home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!olivea!mintaka.lcs.mit.edu!KANCHENJUNGA.LCS.MIT.EDU!tuna
- From: tuna@KANCHENJUNGA.LCS.MIT.EDU (Kirk 'UhOh' Johnson)
- Newsgroups: comp.arch
- Subject: Re: Multi threaded CPUs
- Message-ID: <1992Sep3.021846.4137@mintaka.lcs.mit.edu>
- Date: 3 Sep 92 02:18:46 GMT
- References: <BtxGvn.8MH@pix.com> <BtxHG0.8rz@pix.com> <BtxJDM.CK@rice.edu>
- Sender: news@mintaka.lcs.mit.edu
- Organization: Massachvsetts Institvte of Technology
- Lines: 65
-
- stripes@pix.com (Josh Osborne) wrote:
-
- A little while ago I read about a RISC CPU (SPARC I think) that
- some researchers had designed. It had a cache of threads (I got
- the impression that that includes registers, and possibly MMU
- state as well), four in the current chip, and when there was a
- pipe stall the execution would just switch to another thread.
-
- to which preston@cs.rice.edu (Preston Briggs) added:
-
- Anant Agarwal, at MIT, was working on a modification of the SPARC,
- called Sparcle, to be used as a node in a large-scale
- multiprocessor called Alewife. He gave a talk on the project last
- year at a workshop during Supercomputing '91. Unfortunately, I
- have only a brief abstract here.
-
- My memory from the talk was that they used each set of registers
- in the SPARC's register windows to support a different thread. Can
- switch threads in 11 cycles. Probably use the multithreaded
- behavior to hide communication latency.
-
- this is essentially correct. in SPARCLE, we use the SPARC register
- windows to keep the state (registers) for multiple contexts (four, in
- this implementation) in hardware so we can rapidly switch between
- them (14 cycles). SPARCLE will be used as the integer unit in Alewife;
- as preston suggests, the multithreading feature is used as a
- communication (and synchronization) latency hiding mechanism. note
- that SPARCLE only switches contexts on cache misses and explicit
- program directives; random pipe stalls for other reasons do not cause
- context switches. also note that the SPARCLE contexts do not include
- MMU state.
-
- we've had working SPARCLE silicon (40 MHz) for almost six months now;
- we're furiously hacking to get the rest of the Alewife hardware ready
- to go.
-
- anant did give a SPARCLE talk at hot chips, but alas there was no
- associated paper. there was an early SPARCLE paper in ISCA'90, though:
-
- @InProceedings(Agarwal-april,
- Author = "Anant Agarwal and Beng-Hong Lim and David A. Kranz and
- John Kubiatowicz",
- Title = "{APRIL: A Processor Architecture for Multiprocessing}",
- BookTitle = "Proceedings 17th Annual International Symposium on
- Computer Architecture",
- Pages = "104-114",
- Month = "June",
- Year = "1990")
-
- we are also hoping to have another SPARCLE paper (with quite a bit
- more detail) out RSN.
-
- there were a couple of papers that appeared as posters in the most
- recent ISCA ('92) which talked about issues related to multithreaded
- processors -- one from jim laudon (stanford), one from dave culler
- (berkeley); both of these made interesting reading. if anybody is
- interested and can't get the info themselves, i can probably chase up
- better references, e-mail addresses, or postscript for these.
-
- share and enjoy,
-
- kirk
- --
-
- Kirk Johnson ------ tuna@lcs.mit.edu ------ Yow! Have I graduated yet?
-