I/O Scheduling For Mixed Workloads Using Ultra-Low Latency SSDs

To improve data center resource utilization and reduce costs,different types of applications are often deployed on the same server.However,since throughput-oriented applications often cause significant I/O interference to latency-sensitive applications,it is critical to provide low-latency I/O services for latency-sensitive applications to respond quickly to client requests.Ultra-low latency SSDs(ULL SSD),sush as Optane SSD,can provide extremely low latency compared to traditional SSDs.However,latency-sensitive applications that simply use ULL SSDs in the data center do not take full advantage of the features of ULL SSDs.Since I/O interference occurs in multi-layer I/O stacks,it is worthwhile to discuss how to reduce I/O latency for latency-sensitive applications while avoiding severe performance impact on throughput-oriented applications.Detailed analysis of the read and write paths of the storage stack shows that I/O interference mainly occurs at three layers: compound transactions commit in the file system,resource contention in device,and costly process scheduling.To solve the above problems,FastResponse,a holistic cross-stack approach,identifies latency-sensitive applications on the multi-layer I/O stack and adopt appropriate scheduling methods to reduce I/O contention.First,it proposes a fine-grained journaling scheme at the file system layer to commit filelevel transactions,thus avoiding long latency due to compound transaction.Second,it provides a new I/O scheduler at the block layer to throttle I/O requests of throughputoriented applications,thus reducing resource contention in device.Finally,it redesigns the Completely Fair Scheduler(CFS)to promote the priority of latency-sensitive applications using a preemption approach,allowing them to be scheduled quickly upon wake-up.Experiments show that,compared to the Linux kernel and the state-of-the-art work SelectISR,FastResponse can reduce the average response time of latency-sensitive applications by 18%～70% and 10%～67%,respectively,and reduce the 99 th percentile response time by 58%～80% and 52%～78%,respectively,with no more 6% performance degradation for throughput-oriented applications.