They are orthogonal concepts. Directory-based coherence is about scaling the implementation of coherence efficiently. Clearly, scaling challenges will be very important in a huge distributed shared memory system like Blacklight. However, efficient implementation of coherence is important in smaller scale systems, and that's why you see modern single-chip Intel processors implementing a version of directory-based coherence (See slide 44).
I see. So Symmetric Multi-Processors are more efficient than distributed shared memory on smaller scale systems. But on a bigger scale systems like Blacklight, Distributed Shared Memory system would be more efficient since it can reduce the communication bandwidth.
@aznshodan: In practice about every system today is non-symmetric in some way. The degree of asymmetry grows as the system size grows.