home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.unix.admin
- Path: sparky!uunet!nih-csl.dcrt.nih.gov!helix.nih.gov!rvenable
- From: rvenable@helix.nih.gov (Richard M. Venable)
- Subject: Re: Need info on job queuing system
- Message-ID: <1992Nov20.103234@helix.nih.gov>
- Sender: postman@alw.nih.gov (AMDS Postmaster)
- Reply-To: rvenable@helix.nih.gov (Richard M. Venable)
- Organization: National Institutes of Health
- References: <GERMAIN.92Nov19165405@tohi.DMI.USherb.CA>
- Distribution: comp.unix.admin,comp.unix.questions,comp.unix.aix
- Date: Fri, 20 Nov 1992 15:32:34 GMT
- Lines: 62
-
- In article <GERMAIN.92Nov19165405@tohi.DMI.USherb.CA>,
- germain@tohi.DMI.USherb.CA (Luc Germain - Chimie) writes:
- |> We are looking for a job queuing system that would manage many UNIX
- |> boxes (actually we have 5 RS/6000 running AIX) used for intensive
- |> scientific calculations (mainly fortran programs). We would like
- that
- |> the users submit their programs to a unique job queue, specifying
- some
- |> requirements about their jobs (memory needs, estimated time, special
- |> libraries...) and the queuing system would automatically dispatch
- the
- |> jobs to the most available station on the network that meets the
- |> requirement.
- |>
- |> I heard about "NQS": can this program be configured to meet our
- |> requirements? Is there other queuing systems available? (Public
- domain
- |> or commercial). Anybody want to comment on their experiment with
- such
- |> installations?
-
- ---
-
- NQS does do some of these things, but consider the following scenario:
-
- User A submits 6 jobs, each of which will run for 10 hours and must run
- to completion before the next job starts (sequential).
-
- User B submits 4 jobs, each requiring 12 hours, about a half later,
- with
- the same sequential requirement.
-
- Under NQS, all of user A's jobs will before any of user B's
- (FIFO: first in, first out), allowing user A to monopolize the queue.
-
- Worse yet, user A's second job would start before the first was
- fininshed.
-
- ---
-
- This is why I wrote a simple batch system that (1) interleaves, to stop
- queue hogs, and (2) preserves the sequential nature of submitted jobs.
- We have several HP 9000/7xx machines, with one queue per machine; no
- dynamic load balancing, it's up to the users. If there's interest, I
- may
- put together a slightly more secure version and assemble a distribution
- package; it's mostly /bin/csh scripts and one very short FORTRAN
- program.
-
- Also, IBM is beta testing their own AIX multiple host queuing system,
- code
- named CONDOR and scheduled to be released under the name UNIJES which
- does
- restrict users from monopolizing a queue. To my knowledge, it does not
- handle the sequential job problem.
-
- -----------------------------------------------------------------
- Rick Venable | "Eschew
- FDA/CBER Biophysics Lab | Obfuscation"
- rvenable@helix.nih.gov | -- the Phantom Nerd
- -----------------------------------------------------------------
-
-