Chiming in here. This is Bryan Thompson, the PI for the MapGraph project. We have new results for multiple GPUs with up to 29 billion traversed edges per second on 64 K20 GPUs and 4.3 billion directed edges. The multi-GPU code base is not yet the full MapGraph / GAS API, but we will be fixing that soon. We will be extending the capabilities of the platform in a number of directions, including allowing multiple partitions per GPU so you can operate on graphs larger than the GPU RAM without having to scale-out, and so you can operate on huge graphs on GPU clusters.
GraphLab (a purely memory based system) is much faster than GraphChi (a disk-based system using the parallel sliding window algorithm). Both GraphLab and GraphChi use "vertex cuts" which amount to 2D partitioning, which is what we use in the multi-GPU version MapGraph. This is also what is used by the world record holder for BFS traversal, which is the Blue Gene/Q.
We used the synchronous engine for GraphLab and the specified version of graphlab because that is what existed and worked (without segfaults) when we collected the data for the SIGMOD paper in the summer of 2013. We did cross-check the performance of GraphLab with the GraphLab developers.
One of the main reasons that CPU parallel graph algorithms slow down is because they hit the memory bandwidth of the bus. GPUs have 10x or more memory bandwidth. The K20 has 204 GB/s. The K40 has 288 GB/s. The Pascal GPU (2016) will have over 1 TB/s of memory bandwidth and 24 GB of RAM.
I've listed the MapGraph Publications below. See http://mapgraph.io for more details.
- MapGraph: A High Level API for Fast Development of High Performance Graph Analytics on GPUs - Zhisong Fu, Bryan Thompson, Michael Personick (SYSTAP, LLC). GRADES 2014, June 22, 2014, Snowbird, Utah, USA. Copyright 2014 ACM 978-1-4503-2982-8. DOI:http://dx.doi.org/10.1145/2621934.2621936
GraphLab (a purely memory based system) is much faster than GraphChi (a disk-based system using the parallel sliding window algorithm). Both GraphLab and GraphChi use "vertex cuts" which amount to 2D partitioning, which is what we use in the multi-GPU version MapGraph. This is also what is used by the world record holder for BFS traversal, which is the Blue Gene/Q.
We used the synchronous engine for GraphLab and the specified version of graphlab because that is what existed and worked (without segfaults) when we collected the data for the SIGMOD paper in the summer of 2013. We did cross-check the performance of GraphLab with the GraphLab developers.
One of the main reasons that CPU parallel graph algorithms slow down is because they hit the memory bandwidth of the bus. GPUs have 10x or more memory bandwidth. The K20 has 204 GB/s. The K40 has 288 GB/s. The Pascal GPU (2016) will have over 1 TB/s of memory bandwidth and 24 GB of RAM.
I've listed the MapGraph Publications below. See http://mapgraph.io for more details.
- MapGraph: A High Level API for Fast Development of High Performance Graph Analytics on GPUs - Zhisong Fu, Bryan Thompson, Michael Personick (SYSTAP, LLC). GRADES 2014, June 22, 2014, Snowbird, Utah, USA. Copyright 2014 ACM 978-1-4503-2982-8. DOI:http://dx.doi.org/10.1145/2621934.2621936
- Parallel Breadth First Search on GPU Clusters - Z. Fu, H.K. Dasari, M. Berzins, B. Thompson. SCI Technical Report, SCI Institute, University of Utah. July 29, 2014. URL:http://www.sci.utah.edu/publications/Fu2014a/UUSCI-2014-002....