Article: 80475 of comp.arch From: mash@mash.engr.sgi.com (John R. Mashey) Newsgroups: comp.arch Subject: Re: Memory Wall -- Comments? Date: 27 Apr 1999 00:22:16 GMT Organization: Silicon Graphics, Inc. Lines: 36 Message-ID: <7g2vvo$q0q$1@murrow.corp.sgi.com> References: <7g2989$pu8$1@nnrp1.dejanews.com> <7g2cfg$1m9e@news2.newsguy.com> <7g2gaa$575$2@msunews.cl.msu.edu> <925168594.28359@clint.waikato.ac.nz> 1) Well, latency is certianly *the* worst problem around, given that we'll likely have a lot of machines ~2001 with ~150 clock random accesses to main memory, and 4-6 isntructions/clock, we're certianly headed to 1000-(peak) instruction issues / cache miss. 2) For some codes, vector-processing may resurge. 3) Meanwhile, we really have to rethink algorithms: (a) For CMPSC students, most textbooks I've seen for the classic "Algorithms and Data Structures" undergraduate course don't say much (or anything) about memory hierarchies. [Can anybody point to an undergraduate textbook for this that does a good job on caches & their effects on a algorithmic performance analysis? I'd love to see one. I check the Stanford bookstore fairly often & complained to Hennessy, who said not to worry, Knuth was rewriting his books ... but then I happened to sit next to him at an event, and asked him when he'd be done, and the answer was 2011...] (b) For others who write code, at least a fairly minimal set of hints/warning should be accessible. (c) Many classic numerical codes have been reworked to make good use of caches, but caches don't work as well for random pointer-chasing. Computer micro-architectures can only do so much; it is really time to revisit much of the classic algorithm analysis in the face of the deepening memory hierarchies we'll have. -- -john mashey DISCLAIMER: EMAIL: mash@sgi.com DDD: 650-933-3090 FAX: 650-933-4392 USPS: Silicon Graphics/Cray Research 40U-005, 2011 N. Shoreline Blvd, Mountain View, CA 94043-1389 From surfnet.nl!news.belnet.be!news-feed.inet.tele.dk!bofh.vszbr.cz!enews.sgi.com!news.corp.sgi.com!mash.engr.sgi.com!mash Fri May 7 08:52:46 MET DST 1999 Article: 80972 of comp.arch Path: cwi.nl!surfnet.nl!news.belnet.be!news-feed.inet.tele.dk!bofh.vszbr.cz!enews.sgi.com!news.corp.sgi.com!mash.engr.sgi.com!mash From: mash@mash.engr.sgi.com (John R. Mashey) Newsgroups: comp.arch Subject: Re: Memory Wall -- Comments? Date: 7 May 1999 01:12:14 GMT Organization: Silicon Graphics, Inc. Lines: 63 Message-ID: <7gtele$s93$1@murrow.corp.sgi.com> References: <7g2989$pu8$1@nnrp1.dejanews.com> <7gcat6$k9n$1@news.iastate.edu> <7gcdu0$jma$1@news.doit.wisc.edu> <7gclsk$mhh$1@news.iastate.edu> NNTP-Posting-Host: mash.engr.sgi.com In article <7gclsk$mhh$1@news.iastate.edu>, john@iastate.edu (John Hascall) writes: |> How big will an N GB disk (for modest N) be in 2005? |> My guess, it will fit along with everything else on |> a thick credit-card-sized unit. Well, actually: 1) There are already 340MB drives of this size (IBM), with actual platter the size of a quarter, and the drive is actually 1.7" x 1.4" x .19", compared to a credit card (3.25" X 2" X (.03?)"). See http://www.ibm.com/storage/microdrive, amazing tech. 2) Disk density improved about 30%/year through ~1989, then went to 60% since then, and appear to be accelerating to ~100%/year over the next couple years, which says there should be a 1+ GB microdrive sometime in 2001. 3) Just for fun, at the IDEMA (disk industry) conference yesterday, Dataquest/Gartner's John Monroe predicted the # of Terabytes of disk shipped (with rounding by me): Year # TBs 1997 344,000 1998 704,000 1999 892,000 2000 1,923,000 2001 3,009,000 2002 5,645,000 If these are anywhere near close, there are some amusing conclusions: a) This is 1.75X/year, less than the 2X/year density idea (i.e., more smaller drives). If it gets to 2X/year, that means that every year, we will install *more* disk space than the entire accumulated disk space in the history of computing (i.e., 1, 2, 4 (>1+2), 8 (>1+2+4); we're not quite there yet, you have to repalce "year" by 15 months... b) Each year, the price of a disk of a given form factor drops a little, and the capacity goes up. If you buy the same number of disks each year going forward, and they double every year for a while, then: Year Relative Size % of space in year N N 16 ~52% N-1 8 ~26% N-2 4 ~13% N-3 2 ~06% N-4 1 ~03% Meaning, at least in terms of capacity provided (as opposed to seeks/sec), any disk more than 2-3 years old is almost useless... [I'm reminded of still having, from years ago, a 20MB Mac disk that cost $1000 at the time ... a fine deal.] c) Of course, bandwidth doesn't go up as fast as capacity, and latency ... sigh.... I figure, within 2-3 years, comparing average random access times (average seek plus 1/2 rotation), and peak execution rates (issue rate * Mhz), a random disk access = 20-30M (peak) instruction issues, meanign that you can execute a lot of code to avoid a disk access. "Money can buy bandwidth, but latency is forever". -john mashey DISCLAIMER: EMAIL: mash@sgi.com DDD: 650-933-3090 FAX: 650-933-4392 USPS: Silicon Graphics/Cray Research 40U-005, 2011 N. Shoreline Blvd, Mountain View, CA 94043-1389