Path: utzoo!utgpu!watserv1!watmath!att!tut.cis.ohio-state.edu!zaphod.mps.ohio-state.edu!usc!elroy.jpl.nasa.gov!turnkey!orchard.la.locus.com!fafnir.la.locus.com!fafnir.la.locus.com!richard From: richard@locus.com (Richard M. Mathews) Newsgroups: comp.unix.large Subject: Re: Survey Message-ID: Date: 11 Sep 90 01:41:08 GMT References: <6f7y02Ubc6wm01@amdahl.uts.amdahl.com> <25894@boulder.Colorado.EDU> Organization: Locus Computing Corporation, Los Angeles, California Lines: 38 fwp1@CC.MsState.Edu (Frank Peters) writes: >Well, we are developing a different kind of 'large system' that has >its own unique complexities. Good point -- large networks are also supposed to be part of this newsgroup. Actually, we have both kinds of "large" systems. As I said before, we have many AIX guests running on our 3090. We also have a handful of other 370s, each with a number of AIX guests. Finally we have many PS/2s running AIX. Groups of these are connected via TCF, and we use good old fashioned telnet, rlogin, NFS, etc. to connect the clusters. >(3) Load balancing. In a single box balancing the load among several > CPUs is relatively straitforward (at least in concept). When > your CPUs are spread across a dozen or more machines how do you > avoid the situation of one machine being sunk to its knees while > another is nearly idle. When you add multiple classes of > processor (is a 4/490 at 50% more loaded than a sparstation at > 30%?) or multiple types (how do the above two compare to a > decstation 3100 at 40%?) this issue can become a nightmare. TCF allows processes to migrate between machines, and I know there are others developing similar capabilities. I can send a signal to a process to request that it move to a new site (by default, to the site from which the signal was sent). A load leveling daemon could be written (but one does not come with TCF) which automatically moves processes around in response to varying load. A difficulty is deciding which processes to move -- it would be a shame to waste time moving an I/O bound process which is currently accessing local data. In our environment I have found it quite sufficient to be able to manually move things when the load goes up. Richard M. Mathews Locus Computing Corporation richard@locus.com lcc!richard@seas.ucla.edu ...!{uunet|ucla-se|turnkey}!lcc!richard