Practical implication of analytical models for SSD write amplification

Abstract

A number of analytical models have been proposed to estimate the write amplification of the Flash storage to obtain the expected lifespan. This work is dedicated to examining the practical implication of the four existing analytical models for estimating the write amplification: Coupon Collector, Uniform Distribution, Expected Value and Markov model. Since the models assume uniform random workload in full utilization of an SSD to predict write amplification, they are not applicable in predicting write amplification in general workload. Moreover, the existing models have not been verified with the real SSD. In this work, we compare the write amplification of the models with that of a real SSD. When we use 0.147 as the overprovisioning factor of an SSD while running uniform random workload, the write amplification of Uniform Distribution, Expected Value, Markov model is 3.90, 4.08, and 4.08, respectively. However, write amplification of the real SSD shows 1.19, which is very different from that of the prediction models. Through experiment, we found that write amplification is closely related to the value of overprovisioning factor. To improve the accuracy of existing prediction models, we update the overprovision-ing factor to take account of the ratio of a hot file and the utilization of the storage. We also find that by setting the overprovisioning factor to 1.15, we can obtain write amplification of 1.2 which is close to the write amplification of general workload in a real SSD. © 2017 ACM.