Everhart, Glenn From: Everhart [Everhart@gce.com] Sent: Thursday, January 07, 1999 6:59 PM To: Info-VAX@Mvb.Saic.Com Subject: [Fwd: RE: OVMS Sales.] In response to Rob Young: Thanks, Rob, for the exposition. The DLOCK article makes it clear that what is proposed is that there will be a lock manager in each SCSI device, so an OS can use that for synchronizing access to that device. The problem with this approach of adding more baroque design features to SCSI is however that nobody implements all of SCSI. Disk vendors in particular are well & truly squeezed in profits and system vendors must also decide whether to use SCSI at all, or use IDE. Given that profits are a few dollars per unit, and 90%+ of business is for PC type iron where only minimal SCSI commands are needed or used, who do they imagine will build any disks implementing this? Vendors may write an NT driver for x86 and once that runs, figure they're done. VMS users will recognize how it happens now and again that some SCSI feature VMS needs is missing and so the disk fails under VMS. A lot of those decisions got reworked for 7.1, so VMS tries much harder to get around missing features. However, a vendor relying on this kind of feature will be lucky to find ANY SCSI devices that actually work, and customers will find that almost EVERY SCSI device they try to use with such "cluster" software will fail with unix variants that try to rely on it. Manufacturers won't implement it. (Heck, on tapes and CDs, they mostly won't even bother decoding LUNs and get errors if anything tries to do multi-LUN access...a pure cost issue.) Meanwhile, VMS clusters (and galaxies!) can run happily even with IDE disks. (OK, the IDE bus doesn't allow multiple hosts, but disks work in clusters connected by other means.) Locking is still at fine grain, and no additional I/O bandwidth is thrown away at locking. In a galaxy it's just a bit of memory access...better & better. Thus, if you hear from a vendor whose "clusters" rely on fancy SCSI "standard" locking for fine grained access, be aware that essentially no devices will work with this. If you hear the word "cluster" not associated with the word "VMS", check REAL carefully whether the feature set claimed doesn't depend on this kind of vaporware standard feature. Relying on device locking for data safety is likely to be rather less reliable than relying on prayer.... > -----Original Message----- > From: young_r@eisner.decus.org [SMTP:young_r@eisner.decus.org] > Sent: Thursday, January 07, 1999 12:56 AM > To: Info-VAX@Mvb.Saic.Com > Subject: Re: OVMS Sales. > > In article <3693C56E.49EA30F@gce.com>, "Glenn C. Everhart" > writes: > > > Yes, the concept applies equally to other OSs. The same can however > > be said about clusters. Other OSs could have implemented them like > > VMS did. To date, none has (with the possible exception of the very > > very latest DEC Unix?) in that DLM synchronization is not endemic. > > Sun reps on the SCSI committee regularly propose commands to do > > persistent device reservations. Why? They have no DLM synch, so > > they have to get a processor to reserve disks for exclusive use as > > needed. Other unices that try to share devices in fact do so by > > serving them. > > > > You had mentioned this so often that I finally decided to > go fetch something along these lines. The Unix folks > have a project... a Global File System (why not just call > it a Cluster Common File System and be done with it??): > > http://gfs.lcse.umn.edu/index.html > > What's special here and what caught my eye is something they > call DLOCK (device lock): > > http://gfs.lcse.umn.edu/Pages/dlock.html > > "Find out about our new fine-grain synchronization command for SCSI > devices. > This command was presented by Seagate and the University of Minnesota to > the > SCSI X3T10 committee for inclusion in the SCSI standard on September 16, > 1998." > > I read the paper.. supporting up to 64 nodes. Short of locking > built in at a command level how does one avoid running the same > .sh on seperate nodes at the same time thinking you are all alone, > etc. I am sure there are stickier wickets here , like booting > at the same time, right Glenn ;-)? > > Looking at DUNIX: > > http://www.unix.digital.com/announcements/index.html > > Doesn't look like a CCFS yet. When there is, I would like to > see how binaries run unmodified. Maybe NFS forever, eh? > > DLOCK ... like I said before ... you do the best you can with > what you have to work with .. I guess. > > Rob