On Mar 17, 2006, at 11:17 AM, Mike Marty wrote:
The [M_w and MM_w] states are an unneeded optimization to allow a
processor to retire 1 or more instructions before losing permission
to the block.
What Mike said is correct, and let me add a bit of context.
As Mike said, these wait states are basically performance
optimizations to prevent a block from being stolen away too early.
At one point, we were seeing some differences between our snooping
protocols and other protocols in terms of performance under high
contention. We discovered that our more directory-like protocols had
a built in round-trip latency into their block handoff, allowing the
processor to keep a block for some limited time to perform multiple
operations on the block (for example, both the test and test&set).
In contrast, without these states it would still enforce forward
progress, but allow the block to be stolen too quickly. Of course
any time you're adding extra delay to something, it can also cause
slow downs in some cases.
The main reason these wait states were added was to make all four of
the protocols in my dissertation (snooping, directory, hammer, and
token) act similarly.
- Milo
--
Milo M. K. Martin (milom@xxxxxxxxxxxxx)
http://www.cis.upenn.edu/~milom/
Assistant Professor
Computer and Information Sciences Department
University of Pennsylvania
-----Original Message-----
From: gems-users-bounces@xxxxxxxxxxx [mailto:gems-users-
bounces@xxxxxxxxxxx] On Behalf Of arrvindh shriraman
Sent: Friday, March 17, 2006 9:59 AM
To: gems-users@xxxxxxxxxxx
Subject: [Gems-users] MOESI broadcast protocol
I followed the link in milo's dissertation for a MOESI protocol.
http://www.cs.wisc.edu/multifacet/theses/milo_martin_phd/snooping
I dont quite understand the need for acks and timeouts and the
states M_w
and MM_w which you characterize as wait states. I do comprehend
that M is
analogous to E.
Yahoo! Travel
Find great deals to the top 10 hottest destinations!
_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
--
Milo M. K. Martin (milom@xxxxxxxxxxxxx)
http://www.cis.upenn.edu/~milom/
Assistant Professor
Computer and Information Sciences Department
University of Pennsylvania
|