e07@nikhefh.nikhef.nl (Eric Wassenaar) (12/18/89)
In the next several postings, I will report on my findings installing SR10.1.p on a DSP10020 server node, from a system managers standpoint. This was my first experience with SR10 and I am not satisfied. The setup is as follows: we have a DSP10020 with 32 MB of memory, and 2 * 2 760MB disk drivers, sector striped. Ring connection to a network of 30+ Apollo workstations, ethernet connection to the outer world, NFS to a BSD4.3 remote (non-apollo) host. Access is usually via rlogin. More such machines may be added. They are to act primarily as compute servers for cpu bound fortran jobs processing very large data files (200+ MB). I/O to/from magnetic tape and IBM 3480 cartridge is done on remote machines with access via NFS. In principle, they are pretty well suited for our type of computing. We think that they are ideal to off-load traditional centralized multi-purpose mainframes. Our experience with Apollo workstations started back in 1984, we've seen everything from SR7 to SR9.7.5, and we were mostly very satisfied. In contrast, if SR10.1.p is typical for SR10, then SR10 is unreliable, inconsistent, and unpredictable. If you only want to run your fortran job on a stand-alone single-user DN10000 you are probably ok. But if you want to embed such machine in a larger general multi-user environment and use it as a production machine, you may be confronted with unpleasant surprises. I did not follow every SR10 discussion on the net in full detail, so I may report issues that have been dealt with already. Apologies if I repeat things. Eric Wassenaar -- Organization: NIKHEF-H, National Institute for Nuclear and High-Energy Physics Address: Kruislaan 409, P.O. Box 41882, 1009 DB Amsterdam, the Netherlands Phone: +31 20 592 0412, Home: +31 20 909449, Telefax: +31 20 592 5155 Internet: e07@nikhef.nl