[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference nicctr::kap-users

Title:Kuck Associates Preprocessor Users
Notice:KAP V2.1 (f90,f77,C) SSB-kits - see note 2
Moderator:HPCGRP::DEGREGORY
Created:Fri Nov 22 1991
Last Modified:Fri Jun 06 1997
Last Successful Update:Fri Jun 06 1997
Number of topics:390
Total number of notes:1440

373.0. "Benchmark two times slower with KAP Parallelization" by TBC001::WONG () Mon Mar 24 1997 21:28

I'm having serrious difficulties with Parallelization on a customer benchmark.
The Fortran program supplied by the customer runs twice as slow using KF77 with
-conc and two processors.  It's even worse with KF90 (three times slower).  
Unless we can come up with better results soon, the sales opportunity is in jeopardy.

I'm using the latest SSB versions of compilers on UNIX 3.2G.  The customer code
is too big to be shown here.  It will be most appreciated if someone can telnet to
my machine and take a look.

Telnet Instructions:

	>telnet 16.155.48.56   (or set host to TBC001 via DECnet)
	username: guest2
	password: guest210
	$set host/dte txa10   (if txa10 is not available, try txa11 and up to txa15)
	local> conn a8400a
	login: wong
	password: jw

Program Instructions:
	a8400a> pe    		(this will cd to the program directory 'em')
	a8400a> make -f pmakefile	(makes parallel version. The single processor
					version that runs faster is created with
					the make file 'makefile')
	a8400a> cd ps
	a8400a> nohup time pspml.x < ifdem &

T.RTitleUserPersonal
Name
DateLines
373.1More Info on .0TBC001::WONGMon Mar 24 1997 21:4229
    More information on .0:
    
    The single processor run completed on a AS4100 466MHz system in 414.4
    sec.  The parallel version with PARALLEL=2 completed in 849.4 sec.
    
    The same program ran with different input file (ifdem in 'em/fd'
    instead of ifdem in 'ps/fd') parallelized well.  Single processor run
    was 600 sec whereas dual processor run was 314.5 sec.
    
    Compiler versions were:
    DFABASE410      installed       Digital Fortran 77 V4.1 for Digital
    					UNIX Alpha Systems
    DFACOM410       installed       Digital Fortran V4.1 Tools & their Man
    					Pages
    DFADOC410       installed       Digital Fortran 77 V4.1 Release Notes
    					and Man Page
    DFARTL373                       Digital Fortran RTL #373 for Digital
    					UNIX Alpha (f77 only)
    DFARTL374       installed       Digital Fortran RTL #374 for Digital
    					UNIX Alpha (f90 and f77)
    F9ABASE410      installed       Digital Fortran 90 V4.1 for Digital
    					UNIX Systems
    F9ADOC410       installed       Digital Fortran 90 V4.1 for Digital
    					UNIX Release Notes and Man Page
    KF9BASE310      installed       KAP Fortran 90 for Digital Unix
    					Optimizer
    KPFBASE310      installed       KAP for DEC Fortran for Digital Unix
    					Optimizer
    
373.2observationsMSBCS::SCHNEIDERSay it with ASCIITue Mar 25 1997 12:4814
    I will offer a couple of observations, for what they're worth.
    
    - I personally find working on TBC systems via that DTE path to be
      unwieldy to the point of impossibility.  You might want to package
      up the materials and provide a pointer so that any volunteers can
      work locally.  I see there are only a couple of MB worth of files
      there.
    
    - I see source file names that indicate that FFTs are involved.  If the
      FFTs are the bottleneck, it's likely that the right thing to to is
      use DXML.  Parallel FFTs are available in DXML.
    
    Cheers,
    Chuck Schneider
373.3Response to .2TBC001::WONGTue Mar 25 1997 14:3311
    ref .2
    
    I must apologize for the slowness of the connection.  However, I cannot
    provide a copy of the code on the network because the customer is
    super-sensitive about it.
    
    We tried using DXML before.  It didn't improve performance and it
    created different results.  Seems like the customer's routines are
    non-standard.
    
    Thanks for your comments anyway.
373.4password in notesfile is a no-noPERFOM::HENNINGWed Mar 26 1997 17:478
    >  I cannot
    >   provide a copy of the code on the network because the customer is
    >   super-sensitive about it.
    
    Then for goodness sakes what be ye doing POSTING PASSWORDS in an
    unrestricted notesfile??????????????????????????????????????????
    
    
373.5Response to .4TBC001::WONGThu Mar 27 1997 23:4518
    ref .4
        
        Mr. Hanning,
        
        Seems like you are even more sensitive than my customer!  You must
    have
        had some very bad experiences.  Thanks for your concern anyway. 
    The
        customer files are on another system which can only be reached via
    reverse
        lat.  That makes it difficult (not impossible) to copy them.  It's
    a
        compromise.  Anyway, the account just expired today.  Don't lose
    sleep
        over it!
        
        Joe
    
373.6apology for too many ????PERFOM::HENNINGFri Mar 28 1997 19:453
    sorry, I didn't mean to get so excited.
    
    	/john