SHF: Small: Architectural Support for New Parallel Execution Paradigms Via Agile Threads

在多核处理器时代，微处理器只会在存在大量线程级并行性的情况下继续扩展性能。要实现这个不断扩展软件并行性的目标，显然需要利用新的编译、语言和执行范例。许多被提议的执行范例的可行性和新范例的引入的一个巨大障碍是现代处理器体系结构无法有效地执行短线程。如果这些新的执行模型可以自由地识别和利用长达10到100条指令的并行性，那么它们中的许多都可以非常有效地将并行性暴露给硬件。然而，当前的机器不能很好地执行短线程。本研究的目标是显著降低一个新线程（或一个核心新线程）的启动成本。这反过来又降低了决定一段代码是否可并行化的盈亏平衡点。术语“线程迁移”用于表示大量并行执行操作，所有这些操作都涉及将存储或缓存状态从一个核心移动到另一个核心。这些操作包括分叉线程、用于热管理或负载平衡的迁移/移动线程、循环并行线程、任务级并行、helper线程、事务执行和推测多线程——所有这些操作都将通过本研究在一定程度上得到加速。本研究将攻击线程启动成本的所有来源，包括软件（如操作系统）开销、分支预测器状态、缓存数据和指令状态、提交延迟以及在内核之间传输主线程状态的开销。除了降低并行编程的复杂性，这项研究的更广泛的影响包括研究生和本科生的培训，开源模拟基础设施的可用性。

In the multi-core processor era, microprocessors will only continue to scale in performance in the presence of abundant thread level parallelism. Achieving this goal of continuously scaling software parallelism will clearly require the exploitation of new compilation, language, and execution paradigms. One huge barrier to the viability of many proposed execution paradigms and the introduction of new paradigms is the inability of modern processor architectures to execute short threads efficiently. Many of these new execution models can be highly effective at exposing parallelism to the hardware if they have the freedom to identify and exploit opportunities for parallelism that are 10s to 100s of instructions long. However, current machines are not designed to execute short threads well. The goal of this research is to significantly reduce the startup cost for a new thread (or thread new to a core). This in turn reduces the break-even point that determines whether a piece of code is parallelizable or not.The term "thread migration" is used to indicate a large number of parallel execution operations, all of which involve moving stored or cached state from one core to another. These operations include forked threads, migrated/moved threads for thermal management or load balancing, loop-parallel threads, task-level parallelism, helper threading, transactional execution, and speculative multithreading - all of these operations will be accelerated to some degree by this research. This research will attack all sources of the thread startup cost, including software (e.g., operating system) overheads, branch predictor state, cached data and instruction state, the commit latency, and the overhead of transferring the primary thread state between cores. In addition to reducing the parallel programming complexity, the broader imapcts of this research include graduate and undergraduate student training, availability of an open-souce simulation infrastructure.