Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss
Directory organization in multi level caches with non-uniform block sizes
31 views
Skip to first unread message
NVK
Jun 6, 2013, 2:08:11 AM6/6/13
to
Hi
I'm making a simulation model of a multi-level hierarchical cache system, that uses a directory-based MESI protocol.
Every cache has a directory. The directory entry for a block stores information about the state of the block and the list of lower-level caches that own a copy of the block.
My question is this:
If we have different block-sizes for different levels in the hierarchy (which is typical in real multi core , multiprocessor systems, such as ones using Intel QPI) What does the directory look like?
Does it maintain list of sharers per block, or per sub-block (where sub-block is the block size in child caches)?
thanks for any insights/links
I'm making a simulation model of a multi-level hierarchical cache system, that uses a directory-based MESI protocol.
Every cache has a directory. The directory entry for a block stores information about the state of the block and the list of lower-level caches that own a copy of the block.
My question is this:
If we have different block-sizes for different levels in the hierarchy (which is typical in real multi core , multiprocessor systems, such as ones using Intel QPI) What does the directory look like?
Does it maintain list of sharers per block, or per sub-block (where sub-block is the block size in child caches)?
thanks for any insights/links
Ivan Godard
Jun 6, 2013, 2:41:53 AM6/6/13
to
MitchAlsup
Jun 6, 2013, 11:53:52 PM6/6/13
to
On Thursday, June 6, 2013 1:08:11 AM UTC-5, NVK wrote:
Why do you HAVE to maintain said list. As long as the forgetfulness of
the directory implies that you can't tell and have to go out and snoop,
you can rebuild the lost impfomation, and dramatically save bits in the
directory.
But on the question of different sized blocks, this is a pure trade off
between directory bits (area and power) against the interconnect traffic
when the directory is less than forthcomming.
I, personally, would set up the simulations to get at the heart of the
mater, irrespective of the block size differences in different caches.
Mitch
Why do you HAVE to maintain said list. As long as the forgetfulness of
the directory implies that you can't tell and have to go out and snoop,
you can rebuild the lost impfomation, and dramatically save bits in the
directory.
But on the question of different sized blocks, this is a pure trade off
between directory bits (area and power) against the interconnect traffic
when the directory is less than forthcomming.
I, personally, would set up the simulations to get at the heart of the
mater, irrespective of the block size differences in different caches.
Mitch
0 new messages