Power-Saving in Storage Systems for Cloud Data Sharing Services with Data Access Prediction

Abstract

SUMMARYWe present a power-saving method for large-scale storage systems of cloud data sharing services, particularly those providing media (video and photograph) sharing services. The idea behind our method is to periodically rearrange stored data in a disk array, so that the workload is skewed toward a small subset of disks, while other disks can be sent to standby mode. This idea is borrowed from the Popular Data Concentration (PDC) technique, but to avoid an increase in response time caused by the accesses to disks in standby mode, we introduce a function that predicts future access frequencies of the uploaded files. This function uses the correlation of potential future accesses with the combination of elapsed time after upload and the total number of accesses in the past. We obtain this function in statistical analysis of the real access patterns of 50,000 randomly selected publicly available photographs on Flickr over 7,000 hours (around 10 months). Moreover, to adapt to a constant massive influx of data, we propose a mechanism that effectively packs the continuously uploaded data into the disk array in a storage system based on the PDC. To evaluate the effectiveness of our method, we measured the performance in simulations and a prototype implementation. We observed that our method consumed 12.2% less energy than the static configuration (in which all disks are in active mode). At the same time, our method maintained a preferred response time, with 0.23% of the total accesses involving disks in standby mode.

Analysis of Optimal File Placement for Energy-Efficient File-Sharing Cloud Storage System

“…'s Hadoop cluster, GreenHDFS [7] allocates the disks either hot or cold zone, and replace the data according to the age of data. Hasebe et al [8] used file access traces observed in Flickr to derive the individual file access frequencies and proposed the file exchange algorithm to skew the disk access frequencies. Some recent studies further took into account the correlation of file accesses to determine the file placement for energy saving [36] [37].…”

“…Such web access patterns are commonly observed in other web services and are often approximated by Zipf distributions [40] [41]. In our experimental study in Section 6, we also use the access traces of Flickr investigated in the previous studies [6] [8].…”

“…While Poisson arrivals were observed in real systems [45] [46], the validity of the first assumption A1 on the constant file access rate is arguable, since the popularity of sharing files can change over time as observed in real traces [8]. The solution to the energy-saving file placement problem only gives the best placement under the predicted file access rates for the next period.…”

“…It also clarifies that the necessary condition where the heuristic method achieves the optimal file placement satisfying the performance constraint. To confirm the effectiveness of the optimal file placement on a real storage system, we conducted experiments on our testbed system consisting of four disks with the real file access frequencies obtained from Flickr file access traces [8]. Our experimental results show the optimal file placement can cut 31.8% of energy overheads steadily compared with the baseline file placement in which file accesses are evenly distributed across the hard disks.…”

“…Two essential functions in the implementation of PDC are the prediction of data access frequencies and the data placement method. While the prediction is usually made by analyzing historical workload data, the placement method typically relies on a heuristic algorithm that places data on disks in order of data access frequencies [5] [6] [8] [12]. The heuristic is intuitive, and the effectiveness was validated through some simulation studies [5] [6] [8].…”