home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!swrinde!gatech!usenet.ins.cwru.edu!agate!ucbvax!lrw.com!leichter
- From: leichter@lrw.com (Jerry Leichter)
- Newsgroups: comp.os.vms
- Subject: RE: Pagefile on DECram ramdisk - any comments?
- Message-ID: <9212311339.AA11687@uu3.psi.com>
- Date: 31 Dec 92 12:38:55 GMT
- Sender: daemon@ucbvax.BERKELEY.EDU
- Distribution: world
- Organization: The Internet
- Lines: 77
-
-
- We are currently running a VAX 4000-300 with VMS 5.4-3. We have 128Mb
- of memory. We currently only have one pagefile (50,000 blocks) and
- yes, it's on the system disk (RF31). Our other two disk drives
- (RF72's) are currently being hit so hard that a secondary pagefile on
- those disks is not possible. I have proposed buying an RF35, but that
- will not happen until the Fall of 1993.
-
- We currently have a lot of memory that is not being used (170,000+
- blocks). Working set sizes have been tuned so folks are getting the
- memory they deserve. I have been exploring the possibility of
- installing a secondary pagefile on to a DECram ramdisk of say,
- 60,000-70,000 blocks.
-
- Has anyone ever done this before and if so, are there any issues that
- I should be aware of? If I make the pagefile on the ramdisk
- sufficiently large so that it has the most Reservable pages available,
- I should be able to ensure that the secondary pagefile is used more
- often than not. We're running DECram v1.0, and I'm getting ready to
- install v1.1 this week.
-
- Think about what you are proposing: Pages go to the pagefile when there is
- no room for them (for whatever reason) in memory. You propose to have them
- go to - well, memory, when there is no room in memory.
-
- While this will work, it makes very little sense. If, as you are currently
- configured, effectively all of your physical memory were in use most of the
- time, the effect of installing a ramdisk would be to make your run SHORT of
- memory, thus increasing paging. And to what end? The pages newly forced out
- of memory would actually stay in memory - but they'd be shuffled around from
- memory marked as "memory" to memory marked as "ramdisk", costing you CPU time.
-
- Since your current configuration leaves a significant - in fact, huge! -
- amount of memory free, you can look for other ways to make use of it. On a
- system-wide basis, you might consider decreasing FREELIM/FREEGOAL so as to
- encourage pages to stay on the look-aside lists, getting pretty much the same
- effect as "paging" them to a ramdisk for a LOT less overhead. You might
- consider installing /RESIDENT your most commonly used images - both the VMS
- images and any 3rd-party or locally created, commonly-used images.
-
- However, I suspect the real clue to what's going on here is in your claim that
- "Working set sizes have been tuned so folks are getting the memory they
- deserve". When I hear a system manager talk about what users "deserve", it
- usually implies a rather, ahem, limited view of user/system management inter-
- action. It sounds as if you are letting some arbitrary, probably decade-old
- notion of the "right" amount of memory get in the way of sound system manage-
- ment.
-
- If your system HAS the memory, it's foolish not to let processes use it! I'd
- STRONGLY suggest you first re-examine your policies and standards in this
- direction to see if they make any sense with the equipment and load you have
- TODAY - not whether they made sense when you had 50 people on an 11/780.
- Then set limits that will make effective use of your resources. As a quick
- calculation: Let M be your estimate - it doesn't have to be exact - of the
- maximum number of simultaneous active users your system has to deal with. If
- you have some users who sit at a prompt for long periods of time, don't count
- them against M - or count them as a fraction of a user. (Be generous with M;
- it's better to estimate somewhat too high than somewhat too low.)
-
- Now let P be the number of pages on your system, not counting those reserved
- to VMS; again, this doesn't have to be exact. Then P/(1.5*M) should be a
- reasonable value for the WSQUOTA of a typical user. WSEXTENT should be
- considerably larger - as a first cut, I'd make it 4 times WSQUOTA. (If you
- have distinctly different classes of users, you can of course vary these
- numbers a bit - though typically I'd no more than halve or triple WSQUOTA.
- You can be much more generous with WSEXTENT; there's essentially no reason NOT
- to be generous with it.)
-
- Finally, make sure WSMAX is large enough - at least as large as the biggest
- WSEXTENT, ignoring any WSEXTENT's that you might set to "infinite" (i.e., let
- this user take all the memory he needs, when memory is available). Also, if
- you've modified FREELIM/FREEGOAL, make sure BORROWLIM is set rationally so
- that the WSEXTENT values can contribute something. AUTOGEN will do this for
- you.
-
- -- Jerry
-
-