Just a dimensional analysis point but, it should be P bits/line (since every line has its own vector). Otherwise, you have units of lines x percent.
This comment was marked helpful 1 times.
yingchal
As we have seen from the exam 1, the full bit vector overhead is acceptable sometimes. For example, on a quad-core processor, a full bit vector needs only 4 additional bits for each cache-line, the overhead is less than 1%. Under this situation, we should put effort on other performance issues instead of reducing the overhead of directory.
Overhead calculation:
$$ 100\%*\frac{P\text{ bits}}{(64\text{ bytes/line})(8\text{ bits/byte})} $$
This comment was marked helpful 1 times.
Just a dimensional analysis point but, it should be P bits/line (since every line has its own vector). Otherwise, you have units of lines x percent.
This comment was marked helpful 1 times.
As we have seen from the exam 1, the full bit vector overhead is acceptable sometimes. For example, on a quad-core processor, a full bit vector needs only 4 additional bits for each cache-line, the overhead is less than 1%. Under this situation, we should put effort on other performance issues instead of reducing the overhead of directory.
This comment was marked helpful 0 times.