home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!think.com!enterpoop.mit.edu!mintaka.lcs.mit.edu!ai-lab!zurich.ai.mit.edu!gjr
- From: gjr@zurich.ai.mit.edu (Guillermo J. Rozas)
- Newsgroups: comp.sys.hp
- Subject: Re: clusters, etc.
- Date: 9 Jan 93 08:32:36
- Organization: M.I.T. Artificial Intelligence Lab.
- Lines: 71
- Message-ID: <GJR.93Jan9083236@chamarti.ai.mit.edu>
- References: <28226@dog.ee.lbl.gov> <17780308@hpfcmgw.FC.HP.COM>
- Reply-To: gjr@zurich.ai.mit.edu
- NNTP-Posting-Host: chamartin.ai.mit.edu
- In-reply-to: perry@hpfcmgw.FC.HP.COM's message of Sat, 9 Jan 1993 00:05:46 GMT
-
- In article <17780308@hpfcmgw.FC.HP.COM> perry@hpfcmgw.FC.HP.COM (Perry Scott) writes:
-
- | From: perry@hpfcmgw.FC.HP.COM (Perry Scott)
- | Date: Sat, 9 Jan 1993 00:05:46 GMT
- | I don't work near the people that are investigating NFS Diskless so I'm
- | not an expert. The negative feedback I have gotten from customers is:
- |
- | 1) Not industry standard
- | 2) Diskless doesn't work through IP routers
- | 3) CDFs are yet another implementation. It could be done with a
- | dedicated root file system and symlinks/mountpoints.
- |
- | There are also some internal disadvantages to HP in supporting diskless
- | in addition to industry standard NFS.
-
- We have a cluster with approximately 30 workstations (300s and 700s).
- Before we clusterized our workstations (a few years ago), all
- workstations booted locally and used NFS to mount the server's file
- systems.
-
- We had all sort of kludges for sharing files and for having some
- machines use specific versions, and required great pains and time when
- updating the operating system or major software components.
-
- It was an incredible headache!
-
- The cluster software made our lives considerably easier. I usually
- mention it to everyone who asks about HP equipment as one of the
- strongest selling points.
-
- What does the cluster software buy us:
-
- - A reliable, consistent, fast network file system. I've had horrible
- problems with NFS on HPs, Suns, and other hardware, and NFS is
- considerably slower.
-
- - Easy upgrade and maintenance (changing /etc/netlinkrc, for example)
- because there is only one copy of the system (or two for heterogenous
- clusters). Any other software that allows machines to boot over the
- network but have local swap disks could probably accomplish this as
- well, but the cluster software does this beautifully.
-
- - Easy way to share configuration files and to disambiguate those that
- should not be shared by using CDFs. The MIT AI lab also has a large
- Sun setup, and all the /var directories and mesh of symlinks are a
- very poor solution compared to CDFs.
-
- - Trivial addition of new machines to our setup. When we obtain new
- machines, it usually takes < 10 minutes to have them working as part
- of the cluster with everything set up. Note that this is the only
- thing we use SAM for here, and I wish I did not have to use it for
- this either.
-
-
- I'm not so attached to cluster software that I will be unhappy with
- anything else, but I would consider it a large step backwards to have
- to put up with:
-
- - an unreliable and slow network file system,
-
- - having to upgrade 30 machines every time there is an operating
- system upgrade or configuration change, rather than just one,
-
- - no real way to disambiguate shared files,
-
- - lack of software to conveniently integrate a new machine into our
- setup.
-
-
- In other words, HP, if you do away with cluster software, please make
- sure that you don't throw away the features mentioned above.
-