--------------------------------------------------------------------
CS 757 Parallel Computer Architecture
Spring 2012 Section 1
Instructor Mark D. Hill
--------------------------------------------------------------------

------------
Coherence 3 -- Directories
------------

OUTLINE
* Directories
* MSI
* E and O
* State (e.g., sharers)
* Where? LLC, etc.?


------------------------------

Chapter 8
---------
Directories

Simple idea

Actually, interleaved for different blocks

Figure 8.2 -- state // log2(owner) // sharer list

Figure 8.3 basic transactions

Deadlock and virtual networks

Table 8.1 MSI Directory Protocol -- in detail
* I to S -- 2 common, 1 race
...

Above stalls
* e.g., forward in transient
* no Upgrade

Adding E state
* We make E/M the owner with need to explicit replacement
* Could allow silent reeplacements of E, e.g., SGI Origin 2000 -- in E both
* memory (data) and E-core responses (S,E -- no data, M -- overriding data)

Adding O state

Representing Directory state
* complete sharer bit vector
* coarse sharer list
* limited i-pointers where each cache also has a pointer

Dir[i]B -- if >i broadcast
Dir[i]NB -- force recalls

Directory Organization
* Logically with every block in memory
... now add directory cache ...
* Directory cache -- backed by DRAM -- forget it
... now subtract state at memory
* Directory cache w/ inclusive LLC --  need recalls on LLC replacement (also
* can't have enough associativity) -- Intel
* Standalone directory cache -- dupicates tags, need recalls on LLC
* replacement -- AMD probe filter