[Paper Review] FAST CLOUD: Pushing the Envelope on Delay Performance of Cloud Storage with Coding

Guanfeng Liang, Ulaş C. Kozat|arXiv (Cornell University)|Jan 7, 2013

Advanced Data Storage Technologies15 references6 citations

TL;DR

This paper proposes FAST CLOUD, a load-adaptive coding and chunking framework that significantly reduces cloud storage access delays by combining erasure coding, parallel connections, and dynamic code rate selection. It achieves up to 85% reduction in 99th percentile delay for 2 MB files with only 1.75× storage overhead, outperforming fixed-rate policies through queue-aware adaptation based on real-time backlog thresholds.

ABSTRACT

Our paper presents solutions that can significantly improve the delay performance of putting and retrieving data in and out of cloud storage. We first focus on measuring the delay performance of a very popular cloud storage service Amazon S3. We establish that there is significant randomness in service times for reading and writing small and medium size objects when assigned distinct keys. We further demonstrate that using erasure coding, parallel connections to storage cloud and limited chunking (i.e., dividing the object into a few smaller objects) together pushes the envelope on service time distributions significantly (e.g., 76%, 80%, and 85% reductions in mean, 90th, and 99th percentiles for 2 Mbyte files) at the expense of additional storage (e.g., 1.75x). However, chunking and erasure coding increase the load and hence the queuing delays while reducing the supportable rate region in number of requests per second per node. Thus, in the second part of our paper we focus on analyzing the delay performance when chunking, FEC, and parallel connections are used together. Based on this analysis, we develop load adaptive algorithms that can pick the best code rate on a per request basis by using off-line computed queue backlog thresholds. The solutions work with homogeneous services with fixed object sizes, chunk sizes, operation type (e.g., read or write) as well as heterogeneous services with mixture of object sizes, chunk sizes, and operation types. We also present a simple greedy solution that opportunistically uses idle connections and picks the erasure coding rate accordingly on the fly. Both backlog and greedy solutions support the full rate region and provide best mean delay performance when compared to the best fixed coding rate policy. Our evaluations show that backlog based solutions achieve better delay performance at higher percentile values than the greedy solution.

Motivation & Objective

• To address the high variability and poor delay performance in cloud storage systems like Amazon S3, especially for small-to-medium objects.
• To analyze the impact of chunking, erasure coding, and parallel access on queuing and service delays in cloud storage.
• To design load-adaptive algorithms that dynamically select code rates based on queue backlog thresholds to minimize end-to-end delay.
• To support both homogeneous and heterogeneous workloads with mixed object sizes, chunk sizes, and operation types (read/write).
• To develop a greedy, on-the-fly adaptation strategy that opportunistically uses idle connections and adjusts coding rates in real time.

Proposed method

• Proposes a queuing model that extends M/G/k by incorporating interdependent service times due to erasure coding, where completion of any k out of n coded fragments suffices.
• Derives a closed-form approximation for mean delay in systems with chunking, FEC, and multiple servers, accounting for load-dependent service rates.
• Introduces a load-adaptive algorithm that selects optimal code rates per request using precomputed queue backlog thresholds to minimize delay.
• Develops a greedy online algorithm that dynamically adjusts coding rate based on available idle connections without prior knowledge of load.
• Uses a convex optimization framework to model the trade-off between delay and load, with the objective function incorporating both queueing and service delays.
• Establishes theoretical conditions under which optimal code vectors lie within the feasible region, ensuring convergence and stability of the adaptation policy.

Experimental results

Research questions

• RQ1 How does erasure coding combined with chunking and parallel access affect the delay distribution in cloud storage systems like Amazon S3?
• RQ2 What is the impact of code rate and chunking on the supportable request rate and queueing delay in a cloud storage system?
• RQ3 Can dynamic code rate adaptation based on real-time queue backlog improve delay performance beyond fixed-rate coding policies?
• RQ4 How does the system perform under heterogeneous workloads with mixed object sizes, chunk sizes, and operation types?
• RQ5 What is the trade-off between storage overhead and delay reduction when using adaptive erasure coding in cloud storage?

Key findings

• For 2 MB files, FAST CLOUD reduces mean delay by 76%, 90th percentile delay by 80%, and 99th percentile delay by 85% compared to uncoded storage.
• The proposed load-adaptive backlog-based algorithm achieves better delay performance at higher percentiles than the greedy online method.
• The system supports the full rate region of uncoded storage while maintaining lower average delay, even under high load.
• Erasure coding with chunking improves delay performance only when combined with load-aware rate adaptation; otherwise, it degrades performance due to increased load.
• The optimal code rate is not fixed but depends on current load and queue backlog, with lower code rates extending capacity at high utilization.
• Theoretical analysis confirms that good code vectors are totally ordered by their performance, enabling efficient search and adaptation in real time.