General Question about the 'Best' Platform for NONMEM

At Glaxo we are currently planning to replace our VAX 4000-90 because it is now too slow for our needs. I would be grateful of any information/experience that any of you have on other hardware platforms. We are currently looking at using a bigger, faster VAX, Alpha or UNIX box, possibly incorporating multiple processors. In particular, when purchasing your own systems did you consider different platforms? Did you discover that NONMEM didn't work well on any? Were there any platforms that lacked a suitable FORTRAN compiler? Were there any other pro's and con's that you discovered? What did you decide to buy and why? All information will be gratefully accepted and will be treated confidentially. I look forward to hearing from you. Please reply to GMS4175@GGR.CO.UK or, on CompuServe, to 100425,3600. Many thanks, Response/question from: Ed (166ETG@delphi.com) - 11 Nov 1994 Hi. We are running NONMEM on an IBM ESA 9000 under VM. Some of the scientists here (Boehringer Ingelheim, a pharmaceutical company in Connecticut) were wondering about using NONMEM on an IBM PC under Windows. Do you think this is a good idea ? Some of our jobs now on the shared IBM run 12 hours. Has anyone compared NONMEM on various platforms ? Should I consider a workstation instead ? One more thing. We have found a limit of 400 records under VM on the ESA 9000 due to the loader not being able to go above 16 meg. Does anyone have any info on number of records vs platform ? One scientist here wants to run a 600 record problem. We considered renting time on a CRAY. We here Boeing might have NONMEM on a CRAY. Does anyone know if this is true ?

I will be posting some results on NONMEM benchmarks in the next few days. It will be sent to nmusers and the sources put on the NONMEM ftp site run by Steve Shafer and Jaap Mandema. > One more thing. We have found a limit of 400 records under VM on the ESA 9000 > due to the loader not being able to go above 16 meg. Does anyone have any info > on number of records vs platform ? One scientist here wants to run a 600 record > problem. Can you be a bit more specific about what you mean by a record? The benchmark problem I have has been run on a variety of machines including PCs under Windows. It has 220 subjects and 1336 observations in total. There has never been an indication of a problem with the number of records. Are you referring to the number of records per subject? This has a default value of 50 (I think) and can be increase if necessary but it would require very unusual problems to have 600 records per subject! Can you give me (or nmusers) some better idea of what kind of problem you are attempting and what you mean by the number of records?

> One more thing. We have found a limit of 400 records under VM on the ESA 9000 > due to the loader not being able to go above 16 meg. Does anyone have any info > on number of records vs platform? > One scientist here wants to run a 600 record problem... probably you have followed the instructions in section III, 2.7 of NONMEM guide III in increasing the maximum number of observation records per individual. (When you change the default value make sure that when implementing a NONMEM update the line PARAMETER (NO= ...) again is the same in a l l of these files.) Having done this, the size of NONMEM buffers 1 and 2 (guide III, section III, 2.9) still limits the total number of records per individual to 400. You may change this limit by modifying subroutines DAT1 and DAT2. For example, the lines DIMENSION X(16040) DATA SIZE/16040/ in file DAT1.f, and DIMENSION X(2406) DATA SIZE/2406/ in file DAT2.f allow 800 data records per individual. I hit this limitation when inserting hundreds of EVID=2 lines in a data set in order to get smooth prediction curves. When your predictions can be described by an explicit function (e.g. a sum of exponentials) a better way of producing smooth predictions may be to extract the parameters of this function from NONMEM's report file and to have your graphics program produce the prediction curve. Response/question from: (gasparini@binfc1.jrf.be) - 22 Nov 1994 There has already been lots of discussions on this channel, regarding installation of NONMEM on different platforms. Until now we run NONMEM on VMS (VAX-machine) but we are considering going to UNIX, for obvious reasons (especially machine precision). I know that some of you run NONMEM on UNIX and I would like to know more about your experience on installation on specific machines (Silicon Graphics, SUN, DEC-ULTIRX....). Are there things to consider before and during installation ? Our company tends to choose an HP, but this should not be a fixed choice. Please, kindly reply to GASPARINI@JRF.BE

In reply to Gasparini's question: > > I know that some of you run NONMEM on UNIX and I would like to know more about > your experience on installation on specific machines (Silicon Graphics, SUN, > DEC-ULTIRX....). Are there things to consider before and during installation ? > Our company tends to choose an HP, but this should not be a fixed choice. I can only say that I am a satisfied HP-UX user. I have used an HP730 for 2 years and have just upgraded to an HP735/125 and HP-UX 9.05. Very reliable and fast FORTRAN has made it a good machine for NONMEM. There are a couple of small issues about the HP-UX+f77 that you might want to use - e.g. to return an exit error code from NMTRAN you need the following in CFILEX in the declaration section: $ALIAS exit="exit" (%VAL) and to flush the output buffer from ZXMIN1 in NONMEM put the following in the declaration section: INTEGER UNITID EXTERNAL FFLUSH and the following two lines after the line with the label 6011: 6011 FORMAT (' NUMSIGDIG:',10(8X,F4.1)/(11X,10(8X,F4.1))) UNITID=FSTREAM(6) CALL FFLUSH(%VAL(UNITID)) You may also wish to create a shared code library for most of the NONMEM routines so that when you run multiple instances of NONMEM executables there need be only one common copy of this code in memory. I can give you details of doing that if you want. Response/question from: (gelfring@intnet.upj.com) - 28 Nov 1994 I would be interested in finding out if anyone has done ant front-end work with NONMEM in the IBM/CMS environment that makes the access to NONMEM easier for inexperienced mainframe users.