Previous - Up - Next

18.6   Using g-cache

Let us have a more detailed look at g-cache. It has the following features:

• Configurable number of lines, line size and associativity. Note that the line size must be a power of 2, but the only restriction on the cache structure is that the number of lines divided by the associativity must be a power of two.
• Physical/virtual index and tag.
• Configurable write allocate/write back policy.
• Random, true LRU or cyclic replacement policies. It is easy to add new replacement policies.
• Sample MESI protocol.
• Support for several processors connected to one cache.
• Configurable penalties for read/write accesses to the cache, and read/write accesses initiated by the cache to the next level cache.
• Cache miss profiling.

Transactions are handled one at a time; all operations are done in order, at once, and a total stall time is returned. The transaction is not reissued afterward. Here's a short description of the way g-cache handles a transaction (we assume a write-back, write allocate cache):

• If the transaction is uncacheable, g-cache ignores it.
• If the transaction is a read hit, g-cache returns penalty_read cycles of penalty.
• If the transaction is a read miss, g-cache asks the replacement policy to provide a cache line to allocate.
• The new cache line is emptied. If necessary, a copy-back transaction is initiated to the next level cache. In this case, a penalty of penalty_write_next is counted, added to the penalty returned by the next level.
• The new data is fetched from the next level, incurring penalty_read_next cycles penalty added to the penalty returned by the next level.
• The total penalty returned is the sum of penalty_read, plus the penalties associated with the copy-back (if any), plus the penalties associated with the line fetch.

Note the usage of penalty_read/write and penalty_read/write_next: a write-through cache would always take a penalty_write_next penalty independently of the fact that a write is a hit or a miss, but g-cache always reports hits and misses according to the line lookup, not based on the fact that a transaction is propagated to the next level.

Previous - Up - Next