home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!sdd.hp.com!mips!cprice
- From: cprice@mips.com (Charlie Price)
- Newsgroups: comp.sys.mips
- Subject: Re: pixie/cord info wanted
- Date: 30 Jul 1992 00:40:35 GMT
- Organization: MIPS Technology, Inc
- Lines: 48
- Message-ID: <l7eek3INNols@spim.mips.com>
- References: <1992Jul14.225221.16852@pure.com>
- NNTP-Posting-Host: lloyd.mips.com
-
- In article <1992Jul14.225221.16852@pure.com> rpeck@pure.com (Ray Peck) writes:
- >
- >I'm looking for experiences with pixie and cord. Right now, I work on
- >Sparcs (ugh), but we'll likely be porting some time to MIPS boxes, et.
- >al.
- >
- >I was wondering what performance benefits programmers see from using
- >cord to do function locality improvement. I asked this on comp.arch,
- >expecting to get SPEC timings back w/ and w/o cord, but no-one seems
- >to have any real data.
- >
- >Any experiences out there? How much of a win is cord in real code?
- >
- >Ray Peck rpeck@pure.com
-
- Well, it all depends.
-
- The cord pass of the compilation process doesn't do very much,
- it just arranges that functions are put into the final binary
- in order of the number of cycles they use. This information is
- based on one or more runs of a pixified program with the feedback
- files created as a side effect of profile.
- The intent is to make the high-cycle use routines less likely to
- collide in the Icache.
-
- Cord will improve your execution time if it eliminates Icache collisions.
- Thrashing collisions can happen nearly any time,
- but are statistically less likely as the Icache gets larger.
- For an RC3360 with a 64KB Icache you probably wouldn't see too
- much improvement a lot of the time. Of course a large program
- can just happen to have a deadly thrash collision and cord would
- tend to fix that.
-
- The R3000-based workstations with 32KB Icaches are a bit more sensitive
- to this problem and a corded program would win more there.
-
- Experience: I have seen cord improve performance ~5% for this sort of machine.
- I have seen no amazing wins at this level, but I haven't tried it out
- a lot either.
-
- For an R4000-based system with a primary cache only 8KB and memory a
- long way away we have seen more dramatic results. I've heard about
- a customer application that got a 30% performance improvement from
- being corded.
-
- --
- Charlie Price cprice@mips.mips.com (408) 720-1700
- MIPS Technology Inc. / 928 Arques Ave. MS 1-03 / Sunnyvale, CA 94088-3650
-