Multi-Core Software

The Foundations for Scalable Multi-Core Software in Intel® Threading Building Blocks

CONCLUSION

Intel® Threading Building Blocks is a C++ template library designed to raise the level of abstraction for parallelism as developers port their code to multi-core platforms. Starting with the 2.0 version, Intel TBB is also provided at www.threadingbuildingblocks.org as an open-source project licensed under the GNU Public License.

Two key features of the library are its work-stealing task scheduler and scalable memory allocator. Both of these systems reduce the need of users to understand the many complex issues related to multi-core performance and scalability.

In the TBB Task Scheduler section, we provided an overview of the task scheduler design and outlined several manual optimizations that users can perform to improve the performance of the scheduler when executing fine-grain tasks.

In the Scalable Memory Allocation section, we described the motivation for and implementation of the scalable memory allocator, highlighting the design characteristics that decrease synchronization, increase locality, and avoid false sharing.

In the Experimental Results section, we explored the performance of a number of benchmarks on a server with two Intel® Xeon® processors. We showed that the overhead of work stealing is low for large-grain tasks, and that the manual optimizations described in this paper offer a small but noticeable improvement when scheduling fine-grain tasks.

In our evaluation of scalable memory allocators, the TBB scalable allocator was shown to be competitive with several commercial and research allocators.

In an analysis of an example that studied the combined effects of the scheduling optimizations and the scalable allocator, the use of the scalable allocator showed a large impact for both small- and large-grain tasks. The scheduling optimizations were shown to have a small performance impact for the small-grain tasks and a negligible impact on the scheduling of the larger-grain tasks. This confirms the assertion that memory allocation can sometimes be a limiting factor in the scalability of parallel applications and that a scalable allocator can remove this bottleneck.

With the growing availability of multi-core platforms, it is becoming imperative for performance-oriented developers to thread their code. Intel TBB, built on its work-stealing task scheduler and scalable memory allocator, offers an exciting solution to ease the burden of this transition.

Back to Top