CSR: Small: Running the Kernel Continuously with Simultaneous Multi-Threading

自2002年以来，许多商用计算机处理器都包含一种称为同步多线程（SMT）的功能，由英特尔以Hyperthreading（TM）的名称销售。SMT使单个处理器核心能够并行执行两个或多个任务。到目前为止，操作系统研究人员一直专注于在现有的操作系统设计中适应SMT，研究诸如SMT感知的资源分配、调度和同步方法等主题。在这个项目中，而不是仅仅适应SMT，我们调查了一个新的低层次的操作系统设计，是由SMT硬件。在现代操作系统中，处理器在用于运行应用程序的安全但相当有限的“用户模式”和仅对操作系统可用的不安全但全能的“管理员模式”之间切换。有时，这些开关每秒发生数百万次，提供并行操作的错觉。在我们的新设计中，用户和管理模式实际上是同时活动的，但在单独的SMT线程上。这避免了频繁模式切换的高昂成本，并在整个操作系统中实现了新的效率和设计改进。基于初步的测量，我们发现，传统的SMT提供的吞吐量增益迅速减少与系统中的处理器核心的数量，作为在应用程序中的次线性可扩展性的结果。为了解决这个问题，并更好地利用现有的硬件，我们提出了“cokernel”操作系统的设计原则，其中每个CPU的一个硬件线程专用于连续执行内核。通过从其他硬件线程卸载内核任务，协同内核操作系统可以实现更高的每线程应用程序吞吐量，避免可伸缩性问题。此外，在每个内核上具有连续执行的内核线程使得能够对内核的其他方面进行广泛的改进，例如用消息传递替换大多数系统调用、混合协作-抢占进程调度、内核辅助的异步内核间和套接字间通信以及真正的内核后台任务。

Since 2002, many commodity computer processors include a feature called simultaneous multi-threading (SMT), marketed by Intel as Hyperthreading (tm). SMT enables a single processor core to perform two or more tasks in parallel. Until now, operating systems researchers have focused on accommodating SMT in existing operating system designs, investigating topics such as SMT-aware resource allocation, scheduling and synchronization methods. In this project, rather than merely accommodate SMT, we investigate a new low-level operating system design that is enabled by SMT hardware. In modern operating systems, the processor switches between the safe but quite limited 'user mode' which is used to run applications, and the unsafe but all-powerful 'supervisor mode' which is available only to the operating system. Sometimes, these switches occur millions of times per second, providing an illusion of parallel operation. In our new design, the user and supervisor modes are actually active simultaneously, but on separate SMT threads. This avoids the often high cost of frequent mode switches, and enables new efficiency and design improvements throughout the operating system. Based on preliminary measurements, we find that the throughput gains provided by conventional SMT diminishes rapidly with the number of processor cores in a system, as a result of sub-linear scalability in the application. To counter this, and to make better use of existing hardware, we propose the "cokernel" operating system design principle, where one hardware thread per CPU is dedicated to continuously executing the kernel. By offloading kernel duties from the other hardware thread(s), a cokernel operating system enables higher per-thread application throughput, avoiding scalability concerns. In addition, having a continuously executing kernel thread on each core enables a wide range of improvements to other aspects of the kernel, such as replacing most system calls with message passing, hybrid cooperative-preemptive process scheduling, kernel-assisted asynchronous inter-core and inter-socket communication and true kernel background tasks.