论文阅读——GPU Virtualization and Scheduling Methods: A Comprehensive Survey

一篇关于 GPU 虚拟化 的 survey 文章，发表于 ACM Computing Surveys

Background 里面一段比较有意思的话：

On the contrary, the design of conventional processors is optimized
for reducing the execution time of sequential code on each core, thus adding complexity
to each core at the cost of offering fewer cores in the processor package. Conventional
processors typically use sophisticated control logic and large cache memories to efficiently
deal with conditional branches, pipeline stalls, and poor data locality.

传统处理器的目标是尽可能加速每个核心的串行执行时间，因此每个核心有大量处理分支预测，流水线延迟，data locality 相关的资源。代价是核心数量较少