So if I'm understanding this correctly then won't we have high queue access contention since each thread would try to pull a vertex from the work queue? Is this better here rather than having a work queue per thread and maybe do work stealing?
jrgallag
The work queue could be implemented that way - remember that this is an abstraction presented to the programmer, rather than an implementation. GraphLab might even decide which type to use depending on the parameters of the system.
So if I'm understanding this correctly then won't we have high queue access contention since each thread would try to pull a vertex from the work queue? Is this better here rather than having a work queue per thread and maybe do work stealing?
The work queue could be implemented that way - remember that this is an abstraction presented to the programmer, rather than an implementation. GraphLab might even decide which type to use depending on the parameters of the system.