论文部分内容阅读
Modern applications often require a large amount of memory. Conventional 4KB pages lead to large page tables and thus exert high pressure on TLB address translations. This pressure is more prominent in a virtualized system, which adds an additional layer of address translation. Page walks due to TLB misses can result in a significant performance overhead. One effort in reducing this overhead is to use hugepage. Linux kernel has supported transparent hugepage since 2.6.38, which provides an alternate large page size. Generally,hugepage demonstrates better performance on address translations and page table modifications. This paper first analyzes the impact of hugepage on native system, and then, compares the impact of hugepage on different memory virtualization approaches: hardware-assisted paging(HAP), shadow paging, and para-virtualization.We observe that the current implementation of transparent hugepage is inefficient. It cannot exploit the full performance advantage of hugepages. Worse yet, the conservative strategy of transparent hugepage may conflict with existing OS functions, which can lead to performance degradation. So, we propose a new memory allocation strategy, alignment-based hugepage(ABH) that promotes hugepage allocations. We apply ABH to different paging modes in virtualized systems. The results show that the new allocation strategy can significantly reduce TLB misses and up to 90% page walk cycles due to TLB misses and thus improve the performance in real world applications.
Modern applications often require a large amount of memory. Conventional 4KB pages lead to large page tables and thus exert high pressure on TLB address translations. This pressure is more prominent in a virtualized system, which adds an additional layer of address translation. Page walks due One effort in reducing this overhead is to use huge page. Linux Kernel has supported transparent hugepage since 2.6.38, which provides an alternate large page size. Generally, huge page demonstrates better performance on address translations and page table modifications. This paper first analyzes the impact of hugepage on native system, and then, comparable the impact of hugepage on different memory virtualization approaches: hardware-assisted paging (HAP), shadow paging, and para-virtualization. We observe that the current implementation of transparent hugepage is inefficient. It can not exploit the full performance advantage of hugepages. W orse yet, the conservative strategy of transparent huge page may conflict with existing OS functions, which can lead to performance degradation. So, we propose a new memory allocation strategy, alignment-based hugepage (ABH) that requires huge page allocations. We apply ABH to different paging modes in virtualized systems. The results show that the new allocation strategy can reduce TLB misses and up to 90% page walk cycles due to TLB misses and thus improve the performance in real world applications.