My colleague David Callahan leads a team within Visual Studio that is working on programming models for concurrency. He has pointed out that fundamental concurrency requirements and techniques fall into three basic categories, or pillars. They are summarized in Table 1 . Understanding these pillars gives us a framework for reasoning clearly about all aspects of concurrency from the concurrency requirements and tradeoffs that matter to our current project, to why specific design patterns and implementation techniques are applicable to getting specific results and how they are liable to interact, and even to evaluating how future tools and technologies will fit with our needs.
Let's consider an overview of each pillar in turn, note why techniques in different pillars compose well, and see how this framework helps to clarify our vocabulary.
Pillar 1: Responsiveness and Isolation Via Asynchronous Agents
Pillar 1 is all about running separate tasks, or agents, independently and letting them communicate via asynchronous messages. We particularly want to avoid blocking, especially on user-facing and other time-sensitive threads, by running expensive work asynchronously. Also, isolating separable tasks makes them easier to test separately and then deploy into various parallel contexts with confidence. Here we use key terms like "interactive" and "responsive" and "background"; "message" and "dialogue"; and "timeout" and "cancel."
A typical Pillar 1 technique is to move expensive work off an interactive application's main GUI pump thread. We never want to freeze our display for seconds or longer; users should still be able to keep clicking away and interacting with a responsive GUI while the hard work churns away in the background. It's okay for users to experience a change in the application while the work is being performed (for example, some buttons or menu items might be disabled, or an animated icon or progress bar might indicate status of the background work), but they should never experience a "white screen of death"a GUI thread that stops responding to basic messages like "repaint" for a while because the new messages pile up behind one that's taking a long time to process synchronously. Typically, users keep trying to click on things anyway to wake the application up, and finally either give up and kill the application because they think it crashed (possibly losing or corrupting work, even though the program would eventually have started responding again!) or else wait it out only to experience a flood where all the clicks they tried to perform finally get processed, usually with unsatisfactory or wrong results. Don't let this happen to your application, even if big companies let it happen to theirs.
So what kind of work do we want to ship out of responsiveness-sensitive threads? It can be work that performs an expensive or high-latency computation (background compilation or print rendering, for instance) or actual blocking (idle waiting for a lock, a database result, or a web service reply). Some of these tasks merely want to return a value; others will interact more to provide intermediate results or accept additional input as they make progress on their work.
Finally, how should the independent tasks communicate? A key is to have the communication itself be asynchronous, preferably using asynchronous messages where possible because messages are nearly always preferable to sharing objects in memory (which is Pillar 3's territory). In the case of a GUI thread, this is an easy fit because GUIs already use message-based event-driven models.
Today, we typically express Pillar 1 by running the background work on its own thread or as a work item on a thread pool; the foreground task that wants to stay responsive is typically long-running and is usually a thread; and communication happens through message queues and message-like abstractions like futures (Java
IAsyncResult). In coming years, we'll get new tools and abstractions in this pillar, where potential candidates include active objects/services (objects that conceptually run on their own thread, and calling a method is an asynchronous message); channels of communication between two or more tasks; and contracts that let us explicitly express, enforce, and validate the expected order of messages.
This pillar is not about keeping hundreds of cores busy; that job belongs to Pillar 2. Pillar 1 is all about responsiveness, asynchrony, and independence; but it may keep some number of cores busy purely as a side effect, because it still expresses work that can be done independently, and therefore, in parallel.