When queueing meets coding: Optimal-latency data retrieving scheme in storage clouds

Abstract

Storage clouds, such as Amazon S3, are being widely used for web services and Internet applications. It has been observed that the delay for retrieving data from and placing data into the clouds is quite random, and exhibits weak correlations between different read/write requests. This inspires us to investigate a key problem: can we reduce the delay by transmitting data replications in parallel or using powerful erasure codes?In this paper, we study the problem of reducing the delay of downloading data from cloud storage systems by leveraging multiple parallel threads, assuming that the data has been encoded and stored in the clouds using fixed rate forward error correction (FEC) codes with parameters (n, k). That is., each file is divided into k equal-sized chunks, which are then expanded into n chunks such that any k chunks out of the n are sufficient to successfully restore the original file. The model can be depicted as a multipleserver queue with arrivals of data retrieving requests and a server corresponding to a thread. However, this is not a typical queueing model because a server can terminate its operation, depending on when other servers complete their service (due to the redundancy that is spread across the threads). Hence, to the best of our knowledge, the analysis of this queueing model remains quite uncharted.Real traces from Amazon S3 show that the time to retrieve a fixed size chunk is random and can be accurately approximated as an i.i.d. exponentially distributed random variable. We show that any work-conserving scheme is delay-optimal when k = 1. When k > 1, we find that a simple greedy scheme, which allocates all available threads to the head of line request, is delay optimal, which appears surprising.

Mean-field-analysis of coding versus replication in cloud storage systems

“…Finally, we make a comment on the scenario that is being modeled in our paper and some of the other prior works (e.g., [4], [7], [8], [10]- [12]). Our work views the problem from the point of view of storage service provider.…”

“…Our work views the problem from the point of view of storage service provider. On the other hand, the previous works (e.g., [4], [7], [8], [10]- [12]) view the problem from the point of view of a customer who uses a cloud storage system. Thus, in these other works, the service time of a file is a complicated function of one's own file size, the storage server's speed and the service provided to other customers.…”

Mean-Field Analysis of Coding Versus Replication in Large Data Storage Systems

“…Finally, we make a comment on the scenario that is being modeled in our paper and some of the other prior works (e.g., [5], [8], [9], [11]- [13]). Our work views the problem from the point of view of storage service provider.…”

Joint latency and cost optimization for erasurecoded data center storage

“…However, the approach cannot be applied to a multiple-heterogeneous file storage where each file has a separate folk-join queue and the queues of different files are highly dependent due to shared storage nodes and joint request scheduling. In another work [3], the authors applied this fork-join queue to optimize threads allocation to each request, which is similar to our weighted queue model, however, both proposed greedy/shared scheme would waste system resources because in fork-join queue there will always be some threads have unfinished downloads due to redundant assignment. In addition, in [7], the authors proposed a distributed storage system which analyzed through the Fork-join queue framework with heterogeneous jobs, and provide lower and upper bounds on the average latency for jobs of different classes under various scheduling policies, such as First-Come-First-Serve, preemptive and nonpreemptive priority scheduling policies, based on the analysis of mean and second moment of waiting time.…”

“…Remark 1: Consider the homogeneous case studied in previous work [3,40,45,50] where all nodes have the same service time distribution and where files have the same chunk placement (i.e., |S i | = n i = m ∀i ). It is easy to show that due to symmetry, the optimal scheduling probabilities π i,j minimizing total system latency is π i,j = k i /m for all i, j.…”