Due to the increasingly data-intensive nature of the applications, the storage system performance continues to increase in importance and is often substantially responsible for the overall processing rate of the application. Fortunately, the storage technologies themselves are improving rapidly in numerous ways, from low-level read/write of bits in a device all the way to the management of the entire storage hierarchy in large enterprise and cloud settings. Studying many of the important issues in this entire spectrum often requires storage access traces from the storage server side, but these are often hard to come by. To address this gap, we present a method to generate synthetic traces using a novel generative adversarial network (GAN) architecture that captures the realism and diversity of real storage traces. The generated traces can be used to augment the existing workload traces of interest for a variety of storage system studies. We demonstrate how the proposed method can generate storage traces that have the overall characteristics of the real traces and yet provide behavioral diversity.

In this paper, we address the characterization of similarity between two storage traces and define a three-part measure called SIST (Similarity Index for Storage Traffic). Such a measure is essential for identifying traces that are most appropriate for storage system evaluations. We compare SIST against several other similarity measures in the literature on both the object storage and block storage systems, and show the superiority of SIST in terms of its behavior for known perturbations.

The increase in data generation and the decrease in the cost of storage increases the need for intelligent data management. One avenue that would allow storage systems to manage data better is to feed it with an accurate prediction of how many disk operations are expected on that data. In this paper, we introduce a method to predict the data heat in a storage system. Our method is derived from two insights. The first is that we can use a set of quick to compute signals which constrain the set of possible future heat patterns. The underlying assumption in this is that the signals provide a description of the access pattern and the requests that have similar signals have a similar set of future behavior. The second is that the storage requests can be partitioned into groups based on their signal. Our method generates a unique prediction model for each group from the corresponding heat patterns. This makes the model generation process easier and more precise. The results of our method on public datasets show that it is a viable way to predict heat. Our method is able to accurately almost all inactive regions and provides useful predictions for active regions.