Optimization Techniques and Language Abstractions for GPGPU in Ruby

我们的研究重点是将快速面向对象编程（OOP）引入gpu。我们改进了之前开发的CUDA框架“dynasar”。DynaSOAr的主要组件之一是面向对象的单方法多对象（SMMO）应用程序的动态内存管理器。对DynaSOAr的主要改进是“CompactGpu”，一个内存碎片整理系统。CompactGpu可以通过以更密集的格式存储分配来提高内存受限GPU应用程序的性能。CompactGpu在物理上重新排列内存中的对象，以便将它们存储在几个紧凑的内存块中。对于动态内存分配，这种碎片通常是由不利的分配-释放模式造成的。内存碎片整理被证明对两种应用程序特别有用：(1)只使用少量内存的应用程序，如果存储密集，则大部分应用程序适合L1/L2缓存。(2)已经受益于数组结构（SOA）内存布局的应用程序会变得更快，因为如果数据以紧凑的形式存储，矢量加载/存储会更有效。我们的工作发表在ISMM 2019上，在我们的基准应用程序中实现了高达16%的加速。

Our research was centered around bringing fast object-oriented programming (OOP) to GPUs. We improved our previously developed CUDA framework "DynaSOAr". One of DynaSOAr's main components is a dynamic memory manager for object-oriented Single-Method Multiple-Objects (SMMO) applications.The main improvement to DynaSOAr was "CompactGpu", a memory defragmentation system. CompactGpu can improve the performance of memory-bound GPU applications by storing allocations in a denser format. CompactGpu physically rearranges objects in memory, so that they are stored in few, compact blocks of memory. With dynamic memory allocation, such fragmentation is often caused by unfavorable allocate-deallocate pattern.Memory defragmentation proved to be particularly useful for two kinds of applications: (1) Applications that use only a small amount of memory and fit largely into the L1/L2 cache if stored densely. (2) Applications that already benefit from a Structure of Array (SOA) memory layout get even faster, because vector load/stores are more efficient if data is stored in a compact form. Our work, published at ISMM 2019, achieved a speedup of up to 16% in our benchmark applications.