On Fri, 5 Oct 2012, Beckmann, Brad wrote:
Hi Nilay,
To solve the first problem, can you use the Maybe_Stale AccessPermission
identify the last possible valid copy of the data block? It seems like
if a functional read can't find valid data anywhere else, then you are
guaranteed that the Maybe_Stale block at the directory is valid.
I think this solution will not work. It is possible that two controllers
at two different levels have a block with access permission Maybe_Stale.
In that case, unless we have the hierarchical information, we would not be
able to figure out which block to read. It seems it would not hard to add
this hierarchy information to the existing protocols, and then we can add
a hierarchical list of controllers to the ruby system.
With this change, we should be able to get rid of the physmem backing
store in Ruby, correct? Have you tried that sort of test with this
patch? Especially an application under SE mode that uses a lot of
system calls?
I never thought about removing the backing physical memory. But now that
you mention it, if the cache and directory controllers are the only ones
who should be writing to the memory (which I think is true in the SE
mode), then it should be possible to remove the backing physical memory.
This might lead to bugs being uncovered in the coherence protocols.
45 files modified...have I told you how much I dislike this patch. :)
I agree that a lot of files are being changed. With a high probability,
anyone who has modified Ruby would need to redo his/her patches. But then
I don't see if any changes can be elided. Even if I were to go with your
suggestion of building a flat list of buffers, most of the changes would
still remain as is.
Brad
-----Original Message-----
From: Nilay Vaish [mailto:[email protected]]
Sent: Wednesday, October 03, 2012 10:02 PM
To: Andreas Hansson; Default; Beckmann, Brad
Subject: Re: Review Request: ruby: augment network to support functional
accesses
Some comments on the patch --
* The way the code has been written, functional writes will always succeed,
but functional reads can still fail. This can be illustrated with an example.
Consider the situation when the directory has the actual data, but it believes
that the ownership of the data is with some other controller. This other
controller has relinquished the ownership with out modifying the data. The
info about relinquishment is in the network currently. The read would fail
since we need to know both pieces of info to figure out that the directory
has the data. We would need greater cooperation from protocols for reads
to succeed all the time.
* I have done some testing with the memtest (1 to 8 cores, 160000 loads).
But no testing has been done when dma requests/replies are part of the
system.
* No support for garnet as of now.
* Coherence protocols MI, MOESI CMP token have been changed slightly.
These changes have been vetted only with the memtest.
--
Nilay
--
Nilay
_______________________________________________
gem5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/gem5-dev
