Skip to content

Google Summer of Code (GSoC) 2022

Giannis Gonidelis edited this page Apr 11, 2022 · 36 revisions

Introduction

Welcome to the Google Summer of Code (GSoC) page for the HPX project. Here you can find information about student projects, proposal submission templates, advice on writing good proposals, and links to information on getting started with HPX. The STE||AR Group will apply as an organization and our goal is to have at least five funded students working on HPX related projects.

Requirements

Students must submit a proposal. A template for the proposal is available here. Find hints for writing a good proposal here.

We strongly suggest that students interested in developing a proposal for HPX discuss their ideas on the IRC channel (Libera) or the mailing list to help refine the requirements and goals. Please see [this page] for information on how to access IRC or the mailing list this page. Students who actively plan and discuss projects with developers are generally ranked before those that do not.

We have intentionally left the descriptions of these projects vague and open to interpretation because we expect students to develop their proposals' requirements by doing initial background research on the topic and interacting with the community. In addition, it is important to note that the suggested projects on this page are not binding -- if you have an interest in parallel task-based programming and have an idea for a project that would either improve HPX or demonstrate how well it applies to your problem, then feel free to suggest your idea as a project and write a proposal for it. We will be glad to help you with project goals to improve your proposal if you have ideas, so do not leave them until the last minute.

We will expect students to demonstrate that they have the required level of C++ and CMake knowledge by showing us some of their previous work (e.g., a GitHub repository), or preferably, by them making a small demonstration program using HPX that shows a simple example of something they have created themselves.

Potential Additional Funding

For students who perform at or above expectations on both GSoC evaluations, the Center of Computation and Technology (CCT) at Louisiana State University (LSU) may fund up to an additional four weeks' work on the project for no more than the GSoC rate of pay. This funding is not guaranteed and is independent of the GSoC program. Students accepted for additional funding will be paid through LSU for the additional weeks and affiliated with LSU during that time. Additional paperwork through LSU will be required.

Tips for Prospective Students

Some of our former GSoC students that still contribute to our projects have put together the following list. All of them had to go through the same learning experience. Prospective students most probably face this challenge now, so the list provides pointwise help to get into HPX smoothly.

  1. The first thing we suggest is to build HPX from the source using the CMake build system. An example guide to build HPX is here. Various ways of building HPX (e.g., memory allocators, OTF2 traces, CUDA support) will enable you to understand the capabilities of HPX as a runtime.
  2. Once you're acquainted with the build system, we suggest you read our docs/wiki and try to familiarize yourself with the basic terminology (e.g., locality, LCO, futurization, etc.).
  3. Next, we suggest you watch talks on HPX on YouTube. Doing so should give you a brief overview of the motivations and implementation design of the components within HPX.
  4. At this point, try building and playing with the examples in HPX. Furthermore, we have a basic tutorial that takes you through the features and their usage with code examples.
  5. Going through the examples may be an overwhelming experience, so we suggest you become familiar with our way of writing code through our summer lecture series. (Hint: Pay attention to Lecture #4)
  6. When you're familiar with basic usage, we suggest you try writing demo HPX programs (e.g., matrix-matrix multiplication). Go through our Issue tracker and see if you can find an issue you would like to investigate. Working on bugs is the easiest way to dive into the code base and contribute to HPX.
  7. Dig into our currently active GSoC issues and Pull Requests relevant to them. Furthermore, leave comments and discuss with the corresponding authors.
  8. We highly recommended joining our IRC channel, #ste||ar, on Libera, where you can ask questions, discuss issues, pull requests, and your potential GSoC project. Remember, questions are the key to start contributing!

2022 HPX Project Ideas

There are new projects this year, and also ones revamped from previous years (legacy) that are still of interest. These projects have mentors ready and waiting to help students.

Core HPX Projects

These are projects that involve making changes/improvements/extensions to the core HPX library.

Port HPX to iOS and Mac (M1 architecture)

  • Abstract: HPX has already proven to run efficiently on ARM-based systems. This has been demonstrated with an application written for Android tablet devices. A port to handheld devices running with iOS would be the next logical step! On top of that since the new Apple M1 ARM-based processors have proven to be very efficient the student should consider providing an HPX version for this architecture as well. To run HPX efficiently on there, we need to adapt our build system to be able to cross-compile for iOS and Mac and add a code to interface with the iOS GUI and other system services. A preexisting Mac support infrastructure exists but the student will need to adapt and update it to current releases.
  • Difficulty: Easy/Medium
  • Expected result: Provide a prototype HPX application running on an iPhone or iPad.
  • Knowledge Prerequisite: C++, Objective-C, iOS
  • Mentor: Hartmut Kaiser (hartmut%20kaiser) and Thomas Heller (thomas%20heller)
  • Project Size: 350 hour (large project)

Bring the HPX distributed algorithms up to date

  • Abstract: Along with the standard parallel algorithms provided by the C++ standard, HPX extends its infrastructure by providing (some of) the corresponding distributed versions of those algorithms that run on multiple nodes and on top of that they take care of communication. The set of the implemented algorithms can be found here. Due to lack of maintainance these algorithms do not compile properly to some systems according to our latest continuous integration tests and they are considered the last missing piece for HPX to be a fully integrated and portable library. Part of this project is to investigate the reasons that the tests of these algorithms fail (either on the algorithm source code side or on the integration tests side) and "repair" them. Providing further implementations of the remaining algorithms could facilitate as an extension of the project for the prospective students that are interested in a long-term project. Here you can find the corresponding ticket.
  • Difficulty: Easy
  • Expected result: Repair the segmented algorithms so continuous integration does not fail on them and implement the remaining segmented algorithms.
  • Knowledge Prerequisite: C++, CMake, CircleCI, Jenkins
  • Mentor: Hartmut Kaiser (hartmut%20kaiser) and Giannis Gonidelis (gonidelis [at] hotmail.com)
  • Project Size: 175 hour (medium sized)

Pythonize HPX!

  • Abstract: Python is a widely adopted language due to its simplicity. Providing performant HPX functionality written in C++ with Pythonic APIs would facilitate both usage and ease of learning HPX. The student shall design and implement Python bindings for HPX, exposing all or parts of the HPX functionality with a 'Pythonic' API. This should be possible as Python has a much more dynamic type system than C++. Using Boost.Python and/or Pybind11 seem to be good choices for this.
  • Difficulty: Medium
  • Expected result: Demonstrate functioning bindings by implementing small example scripts for different simple use cases
  • Knowledge Prerequisite: C++, Python
  • Mentor: Hartmut Kaiser (hartmut%20kaiser)
  • Project Size: 350 hour (large project)

Create Generic Histogram Performance Counter

  • Abstract: HPX supports performance counters that return a set of values for each invocation. We have used this to implement performance counters collecting histograms for various characteristics related to parcel coalescing (such as the histogram of the time intervals between parcels). The idea of this project is to create a general-purpose performance counter which collects the value of any other given performance at given time intervals and calculates a histogram for those values. This project could be combined with Add more arithmetic performance counters.
  • Difficulty: Medium
  • Expected result: Implement a functioning performance counter which returns the histogram for any other given performance counter as collected at given time intervals.
  • Knowledge Prerequisite: Minimal knowledge of statistical analysis is required.
  • Mentor: Hartmut Kaiser (hartmut%20kaiser) and Mikael Simberg (mikael%20simberg)
  • See issue #2237 on HPX bug tracker
  • Project Size: 350 hour (large project)

Conflict (Range-Based) Locks

  • Abstract: Some multi-threaded algorithms may require resources that must be held using a lock, but the locking mechanism may be range-based rather than absolute. Consider a large array of N items where a task requires some small subset of the items to be locked while a second task requires a second range. If these tasks are placed into a DAG so that task2 can only run once task1 has been completed, it will be inefficient when the range of items used by task2 does not overlap the range from task1. When many tasks operate on the range, with randomly overlapping or non-overlapping regions, DAG-based task scheduling leads to a highly inefficient strategy. We need a range based lock that can be templated over <items>, and that can then be locked/unlocked on ranges (of those items) and interact with our future<> based scheduling so that items will become ready when the range they need has no locks outstanding, and so that when a task releases a lock, any other tasks that overlap the range are in turn signaled as possibly ready. (For an example of how this is used in conventional HPC programming, look up Byte Range locks in MPI for Parallel IO to a single file). A successful implementation can be extended to multi-dimensional locking *2D/3D etc., ideally templated over dimensions and types).
  • Difficulty: Medium/Hard
  • Expected result: A test application that creates arrays of items and randomly assigns tasks to operate on regions of those items with locking and schedules the tasks to operate in a non-conflicting way.
  • Knowledge Prerequisite: Thread-safe programming. Futures.
  • Mentor: John Biddiscombe (john%20biddiscombe)
  • Project Size: 350 hour (large project)

(Re-)Implement executor API on top of sender/receiver infrastructure

  • Abstract: P0443 will most likely be accepted for C++23. Our executor API (customization points) currently dispatch to an executor interface defined by wg21.link/p0443R3. All HPX facilities related to scheduling tasks (algorithms, future, dataflow, async, etc.) rely on the executor customization points to perform their operations. Although major steps have been taken for the integration of the executors proposal to HPX there is still many facilities that need to be implemented. The project can be correlated with the Coroutine-like interface project project and the P2300 proposed awaitables.
  • Difficulty: Medium
  • Expected result: The result should be functioning executor customization points built upon senders/receivers.
  • Knowledge Prerequisite: Parallel algorithms.
  • Mentor: Hartmut Kaiser (hartmut%20kaiser) and Giannis Gonidelis (gonidelis [at] hotmail.com)
  • See issue #5219 on HPX bug tracker and the correspodning Pull Request that's on the works already.
  • Project Size: 350 hour (large project)

Coroutine-like Interface

  • Abstract: HPX is an excellent runtime system for doing task-based parallelism. In its current form, however, the results of tasks can only be expressed in terms of returning from a function. However, there are scenarios where this is not sufficient. One example would be lazy ranges of integers (for example, Fibonacci, 0 to n, etc.). For those, a generator/yield construct would be perfect (more on coroutines)! Addiotionaly, an option would be to rely on top of the senders/receivers proposed facilities, a completely new interface for execution in standad C++ that may (or not) revolutionize the way we implement concurrency.
  • Difficulty: Medium
  • Expected result: Implement yield and demonstrate on at least one example
  • Knowledge Prerequisite: C++
  • Mentor: Hartmut Kaiser (hartmut%20kaiser) and Thomas Heller (thomas%20heller)
  • Project Size: 350 hour (large project)

Update Build System to automatically fetch HPX dependencies when not available

  • Abstract: HPX currently depends on three libraries. Although our build system supports a CMake flag to fetch and use a default Asio version, there are no analogous options for Boost and HWLOC. Adding these options would make building HPX easier and would attract more prospective users. The idea is to fetch and install these dependencies with HPX if the user does not have them installed. Further improvements and automation on the build system may be proposed. Updates to the HPX documentation wrt build changes are expected. Feel free to check on the related issues #3440, #5728.
  • Difficulty: Medium
  • Expected result: Provide a renewed user-friendly build system environment.
  • Knowledge Prerequisite: CMake
  • Mentor: Nikunj Gupta (nikunj [at] illinois.edu), Hartmut Kaiser (hartmut%20kaiser), and Giannis Gonidelis (gonidelis [at] hotmail.com)
  • Project Size: 175 hour (medium sized)

Add Vectorization to par_unseq Implementations of Parallel Algorithms

  • Abstract: Our parallel algorithms currently don't support the par_unseq execution policy. This project is centered around the idea to implement this execution policy for at least some of the existing algorithms (such as for_each and similar).
  • Difficulty: Medium/Hard
  • Expected result: The result should be functioning parallel algorithms when used with the par_unseq execution policy. The loop body should end up being vectorized.
  • Knowledge Prerequisite: Vectorization, parallel algorithms.
  • Mentor: Hartmut Kaiser (hartmut%20kaiser), Srinivas Yadav (vasu.srinivasvasu.14 [at] gmail.com), Nikunj Gupta (nikunj [at] illinois.edu), Giannis Gonidelis (gonidelis [at] hotmail.com)
  • See issue #2271 on HPX bug tracker
  • Project Size: 350 hour (large project)

New Views and Interactivity for HPX OTF2 Trace Visualization

  • Abstract: HPX traces are collected with APEX and written in as OTF2 files with extensions. These trace files are typically visualized using a Gantt chart or collection of timelines. The present implementation includes a variety of related views to further help summarize or navigate the trace. We would like to add to these capabilities with more views and further kinds of linked highlighting between them. In particular, we would like to add views that summarize area selections in the Gantt chart and extend the kinds of selections (e.g., multi-select, area select, code linking) available. This functionality will require data processing in a Python backend with visualization/interactivity in a Javascript front end. See the Traveler project here: https://github.com/hdc-arizona/traveler-integrated
  • Difficulty: Medium
  • Expected result: A new view will be available through Traveler to show selection summaries. More types of selection and linking will be available in Traveler.
  • Knowledge Prerequisite: Javascript, Python.
  • Mentor: Kate Isaacs (kate%20isaacs)
  • Project Size: 175 hour (medium sized)

Multiple Dataset Performance Visualization

  • Abstract: Traveler-Integrated is a platform for visualization of parallel runtimes performance, such as HPX. The interface allows accessing multiple datasets from different executions. However, several improvements can be made to the interface in managing large numbers of files (e.g., from historical regression runs) and organizing windows to handle the comparison of runs. In this project, you will work on the Javascript front end and Python backend to implement this interface, refining the design through user feedback.
  • Difficulty: Easy-Medium
  • Expected result: Traveler-Integrated will have a newly designed interface for managing data of multiple runs.
  • Knowledge Prerequisite: Python, Javascript.
  • Mentor: Kate Isaacs (kate%20isaacs)
  • Project Size: 350 hour (large project)

Large File Support for HPX OTF2 Trace Visualization

  • Abstract: HPX traces are collected with APEX and written in as OTF2 files with extensions. These trace files are typically visualized using a Gantt chart or collection of timelines. The present implementation reads the entirety of the trace file before generating the visualization. However, the OTF2 interface has support for partial reading of the file and a parallel backend. This project would modify the Gantt chart backend (C++) to utilize these features, thus supporting larger files. The project could also modify the front end to use WebGL (Javascript) when the number of data items is large.
  • Difficulty: Medium-Hard
  • Expected result: Files that require more memory than on a single machine can be run from that machine. The time from program-start to visualization is decreased due to the use of large file features.
  • Knowledge Prerequisite: C++, Javascript.
  • Mentor: Kate Isaacs (kate%20isaacs)
  • Project Size: 175 hour (medium sized)

HPX User Projects

These are projects that improve code that uses HPX. In general, the primary goal with these projects is to improve user uptake of HPX by demonstrating its use in other projects, and only minor fixes/changes/extensions should be necessary for the main HPX library itself.

Implement your favorite Computational Algorithm in HPX

  • Abstract: This is an open project for prospective students who don't want to get their hands dirty into core HPX development. The student shall utilize HPX to implement a short independent project that will utilize HPX for performance boost. The program can implement any given problem that requires heavy computational effort from the literature. Efficient matrix multiplication, sorting, stencil variations or any AI, Physics related problem would be a good candidate. An extensive list of use-case examples is already available in our source code. The goal of this project is for the student to get acquainted with HPX development and contribute to our vast range of applications.
  • Difficulty: Easy
  • Expected result: Implement a standalone program that utilizes HPX for performance.
  • Knowledge Prerequisite: C++
  • Mentor: Hartmut Kaiser (hartmut%20kaiser) and Giannis Gonidelis (gonidelis [at] hotmail.com)
  • Project Size: 175 hour (medium sized)

Conduct a thorough Performance Analysis on HPX Parallel Algorithms (and optimize)

  • Abstract: HPX implements all the C++ standard algorithms along with their ranges counterparts. Conducting extensive performance analysis on the existing implementations and coming up with possible optimizations would improve the efficiency of our parallel algorithms and boost HPX performance in general. The student shall expect to work both on top of HPX by writing custom benchmarks for weak and strong scaling, evaluate the results and perform source optimizations under the hood (core development).
  • Difficulty: Medium
  • Expected result: Boost the performance of at least one C++ standard algorithm in HPX.
  • Knowledge Prerequisite: C++
  • Mentor: Giannis Gonidelis (gonidelis [at] hotmail.com) and Hartmut Kaiser (hartmut%20kaiser)
  • Project Size: 350 hour (large project)

Legacy Project Ideas

These are project ideas from previous Summer of Code years that we are still interested in working on, but it might be harder to find a mentor willing to supervise a student. Therefore, we would expect only very self-motivated and capable students to select a project from the legacy category. We cannot guarantee that we will select a project from this list unless we are quite satisfied that the student can complete the work.

We are looking to fund work on a number of different kinds of proposals (for more details about concrete project ideas, see below):

  • Extensions to existing library features,
  • New distributed data structures and algorithms
  • Multiple competing proposals for the same project

Implement Your Favorite Parcelport Backend

  • Abstract: The HPX runtime system uses a module called Parcelport to deliver packages over the network. An efficient implementation of this layer is indispensable and we are searching for new backend implementations based on CCI, ucx or libfabric. These mentioned abstractions over various network transport layers offer the ability to do fast, one-sided RDMA transfers. The purpose of this project is to explore one of these and implement a parcelport using it.
  • Difficulty: Medium-Hard
  • Expected result: A proof of concept for a chosen backend implementation with performance results
  • Knowledge Prerequisite: C++, Basic understanding of Network transports
  • Mentor: Thomas Heller (thomas%20heller)
  • Project Size: 350 hour (large project)

Implement a Faster Associative Container for GIDs

  • Abstract: The HPX runtime system uses Active Global Address Space (AGAS) to address global objects. Objects in HPX are identified by a 128-bit unique global identifier, abbreviated as a GID. The performance of HPX relies on fast lookups of GIDs in associative containers. We have experimented with binary search trees (std::map) and hash maps (std::unordered_map). However, we believe that we can implement a search data structure based on n-ary trees, tries, or radix trees that exploit the structure of GIDs such that it allows us to have faster lookup and insertion.
  • Difficulty: Medium-Hard
  • Expected result: Various container approaches to choose from together with realistic benchmarks to show the performance properties
  • Knowledge Prerequisite: C++, Algorithms
  • Mentor: Thomas Heller (thomas%20heller)
  • Project Size: 350 hour (large project)

Create A Parcelport Based on WebSockets

  • Abstract: Create a new parcelport that is based on WebSockets. The WebSockets++ library seems to be a perfect starting point to avoid having to dig into the WebSocket protocol too deeply.
  • Difficulty: Medium-Hard
  • Expected result: A proof of concept parcelport based on WebSockets with benchmark results
  • Knowledge Prerequisite: C++, knowing WebSockets is a plus
  • Mentor: Hartmut Kaiser (hartmut%20kaiser) and Thomas Heller (thomas%20heller)
  • Project Size: 350 hour (large project)

All to All Communications

  • Abstract: Design and implement efficient all-to-all communication LCOs. While MPI provides mechanisms for broadcasting, scattering and gathering with all MPI processes inside a communicator, HPX currently misses this feature. It should be possible to exploit the Active Global Address Space to mimic global all-to-all communications without actually communicating with every participating locality. Different strategies should be implemented and tested. A first and very basic implementation of broadcast already exists which tries to tackle the above-described problem. However, more strategies for granularity control and locality exploitation need to be investigated and implemented. We also have the first version of a gather utility implemented.
  • Difficulty: Medium-Hard
  • Expected result: Implement benchmarks and provide performance results for the implemented algorithms
  • Knowledge Prerequisite: C++
  • Mentor: Thomas Heller (thomas%20heller) and Andreas Schaefer (andreas%20schaefer)
  • Project Size: 175 hour (medium sized)

Distributed Component Placement

  • Abstract: Implement an EDSL to specify the placement policies for components. This could be done similar to [Chapels Domain Maps] (http://chapel.cray.com/tutorials/SC12/SC12-6-DomainMaps.pdf). In Addition, allocators can be built on top of those domain maps to use with C++ standard library containers. This is one of the key features to allow users to efficiently write parallel algorithms without having them worried too much about the initial placement of their distributed objects in the Global Address space
  • Difficulty: Medium-Hard
  • Expected result: Provide at least one policy that automatically creates components in the global address space
  • Knowledge Prerequisite: C++
  • Mentor: Thomas Heller (thomas%20heller) and Hartmut Kaiser (hartmut%20kaiser)
  • Project Size: 350 hour (large project)

Port Graph500 to HPX

  • Abstract: Implement Graph500 using the HPX Runtime System. Graph500 is the benchmark used by the HPC industry to model important factors of many modern parallel analytical workloads. The Graph500 list is a performance list of systems using the benchmark and was designed to augment the Top 500 list. The current Graph500 benchmarks are implemented using OpenMP and MPI. HPX is well suited for the fine-grain and irregular workloads of graph applications. Porting Graph500 to HPX would require replacing the inherent barrier synchronization with asynchronous communications of HPX, producing a new benchmark for the HPC community as well as an addition to the HPX benchmark suite. See http://www.graph500.org/ for information on the present Graph500 implementations.
  • Difficulty: Medium
  • Expected result: New implementation of the Graph500 benchmark.
  • Knowledge Prerequisite: C++
  • Mentor: Patricia Grubel (patricia%20grubel), and Thomas Heller (thomas%20heller)
  • Project Size: 350 hour (large project)

Port Mantevo MiniApps to HPX

  • Abstract: Implement a version of one or more mini-apps from the Mantevo project (http://mantevo.org/ "Mantevo Project Home Page") using HPX Runtime System. We are interested in mini-applications ported to HPX that have irregular workloads. Some of these are under development, and we will have access to them in addition to those listed on the site. On the site, MiniFE and phdMESH would be good addition to include in HPX benchmark suites. Porting the mini-apps would require porting the apps from C to C++ and replacing the inherent barrier synchronization with HPX's asynchronous communication. This project would be a great addition to the HPX benchmark suite and the HPC community.
  • Difficulty: Medium
  • Expected result: New implementation of a Mantevo mini-app or apps.
  • Knowledge Prerequisite: C, C++
  • Mentor: Patricia Grubel (patricia%20grubel) and Thomas Heller (thomas%20heller)
  • Project Size: 175 hour (medium sized)

Distributed solver and load balancing for Peridynamics using asynchronous parallelism

  • Abstract: Peridynamics is a reformulation of classical continuum mechanics (e.g., linear elastodynamics). The internal force at any point in the solid results from the interaction of that point with neighboring points within some distance ϵ. Typically, ϵ is much larger than the mesh size. As a result, the computation is more intensive and introduces more substantial data dependencies when partitioning the domain for parallel implementation. This project aims to develop and implement a distributed solver for Peridynamics in an existing codebase [1]. This project will benefit from the last year's GSoC student's effort on a similar goal but for a simplified nonlocal model [2]. In [2], several challenges associated with the parallelization of nonlocal models are highlighted, and algorithms are developed to address the challenges. In this project, we will apply techniques in [2] to the Peridynamics problem; first, we will implement the distributed solver; second, we will optimize the code so that compute node does the information exchange and calculation on the free degree of freedoms (DoFs) simultaneously to minimize the wait time. Finally, if possible, we will add the load balancing algorithm [2]. Here for the given compute node, free DoFs are those DoFs that do not depend on the data owned by other compute nodes. After GSoC, we intend to write a workshop paper based on this project's efforts and possibly present it at a computer science conference.
  • Difficulty: Medium-Hard
  • Expected result: Extend the existing shared memory code to a distributd code
  • Knowledge Prerequisite: C++
  • Mentor: Patrick Diehl (patrick%20diehl) and Prashant K. Jha (prashantk%jha)
  • Project Size: 350 hour (large project)

[1] https://github.com/nonlocalmodels/NLMech

[2] https://arxiv.org/abs/2102.03819

Bug Hunter

  • Abstract: In addition to our extensive ideas list, several active tickets are listed in our issue tracker which are worth tackling as a separate project. Feel free to talk to us if you find something interesting. A prospective student should pick at least one ticket with medium to hard difficulty and discuss how to resolve it.
  • Difficulty: Medium-Hard
  • Expected result: The selected issues need to be fixed
  • Knowledge Prerequisite: C++
  • Mentor: Thomas Heller (thomas%20heller)
  • Project Size: 175 hour (medium sized)

Project: Template

  • Abstract:
  • Difficulty:
  • Expected result:
  • Knowledge Prerequisite:
  • Mentor:
Clone this wiki locally