home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!haven.umd.edu!darwin.sura.net!gatech!ukma!cs.widener.edu!iggy.GW.Vitalink.COM!pacbell.com!network.ucsd.edu!mvb.saic.com!macro32
- From: <brydon@dsny25.sinet.slb.com>
- Newsgroups: vmsnet.internals
- Subject: RE: VMS tuning for a vaxcluster (LAVC)
- Message-ID: <9208281401.AA18734@sndsu1.sinet.slb.com>
- Date: Fri, 28 Aug 92 14:01:37 GMT
- Organization: Macro32<==>Vmsnet.Internals Gateway
- X-Gateway-Source-Info: Mailing List
- Lines: 84
-
- Somebody (looks like gjc@mitech.com) says:
- >This talk about FPL etc has reminded me to ask a question
- >about VAXCLUSTER (LAVC) behavior.
- >
- >The behavior of the VMS memory management with respect to the FPL
- >and installed images is said to reduce the amount of disk activity
- >required when activating commonly used images, as long
- >as the amount of free pages available allows pages from your favorite image
- >to remain in the FPL.
- >
- >Correct?
- >
- >Now. My question is, what good does this do in a VAXCLUSTER environment,
- >specifically a LAVC environment? Here a page fault consists
- >of both a trip over the ethernet and a trip to disk.
- >
- >Obviously if a satelite has the pages of an image already in memory,
- >then you will winning big, avoiding a trip across the ethernet and everything.
- >
- >But what about the trip to disk? Is there a way to avoid that?
-
- I'm having a hard time understanding your question. Please elaborate...
-
- >The LAVC server node may have plenty of free memory, but it may have *never*
- >activated any of the images that people on the LAVC's are interested in.
- >
- >Even if it had activated those images, would there be any optimization
- >available? (That is, does the vaxcluster lavc disk server -know- that
- >pages from a disk file may already be available in memory?)
- >
- >My current situation is that I have a VAXSTATION-3200 with plenty
- >of free memory serving as a LAVC boot node, and I'm interested in
- >any hacks available to speed things up.
-
- Increasing the size of the file system caches on the 3200 would probably do
- something for you. Sysgen parameters ACP_*.
-
- >In VMS 5.4-3 I was using the hack of having all images from SYS$SYSTEM
- >and SYS$SHARE also available on the local hard disk. That is a bit hackish,
- >so I didn't bother re-doing that for VMS 5.5.
-
- I'll bet that your V5.4-3 optimization was actually slowing things down. My
- unproven contention is that paging (and image activation) through ethernet is
- actually faster than to a local disk (assumption: disks on a local workstation
- are slower than on the boot server).
-
- Really the only things that the NI port driver does is the same things that
- the CI port driver does. And all that CI driver does (for this purpose) is
- talk to an HSC-xx device through the CI. When the CPU makes a request for
- disk information, the HSC is really a disk block server, not a file server.
- The logic on a boot server turns it into a file server. The major 'smarts'
- are in VMS, not the HSC. MSCP is some of these 'smarts'. Seems to me that
- paging from a satellite shows up on the boot node as [MSCP?] requests for
- particular disk blocks, not as requests for file 'x' in directory 'y'. I
- don't think the NI port driver on the boot node does any translation (Can
- somebody with source listings confirm or deny this?). But I believe if the
- boot node gets 2 subsequent requests for block 'n', it can pull the second one
- from cache. Correct?
-
- Disk access times (read/write) are measured in milliseconds. Ethernet
- transactions are on a microsecond level. When a satellite pages onto its
- local disk, it is doing I/O's of probably 20-40 milliseconds average (depends
- on a number of things). When a satellite pages through ethernet to another
- system, the data is seen as MSCP info (?) on the boot node and (1) write times
- are essentially the speed of the boot server disk, (2) reads are considered
- user data on the boot server and are cached. Therefore a page read request
- from a workstation for data that is cached on the boot node could be on the
- order of several hundred microseconds, significantly faster than a disk access
- to the local disk.
-
- My claim is that paging through [an unloaded] ethernet is faster than to a
- local disk. Of course if a lot of systems page this way, you will have a
- network contention problem which will slow things down. [By the way, this
- assumes paging to another VMS system. Paging to an Infoserver, as per the
- VXT2000 workstation is essentially the speed of the disk on the infoserver,
- plus you are generating ethernet traffic. There is no caching on an
- Infoserver.]
-
- Any disagreements with the above? Anybody do any quantitative testing on
- this?
- _______________________________________________________________
- Harvey Brydon | Internet: brydon@dsn.SINet.slb.com
- Dowell Schlumberger | P.O.T.S.: (918)250-4312
- I bought my VAXstation from my uncle on his deathbead - I wrote him a check.
-