Diirectory coherence protocols use a directory structure to track all the locations of all copies that exist for every block of memory, or for every block of cached block of memory. For the most common write-invalidate protocols, write requests must invalidate all cache copies of the block being written. Directory coherence protocols make this process easy by providing a list of all such cached copies. Similarly, on a read request, the directory provides an easy mechanism for locating the most recently modified copy of a memory blocks. This simplicity, however, is not without cost. Coherence directories can occupy large amounts of memory and/or on-chip area to store information about all cached copies that exist. Additionally, the latency of accessing a logically centralized directory can introduce latency overheads that reduce overall system performance. Finally, directory coherence protocols require careful verification to ensure atomicity of certain actions and to avoid race conditions that can lead to dead-lock or live-lock conditions. While all three of these properties make directories unattractive options for certain situations, the large size of the directory is especially problematic for CMPs where on-chip area is at a high premium.