Your 15-418/15-618 final project gives you the opportunity to dive deeply into a parallel systems problem of your choosing for the final month of the course. Perhaps more importantly to some of you, it is your big chance to achieve fame, glory, and prizes at the parallelism competition. What you attempt for your project is completely up to you. There are only two requirements: (1) We want your project to be challenging (you should learn something relevant to the themes of this class) and (2) we want your project to be fun (you should be pumped to work on it)!

Choosing a Project

One common way to choose a project is to design a parallel solution to a problem in an application area that is interesting to you. Projects you have attempted in other classes are a good source of ideas. For example, projects in machine learning, AI, graphics, computational photography, and computer vision often stand to benefit greatly from parallelization. If you can convince the course staff that a parallel programming problem in one of these application domains is sufficiently challenging (that is, the solution to get good speedup is not obvious to you from the start), it's likely it will make a good project.

Other project ideas focus on system design or workload evaluation. For example, a project might compare the performance of CPU and GPU implementations of a parallel algorithm, and describe which platform is better suited for the job. Alternatively you could choose to evaluate different versions of an algorithm for different architectures. You could simulate the behavior of code on machines with different SIMD widths, add a feature to the ISPC compiler (its implementation is open source), or develop a parallel debugging tool that helps visualize bottlenecks and performance in parallel programs.

You may implement your project on any parallel platform. The machines in the GHC labs (4 and 6-core machines), GPUs, Blacklight, Amazon EC2, iPhone/iPad/Android SoCs (we can probably get you an NVIDIA Tegra K1 dev kit if you need one), FPGAs, Raspberry Pi, and architecture simulators are all possible and welcome platforms for projects. We also can give you access to a 32-core machine with 256 GB of RAM. You should come talk to us if you have exotic computing needs.

Below is a random sampling of ideas: (You should also look at the list of previous year's projects: Spring 2012 and Spring 2013)

• Application-oriented projects (parallelize an application):
  • Implement a game playing system: Chess, Go, etc.
  • Graphics: extend assignment 2 to achieve high performance under real workloads (render real triangle meshes and more complex shading functions); implement a parallel ray tracer. (those interested in graphics projects should see Kayvon)
  • Physical simulation: implement a high-resolution fluid simulation, rigid body solver, cloth simulation
  • Computer vision: real-time object detection/tracking, image similarity search in a large image database
  • Machine learning: scale up a recent learning algorithm/pipeline, for example - clustering, large scale classification, deep learning etc. (talk to Harry if you are interested!)
  • Machine learning: you may have heard of Google's famous 'cat detector'. This was a ML project designed at building features for tasks such as classification of images, using 16,000 CPU cores. More recently, Coates et al. 2013 showed that similar results can be obtained using commercial off-the-shelf GPU's. Specificially, they managed to get competitive results with google, whilest using just 3 machines. One problem with their result is that it required a lot of complex programming and hand-tuning of parameters. Instead, we will use machine learning to optimize the execution of machine learning algorithms on GPUs without requiring users to be familiar with the machine's hardware; you could say that we are 'fitting' the algorithm to the architecture. NOTE: If you would like to do this project please make sure you have the time to commit to it. You should also be able to write clean C++ code that can be read by others. Please contact Anders Oland (lead on this project) at anderso@cs and Harry.
  • Image processing (for those that have taken Efros' computational photography class)
  • Implement a parallel linear solver (using the conjugate-gradient or multi-grid method)
  • Machine learning algorithms on big data
  • Take a look at Guy Blelloch's problem-based parallel algorithm benchmark suite. Any solution that improves on the algorithms in the performance suite would be a great project.
  • Implement an application on an FPGA (see CoRAM)
  • Compare the performance of different parallel algorithms for the same task on different machines (often different algorithms are best for different platforms)
• Languages/runtimes/compilers:
  • Add high-performance parallel bindings to a system like Javascript or Python. e.g., See PyCuda
  • Annotate the compiled ISPC code with calls to CPU performance monitor instructions and then gather interesting statistics about program execution: cache hits/misses, IPC (and scalar IPC and vector IPC separately). Create a visualization tool for the results. (from Matt Pharr)
  • For the really brave: Add polymorphic functions to ISPC: implement a function template mechanism in the compiler since it gets to be painful to write multiple versions of functions with both uniform and varying parameter types. (from Matt Pharr).
  • Add a GPU backend for a subset of GraphLab, or simply improve the existing parallel runtime for clusters or CPUs.
  • Extend your 411 compiler to generate parallel code.
  • Design a mini domain-specific language (or API, or framework) for a problem domain you are interested in.
• Systems and analysis projects:
  • Study a workload's amenability to SIMD execution. Simulate behavior given multiple SIMD widths.
  • Modify or analyze a workload using GPGPU Sim, PIN, or talk to Yixin about other simulation tools.
  • Modify your 410 kernel to utilize a parallel machine
  • Investigate parallel implementations of malloc
  • Measure the energy consumption of a parallel computer under various loads: (one example is here)
  • Build a real elastic web server using Amazon's actual services.
  • Measure/analyze the energy consumption on a mobile device while certain applications are running.
  • High performance parallel in memory cache.
  • Build a heterogeneous computing system to accelerate the performance of a range of applications using FPGAs.