the address of the old host physical graphics application interface address, the instruction is run, and the contents of each register are restored, so that the original host physical graphics application interface can be correctly executed after the processing function is finished.
  • Step six the thread must not end.
  • the scheduling executor module is bound to the corresponding virtual machine, and after the scheduling communicator sub-module in the scheduling controller module establishes communication with each bound scheduling executor module, the scheduling executor module can transmit the status result. Go to the dispatch controller module and respond to the user commands issued by the dispatch controller module.
  • Step 2 During the run, Handlers will be called multiple times for graphics card status acquisition and delay. GPU HostOps Dispatch instructions and data are sent to the Host GPU API. For each Handlers call, perform steps 2.1 through 2.6.
  • Step 2.1 In the Handlers specified by the resource scheduling algorithm, predict the GPU current consumption time according to the GPU consumption time history corresponding to the host physical graphics application interface.
  • Step 2.2 Using the Host GPU API and the graphics card driver interface, measure the Application GPU Load and the FPS in the current T time, and stop timing the CPU consumption time.
  • Step 2.3 Suspend the execution of the CPU for a period of time.
  • the length of the period is calculated by the scheduling algorithm according to the current consumption time of the CPU and the current consumption time of the GPU.
  • Step 2.4 Start timing
  • the GPU consumes time this time.
  • Step 2.5 call the original host physical graphics application interface.
  • Step 2.6 Stop Timing
  • the GPU consumes this time and updates to the GPU consumption time history corresponding to the host's physical graphics application interface.
  • Step 3 For each T cycle, if a virtual machine VMm does not satisfy the minimum FPS, find and reduce the minimum FPS setting of the virtual machine with the largest and smallest FPS. Reducing the number of FPS depends on the Application GPU Load of several recent frames, and the FPS and the Application GPU Load of several recent frames are linear.
  • Step 4 For each T cycle, if the physical graphics usage does not satisfy the minimum GPU Load, the minimum FPS setting of all virtual machines is increased. Increasing the number of FPS depends on the application GPU Load of several recent frames, and the FPS and the application GPU Load of several recent frames are linear. Step 5, Step 2 to Step 4 remain valid until the user specifies that the algorithm ends or replaces the algorithm or uninstalls the Scheduler Module.
  • the dispatch executor module uninstallation implementation is as follows:
  • Step a after each dispatch executor module receives the uninstall command, it starts the uninstall process from step b to step c.
  • Step b Restoring the host physical card application interface address set loaded by the process, and modifying the code at the address of each host physical graphics application interface to be the content of the original application interface address. The process will run the original application interface logic each time the host uses the physical graphics application interface.
  • Step c Binding the scheduling executor module to the end of the thread inserted in the corresponding virtual machine process, thereby uninstalling the scheduling executor module.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Debugging And Monitoring (AREA)
  • Processing Or Creating Images (AREA)
  • Stored Programmes (AREA)

Abstract

本发明提供一种基于应用程序即时效果反馈的在虚拟机间的物理显卡资源调度系统和方法。本发明描述的系统通过函数钩子的方法在宿主机物理显卡指令发送器和宿主机物理显卡应用程序接口之间插入调度执行器，延迟宿主机物理显卡指令发送器中指令和数据的发送，同时监控虚拟机内显卡应用程序相关的显示性能情况和物理显卡资源使用情况，进而提供反馈给任何基于时间或时序的显卡资源调度算法。本发明无需对虚拟机应用程序、宿主机操作系统、虚拟机操作系统、显卡驱动、虚拟机管理器进行任何修改。本发明无需暂停机器的运行，在重大提升虚拟机间显卡资源调度能力时运行时整体性能损耗低于5%。

Description

基于应用效果即时反馈的显卡虚拟化下资源调度系统、 方法 技术领域
本发明涉及的是一种应用于计算机应用技术领域的系统及方法， 尤其是一种基于应用程序 即时效果反馈的在虚拟机间的物理显卡资源（Graphics Processing Unit, GPU)调度系统方法， 具体涉及基于应用效果即时反馈的显卡虚拟化下的资源调度系统。
背景技术
显卡虚拟化技术（GPU Virtualization )正在被广泛应用于执行显卡计算（GPU Computing) 的数据中心当中，这些计算包括但不限于：云游戏（ Cloud Gaming),视频渲染（ Video Rendering ) 和通用显卡计算 ( General Purpose GPU Computing)。
但是目前缺少一种有效的虚拟机间调度物理显卡资源的系统和方法， 从而能够使并行运行 于多个虚拟机内部的各个应用程序获得较好的显卡加速效果同时有高资源使用率。 现有的显卡 穿透方法（Video Graphics Array Passthrough, VGA Passthrough )将每块可用物理显卡赋予到 运行的每个虚拟机当中。 然而这种方法的缺点是： 第一， 一般商用主板仅支持两到三块显卡， 因此需要特制的主板来同时运行多个需要显卡支持的虚拟机； 第二， 对每个虚拟机内部， 在其 运行期间通常不能用尽其拥有的物理显卡资源， 而在这种技术中， 剩余的显卡资源不能赋予给 其它虚拟机， 因此造成物理显卡资源的浪费。
另外一种办法是利用显卡半虚拟化技术（GPU Paravirtualization )使得多个虚拟机可以共 享一块或多块物理显卡。在 2009年 SIGOPS Operating Systems Review的 Volume 43 Issue 3中 发表了一篇名为 GPU Virtualization on VM ware' s Hosted I/O Architecture的论文提出了这种方 法和系统。之后，在 2008年 Multimedia Computing and Networking ± Bautin Μ·等人在 Graphic Engine Resource Management提出了物理显卡资源在多个应用间均分的调度策略。随后在 2011 年 Usenix ATC 上， Kato等人在 Timegraph: GPU scheduling for real-time multi-tasking environments论文中提出了通过引入显卡资源使用优先级并修改操作系统显卡驱动的方法来提 升物理显卡加速关键用户程序的能力。
上述两个方法能够最大化利用可用的物理显卡资源， 同时为多个虚拟机提供显卡加速的能 力。 然而， 这些方法的缺点是： 一方面， 需要修改操作系统或者显卡驱动， 当应用于虚拟机时 甚至需要修改虚拟机监控器 (Hyper visor )或虚拟机内的应用程序 ( Guest Application ), 因此 存在开发难度大的问题； 另一方面， 由于已有方法不能获得被加速的 Guest Application运行效 果反馈数据， 因此已有物理 GPU资源的调度系统和方法存在盲目性， 获得的资源调度效果一 - 发明内容
本发明针对现有技术存在的上述不足， 提供一种基于应用程序即时效果反馈的在虚拟机间 的物理显卡资源调度系统和方法。 传统 GPU虚拟化技术将虚拟机内的显卡命令和数据通过宿 主机物理显卡指令发送器（GPU HostOps Dispatch )发送到宿主机物理显卡应用程序接口（Host GPU API ) 在此基础上， 本发明提供的方法通过函数钩子的方法在 GPU HostOps Dispatch和 Host GPU API之间插入调度执行器 ( Agent ), 延迟 GPU HostOps Dispatch中指令和数据的发 送， 同时监控 Guest Application相关的显示性能情况和物理显卡资源使用情况， 进而提供反馈 给任何基于时间或时序的显卡资源调度算法。 基于时间或时序的显卡资源调度算法是指显卡资 源的开始使用、 结束使用和使用持续均部分或完全基于绝对或相对时间。 此外， 本发明描述的 系统通过调度控制器（Scheduling Controller )即时接受用户启用或停止各 Agent的决定， 并对 所采用调度方法的选择和参数改变，并据此即时改变各 Agent的相应参数设置。同时 Scheduling Controller 显示或记录当前物理显卡资源调度和使用情况， 各虚拟机内应用程序显卡资源使用 情况等一项或多项内容。
此外，本发明采用高级预测技术（Advanced Prediction ),配合延迟 GPU HostOps Dispatch 中指令和数据的发送， 以达到精确控制帧间延迟（Frame Latency)的目的。该高级预测技术包 含渲染 /显示开销预测（Frame Rendering Performance Prediction )和渲染 /显示命令队列逐帧推 进（Flush Single Queued Frame：)。其中渲染 /显示命令队列逐帧推进技术又包含标记操作（Mark Flush Frame)和提交操作（Commit Flush Frame：)。 标记操作为可选操作， 用于对本虚拟机在 渲染 /显示命令队列的某帧（包含但不限于上一帧或前数若干帧）进行标记， 以说明可被移出显 卡缓冲区的帧（包含但不限于该帧强制显示)。提交操作迫使某帧（之前执行标记操作后为对应 被标记帧）移出物理显卡缓冲区， 使得物理显卡缓冲区拥有空余空间。
本发明描述的系统和方法无需修改宿主机操作系统、 宿主机显卡驱动、 HyperViSOT、 虚拟 机操作系统、 虚拟机显卡驱动或虚拟机内的应用程序。 此外， 本发明描述的系统和方法在运行 时带来的性能开销低于 5%，启用或停止不致招致显著的虚拟机暂停时间 (仅需毫秒级暂停时间)。
本发明是通过以下技术方案实现的- 根据本发明的一个方面，提供一种基于应用效果即时反馈的显卡虚拟化下的资源调度系统， 包括宿主机物理显卡指令发送器、 宿主机物理显卡应用程序接口， 还包括如下模块： 调度执行器， 其连接在宿主机物理显卡指令发送器与宿主机物理显卡应用程序接口之间； 调度控制器， 其连接所述调度执行器，
其中， 调度控制器接收用户命令， 并将用户命令传递给调度执行器； 调度执行器接收来自 于调度控制器的用户命令， 监控应用程序运行状态， 并将应用程序显卡状态结果传输到调度控 制器， 同时根据调度控制器指定使用的调度算法， 周期性 /事件性计算要满足最低应用程序显卡 状态所需的延迟时间， 延迟宿主机物理显卡指令发送器中指令和数据发送到宿主机物理显卡应 用程序接口； 调度控制器接收并处理显示来自于调度执行器的调度结果和调度状态。
优选地， 调度控制器接收用户命令， 解析其中对各调度执行器的操作、 调度算法的配置和 相应参数， 并将用户命令传递给调度执行器模块、 接收来自于调度执行器模块的状态结果并显 示给用户。
优选地， 所述的调度控制器包括如下模块：
控制台， 其用于接收用户命令， 这些用户命令输入关于调度算法的配置和相应参数， 并从 调度通信器中获取调度结果并显示给用户；
调度通信器， 其用于负责调度控制器与一个或多个调度执行器的通信， 负责安装 /卸载调度 执行器、 将用户命令传递给调度执行器、 接收来自于调度执行器的应用程序显卡状态结果。
优选地， 所述的调度执行器包括如下模块- 调度器， 其用于接收用户命令中关于调度算法及其参数配置的指定， 并负责査找相应调度 算法的位置， 配置调度算法并运行相应的调度算法， 按需延迟宿主机物理显卡指令发送器中指 令和数据发送到宿主机物理显卡应用程序接口；
应用程序显卡状态监控器， 其用于采集来自宿主机物理显卡应用程序接口的显卡状态， 并 由此生成应用程序显卡状态结果， 同时将应用程序显卡状态结果反馈给调度器， 并传递给调度 控制器中的调度通信器。
优选地， 所述的应用程序显卡状态包括： 与应用程序种类相关的显卡物理状态和 /或逻辑状 态测度。 物理状态测度如显卡的负载率（GPU Load)、 温度、 电压等。 逻辑状态测度如对于计 算机三维游戏则相应显卡状态测度为每秒帧数 ( Frames per second, FPS)，对于计算机通用 GPU 运算则相应显卡状态测度为计算数每秒（Operations per second, Ops)、本应用程序的显卡负载 率（Application GPU Usage)等等。
根据本发明的一个方面， 提供一种根据本发明所述的资源调度系统的显卡虚拟化下的显卡 资源调度方法， 通过函数钩子的方法在宿主机物理显卡指令发送器和宿主机物理显卡应用程序 接口之间插入调度执行器 ( Agent ), 延迟宿主机物理显卡指令发送器中指令和数据发送到宿主 机物理显卡应用程序接口， 同时监控应用程序相关的显示性能情况和物理显卡资源使用情况， 进而提供反馈给任何基于时间或时序的显卡资源调度算法， 无需对虚拟机应用程序、 宿主机操 作系统、 虚拟机操作系统、 显卡驱动、 虚拟机管理器进行任何修改且性能损耗低。
优选地， 所述方法具体过程为： 一个或多个虚拟机启动完成后， 在客户需要安装所述资源 调度系统时， 通过应用程序运行的方法由调度控制器査找或按用户指定进程将调度执行器绑定 到相应虚拟机； 然后调度控制器中的调度通信器与各个被绑定的调度执行器建立通信； 在对显 卡资源进行调度的时候， 客户发出指令选择调度算法（可以是第三方开发的调度算法） 并提供 相应参数， 控制台在收到客户指令后由调度通信器将用户命令分发给各个调度执行器； 调度执 行器根据用户命令配置运行所选显卡资源调度算法， 延迟宿主机物理显卡指令发送器中指令和 数据发送到宿主机物理显卡应用程序接口； 与此同时， 应用程序显卡状态监控器采集来自宿主 机物理显卡应用程序接口的显卡状态， 并由此生成应用程序显卡状态， 随后周期性 /事件性将应 用程序显卡状态结果反馈给调度器， 并传递给调度控制器中的调度通信器； 在客户需要卸载所 述资源调度系统时， 客户通过调度控制器发出卸载指令， 控制台在收到客户指令后由调度通信 器将用户命令分发给各个调度执行器， 调度执行器收到卸载指令后停止自身运行。
优选地， 在显卡资源调度方法中使用显卡资源使用高级预测方法， 通过配合延迟宿主机物 理显卡指令发送器中指令和数据的发送， 以达到精确控制帧间延迟的目的， 显卡资源使用高级 预测方法包含渲染 /显示开销预测和渲染 /显示命令队列逐帧推进， 其中：
所述渲染 /显示开销预测是根据与本宿主机物理显卡应用程序接口对应的物理显卡资源消 耗时间历史记录预测物理显卡资源本次消耗时间；
所述渲染 /显示命令队列逐帧推进包含标记操作和提交操作， 其中， 标记操作为可选操作， 用于对本虚拟机在渲染 /显示命令队列的某帧（包含但不限于上一帧或前数若干帧）进行标记， 以说明可被移出显卡缓冲区的帧（包含但不限于该帧强制显示），提交操作迫使某帧（之前执行 标记操作后为对应被标记帧）移出物理显卡缓冲区， 使得物理显卡缓冲区拥有空余空间。 优选 地， 所述的将调度执行器绑定到相应虚拟机的步骤， 具体为- 步骤 1.1， 根据用户指定信息査找指定虚拟机图像渲染进程（根据虚拟机管理器设计不同， 这些进程也可以是虚拟机进程)，或选取全部相关虚拟机图像渲染进程，对这些中的每一个虚拟 机进程执行步骤 1.2到步骤 1.6;
步骤 1.2， 在该进程中创建新的线程（Thread ), 并在其中加载调度执行器；
步骤 1.3， 访问调度执行器入口， 初始化调度执行器；
步骤 1.4，査找该进程加载的宿主机物理显卡应用程序接口地址集合，修改其中每一个宿主 机物理显卡应用程序接口地址处的代码， 使其指向调度执行器中相应处理函数的入口并保存各 寄存器内容， 使得进程以后在每次使用宿主机物理显卡应用程序接口时会运行处理函数 ( Handlers);
步骤 1.5， 设定处理函数返回地址为旧的宿主机物理显卡应用程序接口地址之后运行指令， 恢复各寄存器内容， 使得处理函数运行结束后能正确执行原有宿主机物理显卡应用程序接口； 步骤 1.6， 该线程不得结束。
优选地， 所述的延迟宿主机物理显卡指令发送器中指令和数据发送到宿主机物理显卡应用 程序接口的步骤， 具体为-
- 若使用预测方法（技术)， 则包含如下子步骤：
步骤 2.1a， 在资源调度算法指定的处理函数中， 根据与本宿主机物理显卡应用程序接口对 应的物理显卡资源消耗时间历史记录预测物理显卡资源本次消耗时间， 停止计时本次中央处理 器 (Central Processing Unit, CPU) 消耗时间，；
步骤 2.2a， 暂停中央处理器执行一段时间， 这段时间的长度为调度算法根据中央处理器本 次消耗时间和所预测的物理显卡资源本次消耗时间计算而成；
步骤 2.3a， 开始计时物理显卡资源本次消耗时间；
步骤 2.4a， 调用原有宿主机物理显卡应用程序接口；
步骤 2.5a， 停止计时物理显卡资源本次消耗时间， 更新本次物理显卡资源消耗时间到与本 宿主机物理显卡应用程序接口对应的物理显卡资源消耗时间历史记录；
- 若不使用预测方法（技术)， 则包含如下子步骤：
步骤 2.1b， 在资源调度算法指定的处理函数中， 停止计时本次中央处理器消耗时间， 开始 计时物理显卡资源本次消耗时间；
步骤 2.2b， 调用原有宿主机物理显卡应用程序接口；
步骤 2.3b， 停止计时物理显卡资源本次消耗时间；
步骤 2.4b， 暂停中央处理器执行一段时间， 这段时间的长度为调度算法根据中央处理器本 次消耗时间和物理显卡资源本次消耗时间计算而成；
- 若使用显卡资源使用高级预测方法（技术)， 则包含如下子步骤- 步骤 2.1c， 在资源调度算法指定的处理函数中， 执行渲染 /显示命令队列逐帧推进操作的提 交操作， 该操作迫使某帧（如之前存在标记操作则为被标记帧）移出， 使得物理显卡缓冲区拥 有空余空间； 停止计时本次中央处理器消耗时间；
步骤 2.2c， 渲染 /显示开销预测会根据与本宿主机物理显卡应用程序接口对应的物理显卡资 源消耗时间历史记录预测物理显卡资源本次消耗时间；
步骤 2.3c， 暂停中央处理器执行一段时间， 这段时间的长度为调度算法根据中央处理器本 次消耗时间和所预测的物理显卡资源本次消耗时间计算而成；
步骤 2.4c， 开始计时物理显卡资源本次消耗时间；
步骤 2.5c， 调用原有宿主机物理显卡应用程序接口；
步骤 2.6c， 停止计时物理显卡资源本次消耗时间；
步骤 2.7c， 开始计时下次 CPU消耗时间， 可选择执行渲染 /显示命令队列逐帧推进操作的 标记操作，该操作对本虚拟机在渲染 /显示命令队列的某帧（包含但不限于上一帧或前数若干帧） 进行标记， 以说明可被移出显卡缓冲区的帧（包含但不限于该帧强制显示）。更新本次物理显卡 资源消耗时间到与本宿主机物理显卡应用程序接口对应的物理显卡资源消耗时间历史记录。
优选地， 所述的应用程序显卡状态监控器采集来自宿主机物理显卡应用程序接口的显卡状 态的步骤， 具体为- 步骤 3.1，在资源调度算法指定的处理函数中， 调用宿主机物理显卡应用程序接口、操作系 统内核或显卡驱动提供的接口按照资源调度算法需求和用户命令采集显卡状态， 如显卡的负载 率（GPU Load)、 温度、 电压、 FPS、 Ops、 本应用程序的显卡负载率等；
步骤 3.2， 在资源调度算法指定的处理函数中， 调用原有宿主机物理显卡应用程序接口； 优选地， 所述的生成应用程序显卡状态的步骤， 具体为：
步骤 4.1， 用户指定状态汇报频率， 在调度执行器中， 获取该状态汇报频率；
步骤 4.2， 当状态汇报时间点到来时，调度执行器中的应用程序显卡状态监控器将累计的状 态结果传送到调度控制器中的调度通信器中；
步骤 4.3， 调度执行器清空自身的状态结果缓冲区；
优选地， 所述的调度执行器收到卸载指令后停止自身运行的步骤， 具体为- 步骤 5.1， 每一个调度执行器收到卸载指令后， 开始从步骤 5.2到步骤 5.3卸载过程； 步骤 5.2，恢复该进程加载的宿主机物理显卡应用程序接口地址集合，修改其中每一个宿主 机物理显卡应用程序接口地址处的代码为原应用程序接口地址处内容， 使得进程以后在每次使 用宿主机物理显卡应用程序接口时会运行原应用程序接口逻辑；
步骤 5.3， 将调度执行器绑定到相应虚拟机过程中插入的线程结束， 从而卸载调度执行器； 优选地， 所述的资源调度算法具体包括如下步骤- 步骤 6.1， 对于虚拟机群 VM1, VM2-到 VMn， 每个虚拟机的调度执行器中的调度器解析 用户方法配置， 得到需要满足的最小显卡的负载率、 最小每秒帧数 （本专利适用范围不限于计 算机游戏， 对于其它显卡应用， 可以针对不同的状态测量)、 用户指定的检测周期 T;
步骤 6.2，在运行期间，处理函数会被多次调用，对于每次处理函数调用，使用预测方法（技 术）执行步骤 2.1a到步骤 2.5a; 或不使用预测方法（技术）执行步骤 2.1b到步骤 2.4b ;
步骤 6.3， 对于每一个 T周期， 如果某个虚拟机 VMm不满足状态测量， 则寻找并降低拥 有最大最小每秒帧数的虚拟机的最小每秒帧数设定； 降低每秒帧数的数量取决于最近若干帧的 应用程序的显卡负载率（Application GPU Load ),每秒帧数和最近若干帧的应用显卡的负载率 为线性关系；
步骤 6.4， 对于每一个 T周期， 如果物理显卡使用率不满足最小显卡负载率， 则提高所有 虚拟机的最小每秒帧数设定；提高每秒帧数的数量取决于最近若干帧的应用程序的显卡负载率， 每秒帧数和最近若干帧的应用程序的显卡负载率为线性关系；
步骤 6.5， 步骤 6.2到步骤 6.4保持有效， 直到用户指定方法结束或更换方法或卸载调度执 行器。
在本发明的一个优选的具体实施方式中，每台虚拟机对应的 GPU HostOps Dispatch在本发 明中会被安装一个独自拥有的调度执行器。 全局唯一的调度控制器与一个或多个调度执行器相 连。 本发明相对于现有技术具有以下优点： 首先， 无需对虚拟机应用程序、 宿主机操作系统、 虚拟机操作系统、 显卡驱动、 虚拟机管理器进行任何修改。 已有系统通常需要大量改动以上部 分中的一个来实现相近的调度能力， 这种修改会导致已有系统必须不断进化来兼容最新的应用 程序、 操作系统或显卡驱动等。 其次， 本发明在安装或卸载时无需暂停机器的运行， 这种特性 使得本系统易于部署在商业系统中， 尤其适用于需要达到 7X 24小时可用性的商业服务器上。 最后，本发明在重大提升虚拟机间显卡资源调度能力时运行时性能极高，整体性能损耗低于 5%。
附图说明
图 1为本发明模块示意图。
图 2为本发明架构示意图。
具体实施方式
下面对本发明的实施例作详细说明， 本实施例在以本发明技术方案为前提下进行实施， 给 出了详细的实施方式和具体的操作过程， 但本发明的保护范围不限于下述的实施例。
实施例
如图 2所示， 根据本发明提供的基于应用效果即时反馈的显卡虚拟化下的资源调度系统， 包括： 调度执行器模块和调度控制器模块， 其中： 调度控制器模块与调度执行器模块相连接， 传送用户命令到调度执行器当中并接收其返回的显卡状态结果。 调度执行器被插入到宿主机物 理显卡指令发送器 GPU HostOps Dispatch和宿主机物理显卡应用程序接口 Host GPU API之间， 延迟相应调用和数据的向下传输。 调度执行器模块同时负责利用宿主机物理显卡应用程序接口 Host GPU API采集显卡物理状态和 /或逻辑状态测度。 本实施例针对运行于虚拟机中的计算机 游戏， 因此采集的物理状态、 逻辑状态包括： Application GPU Load和 FPS。
如图 1所示， 所述的调度控制器模块包括： 控制台子模块、 调度通信器子模块， 其中： 控 制台子模块用于接收用户命令， 这些命令输入关于调度算法的配置和相应参数。 此外， 控制台 子模块周期性 /事件性从调度通信器子模块中获取调度结果并显示给用户。调度通信器子模块负 责调度控制器模块与一个或多个调度执行器模块的通信， 负责安装 /卸载调度执行器、将用户命 令传递给调度执行器子模块等操作。 所述的事件性指目标事件一次或多次发生， 但发生的时间 间隔并不恒定， 事件在时间上的分布在数学上可表达为非周期性质的时间序列。
如图 1所示， 所述的调度执行器模块包括： 调度器子模块、 应用程序显卡状态监控器子模 块， 其中： 调度器子模块接收用户命令中关于调度算法及其参数配置的指定， 并负责按照配置 运行相应的调度算法， 按需延迟 GPU HostOps Dispatch中指令和数据发送到 Host GPU API。 应用程序显卡状态监控器子模块负责采集来自 Host GPU API的显卡状态， 并由此生成应用程 序显卡状态， 随后周期性 /事件性将应用程序显卡状态结果反馈给调度器子模块， 并传递给调度 控制器模块中的调度通信器子模块。
所述的应用程序显卡状态指与应用程序种类相关的显卡物理状态和 /或逻辑状态测度。在本 实施例中， 采集的物理状态、 逻辑状态包括： Application GPU Load和 FPS。
本实施例针对 VMWare Player 4.0虚拟机管理器系统， 因此指定虚拟机图像渲染进程即虚 拟机进程。 在本实施例中， 仅考虑用户选取全部相关虚拟机图像渲染进程情况。
本实施例应用的基于应用效果即时反馈的显卡虚拟化下的资源调度方法被配置为： 最小 GPU Load=80% 最小 FPS=30、 用户指定的检测周期 T=l秒。
本实施例通过以下方式进行工作：
步骤一， 用户选取全部相关虚拟机进程， 对这些中的每一个虚拟机进程执行步骤二到步骤 六：
步骤二， 在该进程中创建新的线程（Thread), 并在其中加载调度执行器模块。
步骤三， 访问调度执行器模块入口， 初始化调度执行器模块。
步骤四， 査找该进程加载的宿主机物理显卡应用程序接口地址集合， 修改其中每一个宿主 机物理显卡应用程序接口地址处的代码， 使其指向调度执行器模块中相应处理函数的入口并保 存各寄存器内容。 使得进程以后在每次使用宿主机物理显卡应用程序接口时会运行处理函数 (Handlers )。
步骤五， 设定处理函数返回地址为旧的宿主机物理显卡应用程序接口地址之后运行指令， 恢复各寄存器内容， 使得处理函数运行结束后能正确执行原有宿主机物理显卡应用程序接口。
步骤六， 该线程不得结束。
通过以上步骤将调度执行器模块绑定到相应虚拟机， 再经过调度控制器模块中的调度通信 器子模块与各个被绑定的调度执行器模块建立通信后， 调度执行器模块可传送状态结果到调度 控制器模块， 并对调度控制器模块发出的用户命令进行相应。 当之后某一时刻需要对显卡资源 进行调度时， 具体过程如下- 步骤 1， 对于虚拟机群 VM1, VM2…到 VMn， 每个虚拟机的调度执行器模块中的调度器模 块解析用户算法配置， 得到需要满足的最小 GPU Load=80%、 最小 FPS=30、 用户指定的检测 周期 T=l秒。
步骤 2， 在运行期间， Handlers会被多次调用， 用于显卡状态采集和延迟 GPU HostOps Dispatch中指令和数据发送到 Host GPU API, 对于每次 Handlers调用， 执行步骤 2.1到步骤 2.6。
步骤 2.1， 在资源调度算法指定的 Handlers中， 根据与本宿主机物理显卡应用程序接口对 应的 GPU消耗时间历史记录预测 GPU本次消耗时间。
步骤 2.2利用 Host GPU API和显卡驱动接口， 测量本次 Application GPU Load和当前 T 时间内的 FPS， 停止计时本次 CPU消耗时间。
步骤 2.3，暂停 CPU执行一段时间，这段时间的长度为调度算法根据 CPU本次消耗时间和 GPU本次消耗时间计算而成。
步骤 2.4， 开始计时 GPU本次消耗时间。
步骤 2.5， 调用原有宿主机物理显卡应用程序接口。
步骤 2.6， 停止计时 GPU本次消耗时间， 更新到与本宿主机物理显卡应用程序接口对应的 GPU消耗时间历史记录。
步骤 3， 对于每一个 T周期， 如果某个虚拟机 VMm不满足最小 FPS， 则寻找并降低拥有 最大最小 FPS的虚拟机的最小 FPS设定。降低 FPS的数量取决于最近若干帧的 Application GPU Load, FPS和最近若干帧的 Application GPU Load为线性关系。
步骤 4， 对于每一个 T周期， 如果物理显卡使用率不满足最小 GPU Load, 则提高所有虚 拟机的最小 FPS设定。 提高 FPS的数量取决于最近若干帧的 Application GPU Load, FPS和 最近若干帧的 Application GPU Load为线性关系。 步骤 5， 步骤 2到步骤 4保持有效， 直到用户指定算法结束或更换算法或卸载调度执行器 模块。
调度执行器模块卸载实施例如下：
步骤 a， 每一个调度执行器模块收到卸载指令后， 开始从步骤 b到步骤 c卸载过程。 步骤 b， 恢复该进程加载的宿主机物理显卡应用程序接口地址集合， 修改其中每一个宿主 机物理显卡应用程序接口地址处的代码为原应用程序接口地址处内容。 使得进程以后在每次使 用宿主机物理显卡应用程序接口时会运行原应用程序接口逻辑。
步骤 c， 将调度执行器模块绑定到相应虚拟机过程中插入的线程结束， 从而卸载调度执行 器模块。

Claims

权 利 要 求 书
1、一种基于应用效果即时反馈的显卡虚拟化下的资源调度系统，包括宿主机物理显卡指令 发送器、 宿主机物理显卡应用程序接口， 其特征在于， 还包括如下模块- 调度执行器， 其连接在宿主机物理显卡指令发送器与宿主机物理显卡应用程序接口之间； 调度控制器， 其连接所述调度执行器，
其中， 调度控制器接收用户命令， 并将用户命令传递给调度执行器； 调度执行器接收来自 于调度控制器的用户命令， 监控应用程序运行状态， 并将应用程序显卡状态结果传输到调度控 制器， 同时根据调度控制器指定使用的调度算法， 周期性 /事件性计算要满足最低应用程序显卡 状态所需的延迟时间， 延迟宿主机物理显卡指令发送器中指令和数据发送到宿主机物理显卡应 用程序接口； 调度控制器接收并处理显示来自于调度执行器的调度结果和调度状态。
2、根据权利要求 1所述的基于应用效果即时反馈的显卡虚拟化下的资源调度系统，其特征 在于， 调度控制器接收用户命令， 解析其中对各调度执行器的操作、 调度算法的配置和相应参 数， 并将用户命令传递给调度执行器、 接收来自于调度执行器的状态结果并显示给用户。
3、根据权利要求 1所述的基于应用效果即时反馈的显卡虚拟化下的资源调度系统，其特征 是， 所述的调度控制器包括如下模块- 控制台， 其用于接收用户命令， 这些用户命令输入关于调度算法的配置和相应参数， 并从 调度通信器中获取调度结果并显示给用户；
调度通信器， 其用于负责调度控制器与一个或多个调度执行器的通信， 负责安装 /卸载调度 执行器、 将用户命令传递给调度执行器、 接收来自于调度执行器的应用程序显卡状态结果。
4、根据权利要求 1所述的基于应用效果即时反馈的显卡虚拟化下的资源调度系统，其特征 是， 所述的调度执行器包括如下模块- 调度器， 其用于接收用户命令中关于调度算法及其参数配置的指定， 并负责査找相应调度 算法的位置， 配置调度算法并运行相应的调度算法， 按需延迟宿主机物理显卡指令发送器中指 令和数据发送到宿主机物理显卡应用程序接口；
应用程序显卡状态监控器， 其用于采集来自宿主机物理显卡应用程序接口的显卡状态， 并 由此生成应用程序显卡状态结果， 同时将应用程序显卡状态结果反馈给调度器， 并传递给调度 控制器中的调度通信器。
5、根据权利要求 1所述的基于应用效果即时反馈的显卡虚拟化下的资源调度系统，其特征 是， 所述的应用程序显卡状态包括： 与应用程序种类相关的显卡物理状态和 /或逻辑状态测度。
6、 一种根据上述任一权利要求所述的资源调度系统的显卡虚拟化下的显卡资源调度方法， 其特征在于， 通过函数钩子的方法在宿主机物理显卡指令发送器和宿主机物理显卡应用程序接 口之间插入调度执行器， 延迟宿主机物理显卡指令发送器中指令和数据发送到宿主机物理显卡 应用程序接口， 同时监控应用程序相关的显示性能情况和物理显卡资源使用情况， 进而提供反 馈给任何基于时间或时序的显卡资源调度算法， 无需对虚拟机应用程序、 宿主机操作系统、 虚 拟机操作系统、 显卡驱动、 虚拟机管理器进行任何修改且性能损耗低。
7、根据权利要求 6所述的显卡虚拟化下的显卡资源调度方法， 其特征是，所述方法具体过 程为： 一个或多个虚拟机启动完成后， 在客户需要安装所述资源调度系统时， 通过应用程序运 行的方法由调度控制器査找或按用户指定进程将调度执行器绑定到相应虚拟机； 然后调度控制 器中的调度通信器与各个被绑定的调度执行器建立通信； 在对显卡资源进行调度的时候， 客户 发出指令选择调度算法并提供相应参数， 控制台在收到客户指令后由调度通信器将用户命令分 发给各个调度执行器； 调度执行器根据用户命令配置运行所选显卡资源调度算法， 延迟宿主机 物理显卡指令发送器中指令和数据发送到宿主机物理显卡应用程序接口； 与此同时， 应用程序 显卡状态监控器采集来自宿主机物理显卡应用程序接口的显卡状态， 并由此生成应用程序显卡 状态， 随后周期性 /事件性将应用程序显卡状态结果反馈给调度器， 并传递给调度控制器中的调 度通信器； 在客户需要卸载所述资源调度系统时， 客户通过调度控制器发出卸载指令， 控制台 在收到客户指令后由调度通信器将用户命令分发给各个调度执行器， 调度执行器收到卸载指令 后停止自身运行。
8、根据权利要求 6所述的显卡虚拟化下的显卡资源调度方法显卡虚拟化下的显卡资源调度 方法， 其特征在于， 在显卡资源调度方法中使用显卡资源使用高级预测方法， 通过配合延迟宿 主机物理显卡指令发送器中指令和数据的发送， 以达到精确控制帧间延迟的目的， 显卡资源使 用高级预测方法包含渲染 /显示开销预测和渲染 /显示命令队列逐帧推进， 其中- 所述渲染 /显示开销预测是根据与本宿主机物理显卡应用程序接口对应的物理显卡资源消 耗时间历史记录预测物理显卡资源本次消耗时间；
所述渲染 /显示命令队列逐帧推进包含标记操作和提交操作， 其中， 标记操作为可选操作， 用于对本虚拟机在渲染 /显示命令队列的某帧进行标记， 以说明可被移出显卡缓冲区的帧， 提交 操作迫使某帧移出物理显卡缓冲区， 使得物理显卡缓冲区拥有空余空间。
9、根据权利要求 7或 8所述的显卡虚拟化下的显卡资源调度方法，其特征是，所述的将调 度执行器绑定到相应虚拟机的步骤， 具体为：
步骤 1.1，根据用户指定信息査找指定虚拟机图像渲染进程，或选取全部相关虚拟机图像渲 染进程， 对这些中的每一个虚拟机进程执行步骤 1.2到步骤 1.6;
步骤 1.2， 在该进程中创建新的线程， 并在其中加载调度执行器；
步骤 1.3， 访问调度执行器入口， 初始化调度执行器；
步骤 1.4，査找该进程加载的宿主机物理显卡应用程序接口地址集合，修改其中每一个宿主 机物理显卡应用程序接口地址处的代码， 使其指向调度执行器中相应处理函数的入口并保存各 寄存器内容， 使得进程以后在每次使用宿主机物理显卡应用程序接口时会运行处理函数； 步骤 1.5， 设定处理函数返回地址为旧的宿主机物理显卡应用程序接口地址之后运行指令， 恢复各寄存器内容， 使得处理函数运行结束后能正确执行原有宿主机物理显卡应用程序接口； 步骤 1.6， 该线程不得结束。
10、 根据权利要求 8所述的显卡虚拟化下的显卡资源调度方法， 其特征是， 所述的延迟宿 主机物理显卡指令发送器中指令和数据发送到宿主机物理显卡应用程序接口的步骤， 具体为： - 若使用预测方法， 则包含如下子步骤- 步骤 2.1a， 在资源调度算法指定的处理函数中， 根据与本宿主机物理显卡应用程序接口对 应的物理显卡资源消耗时间历史记录预测物理显卡资源本次消耗时间， 停止计时本次中央处理 器消耗时间；
步骤 2.2a， 暂停中央处理器执行一段时间， 这段时间的长度为资源调度算法根据中央处理 器本次消耗时间和所预测的物理显卡资源本次消耗时间计算而成；
步骤 2.3a， 开始计时物理显卡资源本次消耗时间；
步骤 2.4a， 调用原有宿主机物理显卡应用程序接口；
步骤 2.5a， 停止计时物理显卡资源本次消耗时间， 更新本次物理显卡资源消耗时间到与本 宿主机物理显卡应用程序接口对应的物理显卡资源消耗时间历史记录；
- 若不使用预测方法， 则包含如下子步骤- 步骤 2.1b， 在资源调度算法指定的处理函数中， 停止计时本次中央处理器消耗时间， 开始 计时物理显卡资源本次消耗时间；
步骤 2.2b， 调用原有宿主机物理显卡应用程序接口；
步骤 2.3b， 停止计时物理显卡资源本次消耗时间；
步骤 2.4b， 暂停中央处理器执行一段时间， 这段时间的长度为调度算法根据中央处理器本 次消耗时间和物理显卡资源本次消耗时间计算而成；
- 若使用显卡资源使用高级预测方法， 则包含如下子步骤：
步骤 2.1c， 在资源调度算法指定的处理函数中， 执行渲染 /显示命令队列逐帧推进操作的提 交操作， 该操作迫使某帧移出物理显卡缓冲区， 使得物理显卡缓冲区拥有空余空间， 其中， 若 之前存在标记操作则该某帧为被标记帧； 停止计时本次中央处理器消耗时间；
步骤 2.2c， 渲染 /显示开销预测会根据与本宿主机物理显卡应用程序接口对应的物理显卡资 源消耗时间历史记录预测物理显卡资源本次消耗时间；
步骤 2.3c， 暂停中央处理器执行一段时间， 这段时间的长度为调度算法根据中央处理器本 次消耗时间和所预测的物理显卡资源本次消耗时间计算而成；
步骤 2.4c， 开始计时物理显卡资源本次消耗时间；
步骤 2.5c， 调用原有宿主机物理显卡应用程序接口；
步骤 2.6c， 停止计时物理显卡资源本次消耗时间；
步骤 2.7c， 开始计时下次 CPU消耗时间， 可选择执行渲染 /显示命令队列逐帧推进操作的 标记操作， 该操作对本虚拟机在渲染 /显示命令队列的某帧进行标记， 以说明可被移出显卡缓冲 区的帧； 更新本次物理显卡资源消耗时间到与本宿主机物理显卡应用程序接口对应的物理显卡 资源消耗时间历史记录。
11、 根据权利要求 7所述的显卡虚拟化下的显卡资源调度方法， 其特征是， 所述的应用程 序显卡状态监控器采集来自宿主机物理显卡应用程序接口的显卡状态的步骤， 具体为：
步骤 3.1，在资源调度算法指定的处理函数中， 调用宿主机物理显卡应用程序接口、操作系 统内核或显卡驱动提供的接口按照资源调度算法需求和用户命令采集显卡状态；
步骤 3.2， 在资源调度算法指定的处理函数中， 调用原有宿主机物理显卡应用程序接口；
12、 根据权利要求 7所述的显卡虚拟化下的显卡资源调度方法， 其特征是， 所述的生成应 用程序显卡状态的步骤， 具体为- 步骤 4.1， 用户指定状态汇报频率， 在调度执行器中， 获取该状态汇报频率；
步骤 4.2， 当状态汇报时间点到来时，调度执行器中的应用程序显卡状态监控器将累计的状 态结果传送到调度控制器中的调度通信器中；
步骤 4.3， 调度执行器清空自身的状态结果缓冲区。
13、 根据权利要求 7所述的显卡虚拟化下的显卡资源调度方法， 其特征是， 所述的调度执 行器收到卸载指令后停止自身运行的步骤， 具体为- 步骤 5.1， 每一个调度执行器收到卸载指令后， 开始从步骤 5.2到步骤 5.3卸载过程； 步骤 5.2，恢复该进程加载的宿主机物理显卡应用程序接口地址集合，修改其中每一个宿主 机物理显卡应用程序接口地址处的代码为原应用程序接口地址处内容， 使得进程以后在每次使 用宿主机物理显卡应用程序接口时会运行原应用程序接口逻辑； 步骤 5.3， 将调度执行器绑定到相应虚拟机过程中插入的线程结束， 从而卸载调度执行器。
14、 根据权利要求 7所述的显卡虚拟化下的显卡资源调度方法， 其特征是， 所述的资源调 度算法具体包括如下步骤：
步骤 6.1， 对于虚拟机群 VM1, VM2 到 VMn， 每个虚拟机的调度执行器中的调度器解析 用户方法配置， 得到需要满足的最小显卡的负载率、 最小每秒帧数、 用户指定的检测周期 T; 步骤 6.2， 在运行期间， 处理函数会被多次调用， 对于每次处理函数调用， 使用预测方法执 行步骤 2.1a到步骤 2.5a; 或不使用预测方法执行步骤 2.1b到步骤 2.4b;
步骤 6.3， 对于每一个 T周期， 如果某个虚拟机 VMm不满足状态测量， 则寻找并降低拥 有最大最小每秒帧数的虚拟机的最小每秒帧数设定； 降低每秒帧数的数量取决于最近若干帧的 应用程序的显卡负载率， 每秒帧数和最近若干帧的应用显卡的负载率为线性关系；
步骤 6.4， 对于每一个 T周期， 如果物理显卡使用率不满足最小显卡的负载率， 则提高所 有虚拟机的最小每秒帧数设定； 提高每秒帧数的数量取决于最近若干帧的应用程序的显卡负载 率， 每秒帧数和最近若干帧的应用程序的显卡负载率为线性关系；
步骤 6.5， 步骤 6.2到步骤 6.4保持有效， 直到用户指定方法结束或更换方法或卸载调度执 行器。
15、 根据权利要求 8所述的显卡虚拟化下的显卡资源调度方法显卡虚拟化下的显卡资源调 度方法， 其特征在于， 如之前执行标记操作， 所述某帧为对应被标记帧。
PCT/CN2013/077457 2012-07-26 2013-06-19 基于应用效果即时反馈的显卡虚拟化下资源调度系统、方法 WO2014015725A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US15/563,951 US10922140B2 (en) 2012-07-26 2013-06-19 Resource scheduling system and method under graphics processing unit virtualization based on instant feedback of application effect

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201210261862.0 2012-07-26
CN201210261862.0A CN102890643B (zh) 2012-07-26 2012-07-26 基于应用效果即时反馈的显卡虚拟化下的资源调度系统

Publications (1)

Publication Number Publication Date
WO2014015725A1 true WO2014015725A1 (zh) 2014-01-30

Family

ID=47534151

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2013/077457 WO2014015725A1 (zh) 2012-07-26 2013-06-19 基于应用效果即时反馈的显卡虚拟化下资源调度系统、方法

Country Status (3)

Country Link
US (1) US10922140B2 (zh)
CN (1) CN102890643B (zh)
WO (1) WO2014015725A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111524058A (zh) * 2019-02-01 2020-08-11 纬创资通股份有限公司 硬件加速方法及硬件加速系统

Families Citing this family (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102890643B (zh) * 2012-07-26 2015-01-21 上海交通大学 基于应用效果即时反馈的显卡虚拟化下的资源调度系统
CN103645872A (zh) * 2013-12-08 2014-03-19 侯丰花 一种显卡的优化计算方法
CN108345492A (zh) * 2014-04-08 2018-07-31 华为技术有限公司 一种虚拟化环境中的数据通信的方法、装置及处理器
US9690928B2 (en) * 2014-10-25 2017-06-27 Mcafee, Inc. Computing platform security methods and apparatus
CN105426310B (zh) * 2015-11-27 2018-06-26 北京奇虎科技有限公司 一种检测目标进程的性能的方法和装置
JP6823251B2 (ja) * 2016-10-13 2021-02-03 富士通株式会社 情報処理装置、情報処理方法及びプログラム
CN108073440B (zh) * 2016-11-18 2023-07-07 南京中兴新软件有限责任公司 一种虚拟化环境下的显卡管理方法、装置及系统
US11144357B2 (en) * 2018-05-25 2021-10-12 International Business Machines Corporation Selecting hardware accelerators based on score
US10977098B2 (en) 2018-08-14 2021-04-13 International Business Machines Corporation Automatically deploying hardware accelerators based on requests from users
US10892944B2 (en) 2018-11-29 2021-01-12 International Business Machines Corporation Selecting and using a cloud-based hardware accelerator
CN109656714B (zh) * 2018-12-04 2022-10-28 成都雨云科技有限公司 一种虚拟化显卡的gpu资源调度方法
CN109712060B (zh) * 2018-12-04 2022-12-23 成都雨云科技有限公司 一种基于gpu容器技术的云桌面显卡共享方法及系统
CN109934327B (zh) * 2019-02-18 2020-03-31 星汉智能科技股份有限公司 一种智能卡的计时方法及系统
CN110532071B (zh) * 2019-07-12 2023-06-09 上海大学 一种基于gpu的多应用调度系统和方法
CN111522692B (zh) * 2020-04-20 2023-05-30 浙江大学 一种基于虚拟机的多操作系统输入及输出设备冗余保障系统
CN112230931B (zh) * 2020-10-22 2021-11-02 上海壁仞智能科技有限公司 适用于图形处理器的二次卸载的编译方法、装置和介质
US20220180588A1 (en) * 2020-12-07 2022-06-09 Intel Corporation Efficient memory space sharing of resources for cloud rendering
CN113674132B (zh) * 2021-07-23 2024-05-14 中标软件有限公司 一种通过检测显卡能力切换窗口管理渲染后端的方法
CN116157186A (zh) * 2021-09-23 2023-05-23 谷歌有限责任公司 基于游戏交互状态的自动化帧调步
CN113975816B (zh) * 2021-12-24 2022-11-25 北京蔚领时代科技有限公司 一种基于hook的通过DirectX接口使用显卡的显卡分配方法
CN115686758B (zh) * 2023-01-03 2023-03-21 麒麟软件有限公司 一种基于帧统计的VirtIO-GPU性能可控方法
CN116777730B (zh) * 2023-08-25 2023-10-31 湖南马栏山视频先进技术研究院有限公司 一种基于资源调度的gpu效能提高方法

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1700171A (zh) * 2004-04-30 2005-11-23 微软公司 提供从虚拟环境对硬件的直接访问
CN101419558A (zh) * 2008-11-13 2009-04-29 湖南大学 Cuda图形子系统虚拟化方法
CN102890643A (zh) * 2012-07-26 2013-01-23 上海交通大学 基于应用效果即时反馈的显卡虚拟化下的资源调度系统

Family Cites Families (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7673304B2 (en) * 2003-02-18 2010-03-02 Microsoft Corporation Multithreaded kernel for graphics processing unit
US8274518B2 (en) * 2004-12-30 2012-09-25 Microsoft Corporation Systems and methods for virtualizing graphics subsystems
US8341624B1 (en) * 2006-09-28 2012-12-25 Teradici Corporation Scheduling a virtual machine resource based on quality prediction of encoded transmission of images generated by the virtual machine
US7650603B2 (en) * 2005-07-08 2010-01-19 Microsoft Corporation Resource management for virtualization of graphics adapters
US8310491B2 (en) * 2007-06-07 2012-11-13 Apple Inc. Asynchronous notifications for concurrent graphics operations
US8122229B2 (en) * 2007-09-12 2012-02-21 Convey Computer Dispatch mechanism for dispatching instructions from a host processor to a co-processor
US8284205B2 (en) * 2007-10-24 2012-10-09 Apple Inc. Methods and apparatuses for load balancing between multiple processing units
KR100962531B1 (ko) * 2007-12-11 2010-06-15 한국전자통신연구원 동적 로드 밸런싱을 지원하는 멀티 쓰레딩 프레임워크를 수행하는 장치 및 이를 이용한 프로세싱 방법
GB2462860B (en) * 2008-08-22 2012-05-16 Advanced Risc Mach Ltd Apparatus and method for communicating between a central processing unit and a graphics processing unit
US8368701B2 (en) * 2008-11-06 2013-02-05 Via Technologies, Inc. Metaprocessor for GPU control and synchronization in a multiprocessor environment
US8910153B2 (en) * 2009-07-13 2014-12-09 Hewlett-Packard Development Company, L. P. Managing virtualized accelerators using admission control, load balancing and scheduling
US20110102443A1 (en) * 2009-11-04 2011-05-05 Microsoft Corporation Virtualized GPU in a Virtual Machine Environment
CN101706742B (zh) * 2009-11-20 2012-11-21 北京航空航天大学 一种基于多核动态划分的非对称虚拟机i/o调度方法
US8669990B2 (en) * 2009-12-31 2014-03-11 Intel Corporation Sharing resources between a CPU and GPU
EP2383648B1 (en) * 2010-04-28 2020-02-19 Telefonaktiebolaget LM Ericsson (publ) Technique for GPU command scheduling
CN102262557B (zh) * 2010-05-25 2015-01-21 运软网络科技(上海)有限公司 通过总线架构构建虚拟机监控器的方法及性能服务框架
CN101968749B (zh) * 2010-09-26 2013-01-02 华中科技大学 虚拟机过度分配环境下的mpi消息接收方法
US8463980B2 (en) * 2010-09-30 2013-06-11 Microsoft Corporation Shared memory between child and parent partitions
US8970603B2 (en) * 2010-09-30 2015-03-03 Microsoft Technology Licensing, Llc Dynamic virtual device failure recovery
US9170843B2 (en) * 2011-09-24 2015-10-27 Elwha Llc Data handling apparatus adapted for scheduling operations according to resource allocation based on entitlement
US9135189B2 (en) * 2011-09-07 2015-09-15 Microsoft Technology Licensing, Llc Delivering GPU resources across machine boundaries
US8941670B2 (en) * 2012-01-17 2015-01-27 Microsoft Corporation Para-virtualized high-performance computing and GDI acceleration
EP2742425A1 (en) * 2012-05-29 2014-06-18 Qatar Foundation Graphics processing unit controller, host system, and methods

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1700171A (zh) * 2004-04-30 2005-11-23 微软公司 提供从虚拟环境对硬件的直接访问
CN101419558A (zh) * 2008-11-13 2009-04-29 湖南大学 Cuda图形子系统虚拟化方法
CN102890643A (zh) * 2012-07-26 2013-01-23 上海交通大学 基于应用效果即时反馈的显卡虚拟化下的资源调度系统

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111524058A (zh) * 2019-02-01 2020-08-11 纬创资通股份有限公司 硬件加速方法及硬件加速系统
CN111524058B (zh) * 2019-02-01 2023-08-22 纬创资通股份有限公司 硬件加速方法及硬件加速系统

Also Published As

Publication number Publication date
CN102890643A (zh) 2013-01-23
CN102890643B (zh) 2015-01-21
US20180246770A1 (en) 2018-08-30
US10922140B2 (en) 2021-02-16

Similar Documents

Publication Publication Date Title
WO2014015725A1 (zh) 基于应用效果即时反馈的显卡虚拟化下资源调度系统、方法
US11797327B2 (en) Dynamic virtual machine sizing
Qi et al. VGRIS: Virtualized GPU resource isolation and scheduling in cloud gaming
EP3039540B1 (en) Virtual machine monitor configured to support latency sensitive virtual machines
US9405585B2 (en) Management of heterogeneous workloads
EP3161628B1 (en) Intelligent gpu scheduling in a virtualization environment
Zhang et al. vGASA: Adaptive scheduling algorithm of virtualized GPU resource in cloud gaming
US10970129B2 (en) Intelligent GPU scheduling in a virtualization environment
US20090077564A1 (en) Fast context switching using virtual cpus
CN105550040B (zh) 基于kvm平台的虚拟机cpu资源预留算法
US9189293B2 (en) Computer, virtualization mechanism, and scheduling method
Bai et al. Task-aware based co-scheduling for virtual machine system
US10846088B2 (en) Control of instruction execution in a data processor
US20090241112A1 (en) Recording medium recording virtual machine control program and virtual machine system
US20200334075A1 (en) Process scheduling in a processing system having at least one processor and shared hardware resources
US20130125131A1 (en) Multi-core processor system, thread control method, and computer product
Zhao et al. Efficient sharing and fine-grained scheduling of virtualized GPU resources
Yu et al. Colab: a collaborative multi-factor scheduler for asymmetric multicore processors
US20220382587A1 (en) Data processing systems
Lin et al. Improving GPOS real-time responsiveness using vCPU migration in an embedded multicore virtualization platform
Elmougy et al. Diagnosing the interference on cpu-gpu synchronization caused by cpu sharing in multi-tenant gpu clouds
Lee et al. Interrupt handler migration and direct interrupt scheduling for rapid scheduling of interrupt-driven tasks
CN110333899B (zh) 数据处理方法、装置和存储介质
US20230229473A1 (en) Adaptive idling of virtual central processing unit
Xia et al. PaS: A preemption-aware scheduling interface for improving interactive performance in consolidated virtual machine environment

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: 13822194

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205N DATED 07/04/2015)

122 Ep: pct application non-entry in european phase

Ref document number: 13822194

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 15563951

Country of ref document: US