An approach for impact analysis of flash behavior on QoS in DC/Enterprise SSDs

In consumer and enterprise world, SSD Performance is the main quality constraint. SSD performance parameters are classified in terms of lOPS, Throughput, latency and Quality of service(QoS). SSDs processes millions bytes of data with certain latency and throughput for read, write and mixed operations. But Quality of service is not guaranteed by SSD vendors for single user. However, Enterprise/server SSD storage must meet a Quality of Service (QoS) level with the given requirement to ensure the study-state performance over long period of time. Quality of service(QoS) is measurement of overall and consistent performance of services. Which is very critical for data center and cloud services. In this paper, we will discuss about QoS measurement along with statistics of latency(outliers/percentile/confidence values), results and analysis methods. There are few factors like Workload, host interface speed, host I0 stack and NAND characteristics which impacts QoS performance. For results, We measured multiple levels of quantile and percentile (2 9's ~ 7 9's) for read intensive/write intensive/mixed workloads on multiple enterprise SSDs. Also, we will discuss about different confidence values for both Read/Write workloads, which will help us to identify the differences in QoS values of new and previous generation SSDs. QoS has great dependency on the SSD FW policies like over provisioning, garbage collection and endurance. During paper presentation, we intend to present these parameters and their impact on QoS. To demonstrate QoS behavior, We would also like to discuss about following use case in detail "QoS behavioral changes with NAND characteristics". Few important outcomes of this use case as follows 1. MLC based SDs have better QoS performance for read intensive applications. 2. Write intensive and mixed workload QoS performance is good in TLC based SSDs. 3. Enterprise SDs have more over provisioning range ~ 30% than brand SSDs to get sustained QoS 4. Better NAND timing (tR/tPROG/tBERS) parameters have good Quality of service.