On Tuesday at the O’Reilly Open Source Convention (OSCON), Intel announced that the latest version of its Threading Building Blocks (TBB) library, version 2.1, will be available for download this week. TBB is Intel’s contribution to the growing collection of tools designed to help the millions of programmers worldwide exploit parallelism in their applications.
Intel doesn’t make its money making chips — it makes its money making chips that do things people want done. Ultimately this means creating hardware that programmers can use to make their software applications tumble, twirl, and pirouette across the computer desktops of users from Monterey to Minsk. This means that when Intel makes a big architectural change, like the one we are currently experiencing with the introduction of multicore processors, it is in the company’s best interest to help developers manage the change effectively.
A traditional threaded approach to programming that uses, for example, POSIX or Windows threads requires the programmer to manually create and manage the threads used by an application. Managing threads explicitly can be a laborious and error-prone process that challenges even experienced programmers. For those new to parallelism
That’s where Threading Building Blocks, or TBB, comes in. TBB is a C++ template library designed to help programmers express the parallelism hiding in their applications in a way that separates expression from implementation, and without having to worry about thread management details. TBB allows the programmer to focus on managing tasks in the application. The library’s run time engine then handles the scheduling of tasks to threads so that the programmer doesn’t have to, using a work stealing approach that ensures the computational demand on each core stays balanced even in the face of dynamic system load.
Where does it fit in the box of tools available to C++ programmers? TBB is discussed most frequently in the context of POSIX or other native threads implementation and OpenMP.
Compared to parallelism using native threads, TBB offers the advantages already discussed in avoiding the need for the programmer to explicitly manage the threads. For the most part, native threads require that work be assigned to threads through functions and, for users programming in C++ (but not C or Fortran), using native threads may pose challenges that disrupt the flow of the program. TBB is also well-suited to applications with a lot of parallel work that needs to be done that isn’t expressed in a loop structure, and it emphasizes scalability through data parallel programming.
Unlike OpenMP, neither native threads nor TBB requires compiler support. The use of C++ templates means that TBB is available anywhere there is a C++ compiler, and it fits in nicely with object-oriented programming. For applications written in C or Fortran, OpenMP may still be a great solution because it fits well into a structured programming style. (See the examples in this paper for an illustration of how disruptive TBB can be if it doesn’t fit the programming style.) This is especially true if the application is dominated by loop-based processing, where OpenMP really shines. Both OpenMP and TBB are portable, and both adapt easily to spread work to be done over available resources. Native threads share neither of these advantages.
Interestingly all three methods can co-exist in an application, as can MPI, making it straightforward to adopt the model that’s best suited to the work in a particular area of the code — provided you have the necessary programming skills.
There are two versions of TBB: a commercial version, available in shrink wrap from Intel, and the open source version, announced at OSCON last year. I spoke with Intel’s Phil De La Zerda and Vasanth Tovinkere about what’s new in this latest version of the library, and both men emphasized the degree to which the new release benefits from the input of a large and active community on the open source side. Phil is Intel’s director of business development, developer products division and Vasanth is a senior staff engineer in Intel’s performance, analysis and threading lab.
Intel created TBB to provide a robust path to parallel programming for those without a lot of experience in traditional high-end HPC. And Intel has been pleased with its success so far. According to De La Zerda, TBB is included in about 60 percent of Linux distributions that have been paid for, and 80 percent of free distributions. These distributions cover OpenSolaris, Fedora, OpenSuse, Ubuntu, and many others. While De La Zerda declined to be specific on Intel’s download goals, he did mention that in the year since TBB was open sourced, downloads were “in the many thousands,” and the actual numbers more than doubled their original expectations.
With the newly announced version 2.1, TBB is more available than ever before, adding support for Solaris on Intel processors, Apple’s Mac OS X on PowerPC (support for OS X on Intel was already there), and the Xbox 360 from Microsoft. The port to the Xbox was especially important for Ukranian console and PC game developer Deep Shadows, which announced on the 16th of this month that their in-house game engine and games would fully support the use of multicore processors, including the triple-core processor of the popular game console.
Driven by its active user community, Intel has included a number of significant changes in TBB version 2.1. The company talks about them in three logical groupings: “do more,” “do it faster,” and “do it easier.” Not exactly “veni, vidi, vici” memorable, but good enough to organize the rest of this article.
Under the “do more” heading, version 2.1 supports blocking tasks through tbb_thread and the task scheduler. This is important for GUI and I/O events, and allows this type of processing to be handled without interfering with computational work on other cores. TBB also is able to “do it faster” with an affinity partitioner that gets better cache use from chained parallel operations.
Programmers can “do it easier” with new TBB capabilities, such as task cancellation. This is actually a nifty feature that allows one to abort outstanding tasks if they are made unnecessary by other work. An example is a binary search implementation that cancels outstanding tasks once the target node in the tree is found.
Where is TBB going over the next 12 to 18 months? De La Zerda underscores Intel’s continued emphasis on listening to users and engaging the open source community and OS vendors. He says the company is committed to keep on adapting TBB to improve relevance and adoption. Are we likely to see a continued emphasis on OSCON as a key announcement opportunity for TBB? De La Zerda wouldn’t commit, but he did say that he “wouldn’t be surprised” to find himself talking once again about major improvements in the library this time next year.