Efficient SSD Caching by Avoiding Unnecessary Writes using Machine Learning

Abstract

SSD has been playing a significantly important role in caching systems due to its high performance-to-cost ratio. Since cache space is much smaller than that of the backend storage by one order of magnitude or even more, write density (writes per unit time and space) of SSD cache is therefore much higher than that of HDD storage, which brings about great challenges to SSD's lifetime. Meanwhile, under social network workloads, quite a few writes on SSD are unnecessary, e.g., Tencent's photo caching shows that about 61% of total photos are just accessed once whereas they are still swapped in and out of the cache. Therefore, if we can predict this kind of photos proactively and prevent them from entering the cache, we can eliminate unnecessary SSD cache writes and improve cache space utilization. To cope with the challenge, we put forward a one-time-access criteria that is applied to cache space, and further propose a one-time-access-exclusion policy. Based on that, we design a prediction-based classifier to facilitate the policy. Unlike the state-of-the-art history-based predictions, our prediction is non-history-oriented, which is challenging to achieve a good prediction accuracy. To address this issue, we integrate a decision tree into the classifier, extract social-related information as classifying features, and apply cost sensitive learning to improve classification precision. Due to these techniques, we attain a predication accuracy over 80%. Experimental results show that one-time-access-exclusion approach makes caching performance outstanding in most aspects, taking LRU for instance, hit rate is improved by 17%, cache writes are decreased by 79%, and the average access latency is dropped by 7.5%.