Parallel programming has advanced many of today's scientific computing projects to a new level. However, writing a program that utilizes parallel and heterogeneous computing resources is not easy because you often need to deal with a lot of technical details, such as load balancing, programming complexity, and concurrency control. Taskflow streamlines this process and helps you quickly create high-performance computing (HPC) applications with programming productivity.
As the user community of Taskflow continues to grow (e.g., over 1.5M downloads), we have identified two specific projects to enhance Taskflow's infrastructure to facilitate its research and development:
Timeline of GSoC is available here.
The project aims to enhance the benchmarking environment of Taskflow by adding more realistic and reproducible testcases to the repository.
The current benchmarking environment of Taskflow has been narrow on electronic design automation (EDA) problems. Most of these benchmarks are embedded in applications themselves and are difficult to reproduce. Moreover, we need benchmarks from other applications, such as streaming, video processing, and high-performance computing (HPC) to make the broader computing community study the performance of Taskflow. We have identified three specific tasks to enhance the benchmarking of Taskflow:
- Implement the standard PaRSEC benchmarks using Taskflow
- Implement a standalone benchmark environment for our EDA applications (OpenTimer, DREAMPlace, etc.)
- Implement baseline methods using mainstream programming systems (OpenMP, TBB, StarPU, etc.)
We will place the result in a new repository, namely Benchmarks,
under the Taskflow Organization.
The repository will contain comprehensive instructions for reproducing the results.
We will also encourage participants to publis these results in related parallel computing conferences
or arXiv journals.
Participants should have decent C++14/17 programming experience. Basic knowledge about parallelism is preferred.
The Taskflow Team will mentor this project throughout the course of summer code. Participants should expect weekly project meeting to sync up the progress.
We expect 350 hours for this project.
We rate this project an medium level of difficulty. This project primarily focuses on using Taskflow to implement parallel algorithms and applications, rather than developing the Taskflow core functionalities. Participants in this project will gain much practical and hands-on experience of parallel programming and underatand the pros and cons of mainstream parallel programming tools.
The project aims to enhance the the pipeline programming infrastructure of Taskflow by adding a data abstraction layer with high scheduling performance.
Taskflow has introduced a new task-parallel pipeline programming framework in v3.3. The current pipeline design is primitive and does not provide any data abstraction. For many data-centric pipeline applications, this can be inconvenient as users need to repetitively create data arrays. Therefore, the goal of this project is to derive a pipeline class with data abstraction to streamline the implementation of data-centric pipeline applications. We have identified three specific tasks:
- Implement a new pipeline class with a data abstraction layer
- Implement unit tests
- Implement algorithms to avoid false sharing of data
The implementation will be integrated into the core of Taskflow as a new algorithm module,
together with its documentation in our handbook pages.
We will also encourage participants to publis these results in related parallel computing conferences
or arXiv journals.
Participants should have decent C++14/17 programming experience. Basic knowledge about pipeline parallelism is preferred.
The Taskflow Team will mentor this project throughout the course of summer code. Participants should expect weekly project meeting to sync up the progress.
We expect 350 hours for this project.
We rate this project a difficult level of difficulty, since it involves both implementation and algorithm challenges. However, participants in this project will not just learn how to implement a real module atop Taskflow but also gain practical research knowledge about parallel scheduling algorithm.