SBIR Phase I: A Multithreaded Storage Engine using Highly-Concurrent Fractal Trees

这个小型企业创新研究第一阶段项目将研究在多核计算机上实现高性能数据库的技术，重点是如何支持并发活动，并在“分形树”数据库中提供可证明的良好线程调度。 当今的数据库存在存储带宽、磁盘寻道速率和CPU核心容量之间的资源不平衡问题，导致性能不佳、解决方法繁琐和能源效率低下。 该公司为MySQL开发了一个高性能的存储引擎，它在实时数据上维护索引的速度比传统引擎快100倍。该方法采用了缓存无关的分形树索引，其规模与存储带宽，而不是寻求率，从而解决带宽和磁盘寻道率之间的不平衡。 如果成功的话，这项研究将产生一个数据库的实现，为每个查询，要么饱和的CPU内核，饱和的磁盘带宽，或消耗所有的固有的并行查询。目标市场包括具有非常大的数据库和插入和查询占主导地位的工作负载的组织。 有许多应用领域不使用数据库，因为它们的性能太慢。数据库的数量级加速可以帮助市场增长。 金融、零售、国土安全、电信和科学计算领域的应用将受益于高性能数据库。此外，研究人员希望引导所有数据库实现者进入多核领域。 建议的研究将进一步了解如何调度数据库查询时，数据在磁盘上布局良好。随着用户对数据的需求继续超过快速内存的可用性，组织基于磁盘的数据的多线程查询以提高性能的重要性只会越来越大。

This Small Business Innovation Research Phase I project will investigate techniques for implementing high-performance databases on multi-core computers by focusing on how to support concurrent activity with provably good thread scheduling in "Fractal Tree" databases. Today's databases suffer from resource imbalances between storage bandwidth, disk-seek rate, and CPU core capacity, leading to underperformance, cumbersome workarounds, and energy inefficiency. The company has developed a high-performance storage engine for MySQL that maintains indexes on live data 100 times faster than traditional engines. The approach employs cache-oblivious Fractal-Tree indexes, which scale with storage bandwidth rather than seek rate, thus addressing the imbalance between bandwidth and disk-seek rate. If successful, this research will produce a database implementation that for each query that either saturates the CPU cores, saturates disk bandwidth, or consumes all of the inherent parallelism in the query.The target market comprises organizations that have very large databases and a workload dominated by insertions and queries. There are many application areas that do not employ databases because their performance is too slow. Orders-of-magnitude speedup for databases can help grow the market. Applications in finance, retail, homeland security, telecommunications, and scientific computing will benefit from high-performance databases. Furthermore the researchers hope to lead all database implementers into the multi-core realm. The proposed research will further the understanding of how to schedule database queries when data is well laid out on disk. As users' appetite for data continues to outstrip the availability of fast memory, organizing multithreaded queries on disk-based data for performance will only grow in importance.