Keichi Takahashi

Keichi Takahashi

I am a computer scientist at Nara Institute of Science and Technology (NAIST) working on high-performance computing.

Location Japan

Activity

  • Keichi Takahashi made a comment
    Pros and cons of MPI

    Pros: MPI is a highly optimized library with a long-standing history. It's hard to beat MPI's performance and functionality with your own implementation using sockets, verbs, etc..

    Cons: MPI is too low-level. The fact that MPI gives you full control over what data is communicated between which processes also means that you need to describe every...

  • Keichi Takahashi made a comment
    Hands-on: Non-blocking communication

    I wonder why Isendrecv (a non-blocking Sendrecv) does not exist?

  • Keichi Takahashi made a comment
    Hands-on: Message exchange

    Maybe it's worth noting that the same code does not hang if the message is small enough (< 2KB in my environment) since the message is sent with the eager protocol.

  • Keichi Takahashi made a comment
    Different ways to use compiled code with Python

    Cython is a good choice if you have an existing Python code base that you need to make "moderately" faster. However, if you want to squeeze out every last bit of performance, using CFFI or f2py makes more sense.

    To maximize the performance using Cython, you need to statically type variables, disable bounds check, etc. In the end, you lose all the nice...

  • Keichi Takahashi made a comment
    Hands-on: utilizing Fortran code

    Cython: 0.032s
    CFFI: 0.015s
    f2py: 0.019s

    Even though f2py resulted in slightly lower performance than CFFI, it's probably the easiest among them requiring minimal code change.

  • Keichi Takahashi made a comment
    Hands-on: utilizing C code

    Cython: 0.032s
    CFFI: 0.015s

    ~2x speedup using CFFI. Considering that the C library is built with -O3, Cython is actually not bad!

  • Keichi Takahashi made a comment
    Hands-on: Optimising heat equation solver

    Pure Python: 17.34 s
    Cythonized and optimized: 0.03 s

    It took me a while to figure out I had to cdef the loop counters...

  • Keichi Takahashi made a comment
    Hands-on: Using C-functions

    Here's an iterative implementation that I came up with: https://gist.github.com/keichi/4cb14484ab68c685ec729a6cf8232530

  • Keichi Takahashi made a comment
    Avoiding explicit loops with NumPy arrays

    Loops with dependencies (reductions, scans, etc.) are difficult to vectorize.

  • Keichi Takahashi made a comment
    Hands-on: Numerical integration

    dx=0.1 S=0.9296501041259554
    dx=0.01 S=0.9992078366500485
    dx=0.001 S=0.9992037149227544
    dx=0.0001 S=0.9999036736218789
    dx=1e-05 S=0.9999936732092701
    dx=1e-06 S=0.9999996732051449

  • Keichi Takahashi made a comment
    Hands-on: Performance analysis of heat equation solver

    Total runtime was 16.721 s on my Ubuntu VM (running on a MacBook Pro, Intel Core i7-8559U). 16.631 s was spent in evolve().

  • Keichi Takahashi made a comment
    Hands-on: Performance analysis of heat equation solver

    I couldn't find heat_simple.py in the repository. Maybe heat_main.py?