I believe the order in which nodes are processed is critical to achieve early convergence?
zhanpenf
More implementation details about GraphLab are in the following link: https://github.com/dato-code/PowerGraph
It says that for GraphLab, inter-process communication is accomplished over TCP-IP and MPI is used to launch and manage GraphLab programs. Each process uses pthreads to utilize multicore resources for parallel computation.
grose
@afa4 if it is, I wonder if it'd be good to have a way to specify the ordering somehow, or if GraphLab can figure that out in some/most/all cases
zhiyuany
@afa4, what you said really depends on the algorithm. GraphLab just assume your computation will converge and then compute for enough number of iterations.
I believe the order in which nodes are processed is critical to achieve early convergence?
More implementation details about GraphLab are in the following link: https://github.com/dato-code/PowerGraph
It says that for GraphLab, inter-process communication is accomplished over TCP-IP and MPI is used to launch and manage GraphLab programs. Each process uses pthreads to utilize multicore resources for parallel computation.
@afa4 if it is, I wonder if it'd be good to have a way to specify the ordering somehow, or if GraphLab can figure that out in some/most/all cases
@afa4, what you said really depends on the algorithm. GraphLab just assume your computation will converge and then compute for enough number of iterations.