Overview / Usage

The goal of the project was to become familiar with Intel oneAPI, specifically DPC++ and oneTBB, and then evaulate their suitability for very fine-grained parallelism. An example of a very-fined grained workload is recursive Fibonacci, which creates a task for each step of the computation that may consist of only a few instructions. Some parallel programming libraries struggle performance-wise when dealing with tasks of this size, and we wanted to determine if oneTBB fit into this category or not.

We concluded that DPC++ and oneTBB are both well-suited for very-fined grained workloads, and specifically that oneTBB outperformed similar parallel programming libraries in many of our tests. We have written a technical report detailing the results of our testings and our findings.

Methodology / Approach

For our testing, we created a series of benchmarking programs. We used our benchmarks to measure the maximum theoretical throughput in tasks per second, using noop tasks; measured the load balancing capabilities of the scheduler by keeping track of how many tasks each logical thread was processing during execution; and compared the performance of oneTBB to other parallel programming libraries. Specifically, we looked at the GNU and LLVM implementations of OpenMP and XQueue, which is an OpenMP implementation that uses a completely lock-free scheduler. We converted several OpenMP benchmarking suites to use oneTBB.

We tried investigating oneTBB's source code to learn how it is able to achieve its level of performance. Although we had some struggles doing so, after the report was written, we were able to meet with Intel employees to get more details about oneTBB's implementation which explained its performance, and learned why our changes did not have the expected effect.

Technologies Used

As mentioned previously, we used DPC++ and oneTBB. We also consulted Intel vTune during our investigation into oneTBB's source code. We ran all of our benchmarks on a 192-core, 384-thread, Xeon-powered supercomputing node at our university, which allowed us to run our tests at very high levels of parallelism.

Documents and Presentations

Other links

• Source code for the modified GAP benchmark suite. (https://github.com/owm111/gapbs)

• Source code for miscellaneous benchmarks. (https://github.com/owm111/oneapi-misc)

• Attempted modifications to oneTBB's source code. (https://github.com/owm111/oneTBB/tree/locks-testing)

• Source code for the modified BOTS benchmark suite. (https://github.com/owm111/bots)