Can'T We All Get Along? Redesigning Protection Storage For Modern Workloads

PROCEEDINGS OF THE 2018 USENIX ANNUAL TECHNICAL CONFERENCE(2018)

引用 24|浏览154
暂无评分
摘要
Deduplication systems for traditional backups have optimized for large sequential writes and reads. Over time, new applications have resulted in nonsequential accesses, patterns reminiscent of primary storage systems. The Data Domain File System (DDFS) needs to evolve to support these modern workloads by providing high performance for nonsequential accesses without degrading performance for traditional backup workloads.Based on our experience with thousands of deployed systems, we have updated our storage software to distinguish user workloads and apply optimizations including leveraging solid-state disk (SSD) caches. Since SSDs are still significantly more expensive than magnetic disks, we make our system cost-effective by caching metadata and file data rather than moving everything to SSD. We dynamically detect access patterns to decide when to cache, prefetch, and perform numerous other optimizations. We find that on a workload with nonsequential accesses, with SSDs for caching metadata alone, we measured a 5.7x improvement on input/output operations per second (IOPS) when compared to a baseline without SSDs. Combining metadata and data caching in SSDs, we measured a further 1.7x IOPS increase. Adding software optimizations throughout our system added an additional 2.7x IOPS improvement for non-sequential workloads. Overall, we find that both hardware and software changes are necessary to support the new mix of sequential and nonsequential workloads at acceptable cost. Our updated system is sold to customers worldwide.
更多
查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要