III: Small: Avoiding Contention on Multicore Machines

为了充分利用多核机器提供的并行性，必须编写并行代码。编写并行代码是困难的。即使编写了正确的代码，也会有许多性能缺陷。例如，一个无法识别的数据热点可能意味着所有线程都有效地序列化了它们对该热点的访问，吞吐量将大大降低。该项目旨在提供一个通用框架，用于并行执行某些类型的并发操作。基础设施提供了在多核机器的可用线程上以可扩展的方式执行这些操作，自动响应热点和其他性能危害。我们的目标不是榨取特定平台的最后一滴性能。相反，对于计划好的系统，程序员可以在没有详细的并发和并行编程知识的情况下，开发出有效利用多核机器的代码。该项目涉及算法和数据结构的开发，旨在有效地并行执行通用代码片段。主要关注的是数据密集型操作，通常会在内存数据库引擎中找到。关键的研究问题包括如何设计可应用于一系列计算的通用多线程操作符，如何避免缓存刷散，以及如何以一种可在各种硬件平台上工作的方式实现框架。相对于受争用影响的原始解决方案，期望吞吐量的性能提高一个数量级。该项目的目标是实现接近手工定制的专家编写的并行代码的性能，而编码工作要少得多。这个项目在商业和公共领域数据库系统中都有直接的应用，其中性能改进将增强数据库系统用户的体验，并减少给定性能水平下的硬件和能源需求。可编程性的改进将允许没有并行编程专业知识的程序员有效地使用多核机器。该项目还为多核机器数据库系统实现的高级课程提供了重点。软件基础设施将提供给其他人进行研究使用。

To take full advantage of the parallelism offered by a multicoremachine, one must write parallel code. Writing parallel code isdifficult. Even when one writes correct code, there are numerousperformance pitfalls. For example, an unrecognized data hotspot couldmean that all threads effectively serialize their access to thehotspot, and throughput is dramatically reduced.This project aims to provide a generic framework for performingcertain kinds of concurrent operations in parallel. Infrastructure isprovided to perform those operations in a scalable way over theavailable threads in a multicore machine, automatically responding tohotspots and other performance hazards. The goal is not to squeezethe last drop of performance out of a particular platform. Rather,with the planned system a programmer can, without detailed knowledgeof concurrent and parallel programming, develop code that efficientlyutilizes a multicore machine.The project involves the development of algorithms and data structuresdesigned for the efficient parallel execution of generic codefragments. The primary focus is on data intensive operations as wouldtypically be found in an in-memory database engine. Critical researchquestions include how to design generic multi-threaded operators thatcan be applied to a range of computations, how to avoid cachethrashing, and how to implement the framework in a way that works on avariety of hardware platforms. Performance improvements in throughputof an order of magnitude are expected relative to naive solutions thatsuffer from contention. The project aims to achieve performance closeto that of hand-tailored expert-written parallel code, with far lesscoding effort.This project has immediate applications in both commercial andpublic-domain database systems where performance improvements wouldenhance the experience of database system users, and reduce hardwareand energy requirements for a given level of performance.Programmability improvements would allow programmers without expertisein parallel programming to effectively use multicore machines. Theproject also provides the focus for an advanced-level course ondatabase system implementation for multicore machines. The softwareinfrastructure will be made available for research use by others.