基于虛擬機(jī)的GPU計(jì)算資源管理系統(tǒng)
發(fā)布時(shí)間：2019-03-16 16:30
【摘要】：統(tǒng)一計(jì)算架構(gòu)讓圖形處理器（GPU）升級(jí)為通用圖形處理器，，分擔(dān)原來(lái)需由中央處理器處理的通用計(jì)算任務(wù)，并獲取相對(duì)中央處理器數(shù)倍甚至數(shù)十倍的加速比。與此同時(shí)，隨著虛擬化技術(shù)日趨成熟，虛擬化技術(shù)和GPU計(jì)算的結(jié)合越來(lái)越受到高性能計(jì)算的青睞。但是目前沒(méi)有基于虛擬機(jī)的GPU計(jì)算管理系統(tǒng)，或者在虛擬化環(huán)境下不能有效使用GPU的強(qiáng)勁計(jì)算能力，或者基于物理機(jī)的集群中資源不便有效管理、容易受到惡意用戶(hù)攻擊且崩潰后不能迅速恢復(fù)。 針對(duì)基于虛擬化技術(shù)的GPU計(jì)算，基于虛擬機(jī)的GPU計(jì)算資源管理系統(tǒng)（簡(jiǎn)稱(chēng)VMGPURMS）主要包含支持虛擬機(jī)中GPU計(jì)算，并采用CPU-GPU混合調(diào)度策略提高系統(tǒng)利用率。系統(tǒng)采用共享內(nèi)存方式實(shí)現(xiàn)虛擬機(jī)和特權(quán)域之間的數(shù)據(jù)通訊，減少數(shù)據(jù)通訊帶來(lái)的GPU計(jì)算性能的損耗；提出了CPU-GPU混合調(diào)度的策略：針對(duì)CPU資源遠(yuǎn)遠(yuǎn)多于GPU資源的現(xiàn)狀，采用最少資源策略合理調(diào)度CPU和GPU作業(yè)，減少資源無(wú)益占據(jù)GPU作業(yè)的等待時(shí)間；針對(duì)GPU作業(yè)運(yùn)行需要CPU協(xié)助的特點(diǎn)，提出了處理器資源預(yù)留策略，為GPU作業(yè)準(zhǔn)備好配套的處理器和GPU資源；針對(duì)CPU作業(yè)占據(jù)GPU節(jié)點(diǎn)處理器資源，提出了GPU作業(yè)搶占的策略，把CPU作業(yè)遷移到其他節(jié)點(diǎn)上釋放GPU資源給GPU作業(yè)，既不影響CPU作業(yè)的運(yùn)行，也避免了GPU作業(yè)的等待。 基于虛擬機(jī)的GPU計(jì)算資源管理系統(tǒng)實(shí)現(xiàn)了在虛擬機(jī)層次上對(duì)處理器資源和GPU資源的協(xié)同管理和調(diào)度，有效結(jié)合利用虛擬機(jī)的良好特性和GPU的強(qiáng)勁計(jì)算能力。通過(guò)對(duì)系統(tǒng)的測(cè)試顯示，虛擬化環(huán)境中的GPU計(jì)算能力保持在物理機(jī)的85%；采用CPU-GPU混合調(diào)度策略對(duì)CPU和GPU協(xié)同調(diào)度策略后，整個(gè)系統(tǒng)的利用率提升了17%。
[Abstract]:The unified computing architecture allows the graphics processor (GPU) to upgrade to a general-purpose graphics processor, sharing the common computing tasks that were originally handled by the central processing unit (CPU), and obtaining several or even tens of times the acceleration ratio of the CPU. At the same time, with the maturity of virtualization technology, the combination of virtualization technology and GPU computing is more and more popular in high-performance computing. However, at present, there is no virtual machine-based GPU computing management system, either can not effectively use the powerful computing power of GPU in a virtualized environment, or the physical machine-based cluster is inconvenient and effective management of resources, Vulnerable to a malicious user and unable to recover quickly after crashing. For GPU computing based on virtualization technology, GPU computing resource management system (VMGPURMS) based on virtual machine mainly includes supporting GPU computing in virtual machine, and adopts CPU-GPU mixed scheduling strategy to improve system utilization. The system uses shared memory to realize the data communication between virtual machine and privileged domain, and reduces the loss of GPU computing performance caused by data communication. This paper proposes a hybrid scheduling strategy for CPU-GPU: in view of the fact that CPU resources are far more than GPU resources, the least resource strategy is adopted to schedule CPU and GPU jobs reasonably, so as to reduce the waiting time that resources do not benefit to occupy GPU jobs; According to the characteristic that CPU is needed in the operation of GPU job, a resource reservation strategy of processor is put forward to prepare the supporting processor and GPU resource for GPU job. In order to solve the problem that CPU jobs occupy the resources of GPU node processor, the strategy of preemption of GPU jobs is proposed. The CPU jobs are migrated to other nodes to free GPU resources to GPU jobs, which does not affect the running of CPU jobs and avoids the waiting of GPU jobs. The virtual machine-based GPU computing resource management system realizes the cooperative management and scheduling of processor resources and GPU resources at the virtual machine level, which effectively combines the good characteristics of virtual machines and the powerful computing power of GPU. The test of the system shows that the GPU computing power in the virtualized environment is kept at 85% of that of the physical machine, and the utilization of the whole system increases by 17% after the CPU-GPU hybrid scheduling strategy is applied to the CPU and GPU co-scheduling strategy.
【學(xué)位授予單位】：華中科技大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2012
【分類(lèi)號(hào)】：TP315
本文編號(hào)：2441733
[Abstract]:The unified computing architecture allows the graphics processor (GPU) to upgrade to a general-purpose graphics processor, sharing the common computing tasks that were originally handled by the central processing unit (CPU), and obtaining several or even tens of times the acceleration ratio of the CPU. At the same time, with the maturity of virtualization technology, the combination of virtualization technology and GPU computing is more and more popular in high-performance computing. However, at present, there is no virtual machine-based GPU computing management system, either can not effectively use the powerful computing power of GPU in a virtualized environment, or the physical machine-based cluster is inconvenient and effective management of resources, Vulnerable to a malicious user and unable to recover quickly after crashing. For GPU computing based on virtualization technology, GPU computing resource management system (VMGPURMS) based on virtual machine mainly includes supporting GPU computing in virtual machine, and adopts CPU-GPU mixed scheduling strategy to improve system utilization. The system uses shared memory to realize the data communication between virtual machine and privileged domain, and reduces the loss of GPU computing performance caused by data communication. This paper proposes a hybrid scheduling strategy for CPU-GPU: in view of the fact that CPU resources are far more than GPU resources, the least resource strategy is adopted to schedule CPU and GPU jobs reasonably, so as to reduce the waiting time that resources do not benefit to occupy GPU jobs; According to the characteristic that CPU is needed in the operation of GPU job, a resource reservation strategy of processor is put forward to prepare the supporting processor and GPU resource for GPU job. In order to solve the problem that CPU jobs occupy the resources of GPU node processor, the strategy of preemption of GPU jobs is proposed. The CPU jobs are migrated to other nodes to free GPU resources to GPU jobs, which does not affect the running of CPU jobs and avoids the waiting of GPU jobs. The virtual machine-based GPU computing resource management system realizes the cooperative management and scheduling of processor resources and GPU resources at the virtual machine level, which effectively combines the good characteristics of virtual machines and the powerful computing power of GPU. The test of the system shows that the GPU computing power in the virtualized environment is kept at 85% of that of the physical machine, and the utilization of the whole system increases by 17% after the CPU-GPU hybrid scheduling strategy is applied to the CPU and GPU co-scheduling strategy.
【學(xué)位授予單位】：華中科技大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2012
【分類(lèi)號(hào)】：TP315
【參考文獻(xiàn)】
相關(guān)期刊論文 前1條
1 劉真;;虛擬機(jī)技術(shù)的復(fù)興[J];計(jì)算機(jī)工程與科學(xué);2008年02期
本文編號(hào)：2441733
本文鏈接：http://sikaile.net/kejilunwen/jisuanjikexuelunwen/2441733.html
最近更新
教材專(zhuān)著
