Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!mailrus!uwm.edu!psuvax1!psuvm!GECRDVM1!MOORE From: MOORE@GECRDVM1.BITNET (Hugh Kelsea Moore) Newsgroups: bit.listserv.vmxa-l Subject: (no subject given) Message-ID: Date: 28 Feb 90 12:18:39 GMT Sender: VM/XA Discussion List Reply-To: VM/XA Discussion List Organization: General Electric Corporate Research & Development Lines: 39 Approved: NETNEWS@PSUVM.BITNET Gateway Comments: To: Martin Doyle In-Reply-To: Message of Mon, 26 Feb 90 13:10:02 EST from On Mon, 26 Feb 90 13:10:02 EST said: >Subject: A Tuning Experiment >One thing that I have been tinkering with is a program that is running >in a disconnected service machine that issues the indicate command >every 30 seconds and tries to see if we are constrained for either >CPU, storage, or paging (DASD paging only, I'm not looking at XSTOR >yet). If we are CPU bound, it alters the DSPBUF parameters to >discriminate against Q2 and Q3 users forcing them into E2 and E3. If >we are suffering from a storage constraint the STOREBUF parameters >are changed to force users from Q2 and Q3 into E2 and E3. If we are >paging to DASD too much, the LDUBUF parameters are altered to force >Q2 and Q3 users into E2 and E3. All parameters are altered by some >predetermined percentage and are forced to stay within a predetermined >minimum and maximum range. When we reach a point where the system is >no longer constrained for a particular resource, the SRM parameters that >control use of that resource are slowly bumped back up to their >maximum value. I think you have the start of some good ideas. You might want to add a delay after altering any of the three values so that you don't alter it again for about 7-10 minutes. It sometimes takes that long for the boat to stop rocking. You might want to add a message receiving facility for the svm to accept smsg's from you--- either to comment that the performance is good at this time or otherwise. Then you can look at the log to see what the settings were. If you save the important info about paging, storage and the number of users in various dispatch lists, you can set up a database that can will show you what srm values you have set for what conditions. If you add to that the messages you sent to the svm, you may have the makings of an expert system that 'learns'. Oh, just another idea. Try writing an exec that issues various cp or cms commands and timing them (using REXX). If the 'set' of commands completes in a specified time period, that indicates a 'good' performance window and send the message to the svm. Hope this helps. Hugh Kelsea Moore Project Analyst General Electric Corporate R&D Schenectady, New York (518) 387-6579