That is a really open question :-)
It depends on what you the application is.
For instance, coordinate descent algorithms can fit in GraphLab assuming the objective function is sufficiently sparse (For instance, matrix factorization). In some other cases the dual problem is sparse and we get a "message passing-like" algorithm like Belief Propagation and that fits in GraphLab quite nicely. Ant Colony optimization might indeed fit quite nicely for TSP solving (though I have not really looked at it closely), not sure about other cases: might be case by case.