CAJ | 학술논문

虚拟化技术能够以较低的成本和能源消耗共享有效的资源，一些应用程序往往需要利用图形处理器来加快它们的计算以提高性能。但是由于虚拟化本身的特点，在GPU虚拟化环境下进行CUDA应用开发会带来很大的性能开销。此外，当采用多GPU并行处理大规模的程序时，传统的GPU之间的数据交互方式是通过CPU来中转，不仅会带来“路程”上的开销，同时PCI-E相对于GPU显存的低带宽更是限制了数据传输的速率。针对以上问题，文中在Xen和VMware虚拟化平台下，针对CUDA应用的延迟和吞吐率找出最优的虚拟机间通讯方式，针对GPU之间不同的数据传输方式，找出最优通信方案，并从理论上和实验中分析出影响多GPU协同运算效率的因素。
허의화기술능구이교저적성본화능원소모공향유효적자원，일사응용정서왕왕수요이용도형처리기래가쾌타문적계산이제고성능。단시유우허의화본신적특점，재GPU허의화배경하진행CUDA응용개발회대래흔대적성능개소。차외，당채용다GPU병행처리대규모적정서시，전통적GPU지간적수거교호방식시통과CPU래중전，불부회대래“로정”상적개소，동시PCI-E상대우GPU현존적저대관경시한제료수거전수적속솔。침대이상문제，문중재Xen화VMware허의화평태하，침대CUDA응용적연지화탄토솔조출최우적허의궤간통신방식，침대GPU지간불동적수거전수방식，조출최우통신방안，병종이론상화실험중분석출영향다GPU협동운산효솔적인소。
Virtualization,as a technology that enables easy and effective resource sharing with a low cost and energy footprint,applications with stringent performance often need to make use of graphics processors for accelerating their computations. Due to the characteristics of virtualization itself,it will be brought significant performance overhead when doing CUDA application development under GPU virtualiza-tion environments. Besides,it will be studied when processing a large-scale data by using multi-GPU in parallel way. The traditional way of data exchange between the GPUs is transferring by the CPU. Thus it not only will bring the"walk" cost,but also limit the bandwidth of the data transfer rate by using PCI-E with respect to the GPU bandwidth. To solve the above problems,try to identify the optimal com-munication between virtual machines under Xen and VMware virtualization platform for CUDA applications in this paper. For different data transmission between the GPUs,try to find the optimal communication scheme,and to analyze the factors that affect the efficiency of multi-GPU collaborative computing through theory and experiment.