CSR: Small: Cost Effective, High Performance Solutions Using Erasure Codes for Big Data Management in Large Data Centers

在todayÕs大型数据中心中，数据量和复杂性都在急剧增加，数据和I/O可用性日益受到关注。大多数现有的解决方案都基于多复制技术来提供数据冗余，其中数据块在存储服务器节点之间复制。然而，多重复制技术不足以管理大数据：itÕs有效地复制数十到数百pb的数据集的N个副本是一个巨大的挑战！作为一种替代解决方案，允许多次故障的擦除码可以以低得多的成本提供可靠性和可用性。然而，使用纠删码管理大数据面临的最大挑战是由于编码/解码操作复杂而导致的性能问题，这限制了纠删码在大规模数据中心的应用。本项目通过多项研究创新，开发具有成本效益的技术，利用擦除码实现高可用性，提高大型数据中心的性能，从而有效地管理大数据。该项目集中研究了如何利用适当的空间成本和系统/架构技术来提高基于擦除代码的服务器集群的整体数据访问性能。本研究为有效部署使用擦除码的数据中心铺平了道路。它有可能使许多大数据应用程序受益，例如在线搜索、社交网络、电子商务、医疗保健等典型的数据密集型应用程序。

Data and I/O availability is an increasing concern in todayÕs large data centers where both data volume and complexity are increasing dramatically. Most existing solutions are based on multi-replication techniques to provide data redundancy, where data chunks are replicated across storage server nodes. However, multi-replication techniques are insufficient to manage big data: itÕs a big challenge to efficiently replicate N copies of a data set of tens-to-hundreds of petabytes! As an alternative solution, erasure codes tolerating multiple failures can provide reliability and availability at much lower cost. However, the biggest challenge using erasure codes to manage big data is the performance problem due to the complex encoding/decoding operations, which limits the application of erasure codes in large-scale data centers. This project develops cost effective techniques to exploit erasure codes to achieve high availability and enhance performance in large data centers to efficiently manage big data via several research innovations. This project cohesively investigates how to utilize proper spatial cost and system/architecture techniques to improve the overall data access performance of server clusters built upon erasure codes. This research has fundamental contributions to pave the way to efficiently deploy data centers using erasure codes. It has potential to benefit numerous big data applications such as online searching, social network, e-business, health care, and so on which are typically data intensive.