webbhorn / libdsmu

6.824 final project

libdsmu

Distributed Shared Memory in Userspace Library for Linux Applications. Modeled after the Ivy system. Final project for 6.824.

Authors

• Webb Horn
• Tim Donegan
• Julian Gonzalez
• Ameesh Goyal

Setup

libdsmu was tested on Ubuntu 14.04 x86/64 running the 3.13.0-24-generic kernel. Older kernels on a non-x86/64 machine probably will not work because the code that detects fault types is non-portable.

Third party libraries:

$ sudo apt-get install libb64-dev

Build

$ cd src
$ make

Running the manager

$ python manager/manager.py &

Running the Matrix Multiply Benchmark

matrixmultiply uses the alternation-style sharding algorithm for nodes-to-rows mapping, while matrixmultiply2 allocates rows to nodes in contiguous chunks to reduce the effects of page thrashing. They are both launched the same way.

First, start the manager:

$ python manager/manager.py &

Next, start up instances of matrixultiply in parallel. To start three instances, launch them like this:

$ cd src
$ make
$ ./matrixmultiply2 127.0.0.1 4444 1 3 && \
  ./matrixmultiply2 127.0.0.1 4444 2 3 && \
  ./matrixmultiply2 127.0.0.1 4444 3 3: