From: young_r@eisner.decus.org Sent: Friday, July 23, 1999 12:44 AM To: Info-VAX@Mvb.Saic.Com Subject: Re: Whither VMS? In article <7n7jgq$aou$1@pyrite.mv.net>, "Bill Todd" writes: > Since this sub-thread tied into one I was having in comp.arch ("Fine-grained > locking"), I brought it up (in somewhat less detail) there to try to get > more views on the matter (only one so far). (Of course, carrying on here is > fine too...) > Bill, Saw your other posts in comp.arch... I am flattered that you would even suggest I have helped change your thinking on this subject. If you haven't do pick up Greg Pfister's book "In Search of Clusters", easy lookup on Amazon, etc... This caching issue has been kicked around for quite a while.. at the risk of looking slightly silly, here is a post from Dec. 1997... Greg Pfister isn't a fan of Simple-COMA, you can find links to COMA here and there. It will be interesting how all this plays out.. I am leaving this thread for now unless someone responds and adds something.. you and I done whipped this horse so long and hard he could blow past Secretariat. Note, I couldn't find Andrew's original post referencing Sun's COMA "plans" doing a http://www.deja.com/ search, there may still be something on Sun's Website. Actually found something from a search for Andrew and Coma and a post he may have referenced: http://playground.sun.com/pub/S3.mp/simple-coma/isca-94/paper.html Big systems for commercial use, 64 dollar question: Where is Sun headed? What will Sarenghetti do with 1000 CPUs? Is Sarenghetti spelled like spaghetti? Will API and their cheetah campaign, totally muddle the UltraSparc III (aka "Cheetah") tie-in to Sarenghetti? Where is HP headed? IBM and Sequent appear more easily discernable. By the way, the reason there isn't much interplay on this topic (IMHO) is because it is a VERY important area for all involved and folks don't blab until the ink is dry on the patent applications, project plans well underway, etc. (my opinion again). Rob P.S. Always slightly embarassing to read one's post (especially if one is posturing quite a bit) a year or two out. Also, it might be 2 years MORE before data caching gets into shared memory (if it does that is ;-) , I wrote the below as if it was right around the corner. --------- From: young_r@eisner.decus.org (Rob Young) Subject: Re: Review of Forest Baskett's talk at UW, 11/13/97 Date: 02 Dec 1997 Newsgroups: comp.arch In article <65um55$bdi@spool.cs.wisc.edu>, schoinas@fox.cs.wisc.edu (Ioannis Schoinas) writes: >young_r@eisner.decus.org (Rob Young) writes: > >+ Wrap a VMS cluster around the whole thing? Humm. Don't think >+ that it is that.. and no it isn't "shared memory" because it >+ doesn't have to be. It can be shared nothing: > >+http://www.people.memphis.edu/~flowers/galaxy/sld021.htm >+http://www.people.memphis.edu/~flowers/galaxy/sld013.htm > >+ No cluster interconnect in slide 13, a single box with >+ multiple VMS instances inside. > >BTW, how is this different than: >http://www.unixreview.com/backissu/sun9706.htm >which seems to have been available for Sun servers >for quite some time now. How does the share nothing model of Sun's Domains compare to the share nothing model of APMP? How does a Huyndai compare to a Mercedes? They are both cars after all. . . Andrew Harrison and myself beat this around here at the end of May. Sun's future beyond their initial stab at Domains is tilted towards COMA. You can search "old" http://www.dejanews.com/ for: Andrew Harrison AND COMA to confirm that. COMA is all about attractor memories and sort of indirectly answers Andy Glew's inquiry as to how Sun's Domain model shares memory among/across domains in that if they could share memory now, what would a COMA gain you? Besides, reading the Sun whitepaper on page 6 you find: http://www.sun.com/servers/ultra_enterprise/10000/wp/ "When more than one domain is active, domain filtering restricts communications among different domains; that is, no data from outside a domain can reach the domain." But back to your question... Multiple instances of VMS do much more sharing whether internal to a box (futures) or cluster-wide. Individual instances can maintain data cache and communicate at much higher [future] speed internal to the same physical box as today the best a VMS cluster can do is over Memory Channel. Regardless, the APMP "share nothing" lends itself to running a parallel RDBMS on each instance sharing the same database as locking/sharing is coordinated among all instances in traditional VMS Distributed Lock Mastering fashion. The Sun marketing white paper makes no mention of such a thing as it isn't yet possible (see above). There are advantages to each APMP model. But focusing on something that is pointed out that looks like a management headache is seen on page 9: "Finally, because partitions stand as a self-contained systems, each domain requires its own boot disk and network connection." Would anyone look forward to when Domains number beyond 5 to 16 and you get to maintain 16 boot disks? Also, Domains add/remove 4 CPUs at a time. That isn't near the granularity of APMP (dragging and dropping a single CPU). How do they compare overall? APMP promises: 1) Share Nothing OR 2) Partially Share OR 3) Share Everything Sun's future COMA promises attractor memories. One wonders if the folks receiving very HEAVY Non-Disclosure right now see very good futures for APMP. One final tweek here... in that Sun white paper you read this: "Shared-Nothing Environments -- An RDBMS Advantage" Really? In what regard? Protection? How about an APMP tagline: "Shared-Everything Environments -- An RDBMS Advantage" Now THAT is more like it!!! One would want to take advantage of several hundred Gigabytes of data cache (seen by all instances). Wouldn't YOU? Rob