phil@amdcad.UUCP (Phil Ngai) (01/12/87)
Have any studies been done on how the kernel's locality of reference
compares to that of a typical application such as the compiler or
nroff? I am interested in cache design issues. How big a cache? How
associative? What kind of hit rates? How big a line size? How much
time is spent in reload? What is the total running time? Writeback
policies? Most of these are general cache design issues but I am
especially interested in whether the kernel acts significantly
differently than the average program the user might write and what
parameters might be changed to allow for differences.
--
My left foot is digital, my right foot is analog.
Phil Ngai +1 408 749 5720
UUCP: {ucbvax,decwrl,hplabs,allegra}!amdcad!phil
ARPA: amdcad!phil@decwrl.dec.comgnu@hoptoad.uucp (John Gilmore) (01/16/87)
The January 1987 (vol 38 #1) issue of the H-P Journal contains an
article on the HP3000 series 70 cache, which goes into great detail on
how they instrumented the older models to determine what cache
parameters to change in the new model. One quote from the article,
and a reference:
"...it has been shown that operating systems have rather poor locality
compared to user programs." They don't give a derivation for this finding.
The reference:
A. J. Smith, "Cache Evaluation and the Impact of Workload Choice",
Reprot UCB/CSD85/229, March 1985, Proceedings of the 12th International
Symposium on Computer Architectures, June 1985, pp 64-75.
To subscribe to the HP Journal, send your request to: HP Journal, 3200
Hillview Ave, Palo Alto CA 94304 USA. "Allow 60 days." They'd probably
just send this issue if you asked. The journal seems to have an assortment of
technical articles on their computers (especially the RISC machines), software,
and instruments, each about 1/3. It's free and I find it worth reading.
--
John Gilmore {sun,ptsfa,lll-crg,ihnp4}!hoptoad!gnu gnu@ingres.berkeley.edu
/* No comment */mash@winchester.UUCP (John Mashey) (01/19/87)
In article <1666@hoptoad.uucp> gnu@hoptoad.uucp (John Gilmore) writes: >The January 1987 (vol 38 #1) issue of the H-P Journal contains an >article on the HP3000 series 70 cache, which goes into great detail on >how they instrumented the older models to determine what cache >parameters to change in the new model. One quote from the article, >and a reference: >"...it has been shown that operating systems have rather poor locality >compared to user programs." They don't give a derivation for this finding. >The reference: >A. J. Smith, "Cache Evaluation and the Impact of Workload Choice",... Another good reference is: Douglas W. Clarke and Joel S. Emer, " Performance of the VAX 11/780 Translation Buffer: Simulation and Measurement", ACM Trans. on Computer Systems 3(1), pp31-62 (Feb 1985). Of course, this is about locality in the TLB, which is different from that of cache locality. However, it does give good numbers and analysis. [Kernels are worse in both cases, as it happens.] -- -john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc> UUCP: {decvax,ucbvax,ihnp4}!decwrl!mips!mash, DDD: 408-720-1700, x253 USPS: MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086
josh@polaris.UUCP (01/22/87)
In article <14295@amdcad.UUCP> phil@amdcad.UUCP (Phil Ngai) writes: >Have any studies been done on how the kernel's locality of reference >compares to that of a typical application such as the compiler or >nroff? I am interested in cache design issues. How big a cache? ... You might try looking at "Cache Memory Performance in a Unix Environment" by Cedel Alexander, William Keshlear, Furrokh Cooper and Faye Briggs, Computer Architecture News, Volume 14, Number 3, June 1986, pp. 41-70. They have lots of numbers and graphs and the traces they use to drive the cache models are obtained from a system running Genix and include the kernel, unlike most such studies. The traces are obtained by modifying the hardware of a National 32016 based box. -- Josh Knight, IBM T.J. Watson Research josh@ibm.com, josh@yktvmh.bitnet, ...!philabs!polaris!josh