Performance measurement of numerical calculation

Wed Sep 19 01:16:46 AEST 1990

  In my measurement, it is worth to say that RISC/6000 is fast in the 
numerical calculations by the fortran program. I have also measured using 
various fortran programs (which are written not for the benchmark test, 
but for the nuclear physics research) and my results are shown below.  For 
other programs, my feeling of the relative performance of the numerical 
calculation of RISC/6000 is very close to the value of the average 
performance listed in table 2. I can state that is "next to" 
the (scaler-type) Maim Frame computer.

   It is also true that we have encountered a lot of problems in the 
installation of the Power Station. Most of them are in the network
connection to the non-IBM machine using TCP/IP and UUCP. Even in the
telnet we have encountered many problems !  Although most of the problems 
have been resolved so far, some problems are still not resolved. (eg. 
Aixterm to VAX-3100 of my previous posting is still not resolved. 
>  aixterm: BadValue, integer parameter out of range for operation
>   Request code 1, minor code 0, serial #70, resource id -2 
I'm asking for the suggestion from the specialist of the IBM side)  
I hope every problems will be fixed soon.

===========================================================================

     Performance evaluation of the ** numerical calculations ** of the
   recent RISK work stations, Main frames and VAX'es using some typical 
   Fortran programs for the nuclear physics research. 

   NB:  Following data are only some examples of the performance 
        evaluations and they strongly depend on the application programs 
        and environment (workload level, real memory size, program size..) 

	Ver 1.3	 14-Sep-1990  by Takashi Ichihara (RIKEN)

(1) Result of some bench mark tests measured at RIKEN.
--------------------------------------------------------------------------
 	        |   Linpack (*1)      Linpack (*1)    Dhrystone V1.1 (*2)
                | Single Precision  Double Precision  (MIPS) (dhrystones/s)
----------------+---------------------------------------------------------
PowerStation 530|  11.0  MFLOPS     10.8  MFLOPS      37.3 MIPS  (65600)
SG 4D/220(R3000)|   6.4  MFLOPS      3.8  MFLOPS      19.5 MIPS  (34300) 
SparcStation 330|   2.5  MFLOPS      1.47 MFLOPS      14.4 MIPS  (25300)
SparcStation 1  |   1.94 MFLOPS      1.09 MFLOPS      11.5 MIPS  (20200) 
                |                          
VAX-6410 (VMS)  |   1.68 MFLOPS      1.12 MFLOPS      10.6 MIPS  (18700)
VAX Station 3100|   0.90 MFLOPS      0.54 MFLOPS       5.1 MIPS  ( 8900) 
VAX-8250 (VMS)  |   0.27 MFLOPS      0.16 MFLOPS       1.8 MIPS  ( 3100)
Micro VAX II    |   0.17 MFLOPS      0.12 MFLOPS       1.0 MIPS  ( 1800) 
                |                       
FACOM M780 (MSP)|  19.7  MFLOPS     14.9  MFLOPS      (not measured yet)
FACOM M380 (MSP)|   8.0  MFLOPS      7.6  MFLOPS      (not measured yet)
----------------------------------------------------------------------------

(2) Result of some fortran programs of numerical calculation of nucler
    physics research measured at RIKEN.
-----------------------------------------------------------------------------
                |  CPU time in second (*3)   |   Relative performance (*4)
                |                            |                           (*5)
                |CALC-1 CALC-2 CACL-3 CALC-4 |CALC-1 CALC-2 CALC-3 CALC-4 Avr
----------------+----------------------------+-------------------------+-----
PowerStation 530|  6.98   6.63   4.00   61.8 | 63.9  83.9 138.0  51.1  |74.7
SG 4D/220(R3000)| 19.3   26.0    6.3    60.3 | 23.1  21.4  87.6  52.4  |33.7
SparcStation 330| 32.5   51.8    9.8   104.5 | 13.7  10.7  56.3  30.2  |18.7
SparcStation 1  | 48.8   87.7   14.4   176.3 |  9.2   6.3  38.3  17.9  |11.6
                |                            |                         |
VAX-6410 (VMS)  | 31.36  48.26  14.27  122.8 | 14.2  11.5  38.6  25.7  |18.2
VAX Station 3100| 62.31 105.96  29.71  254.4 |  7.2   5.2  18.5  12.4  | 8.7
VAX-8250 (VMS)  |168.77 320.33  87.75  754.4 |  2.6   1.7   6.3   4.2  | 3.0
Micro VAX II    |320.40 485.16 138.65 1049.8 |  1.4   1.1   4.0   3.0  | 1.8
                |                            |                         |
FACOM M780 (MSP)|  2.14   2.74   2.72   16.8 | 208.   203.  203.  188. | 201.
FACOM M380 (MSP)|  4.46   5.56   5.52   31.6 | 100.   100.  100.  100. | 100.
-----------------------------------------------------------------------------

 [Main part of the calculation]

  CALC-1  simple single-precision(SP) floating point calculation
  CALC-2  simple double-precision(DP) floating point calculation
  CALC-3  many subroutine calls + single-precision FP cal.
  CALC-4  many nesting conditional branches + double-precision FP cal.

  Fortran Compiler Option:

         IBM Power Station 530           xlf -O    (Optimize)
         Silicon Graphic 4d/220 (R3000)  f77 -O3   (Optimize 3)
         Sparc Station 1,330 (SUN 4)     f77 -O    (Optimize 3)
         VAX/VMS                         fort/opt  (Optimize)
         FACOM MSP (IBM/370 compatible)  FORT7CLG,OPT=3  (Optimize 3)

(*1) Linpack benchamark test for 100*100 system of linear equations.
(*2) C Compiler option in compling dhrystone V1.1 program:
         UNIX machine:  cc -O -DREG=register 
         VAX/VMS     :  cc/opt               
(*3) CPU time for executing the load module (excluding cmpiling/linking time) 
(*4) Relative performance  (FACOM M380 = 100)
(*5) Performance for the total CPU time = CPU(CAL1)+CPU(CAL2)+CPU(CAL3)+
                                          CPU(CAL4)/6

-------------------------------------------------------------------------
  Takashi Ichihara  (Ph.D)
  RIKEN (The Institute of Physical and Chemical Research)
  2-1, Hirosawa, Wako, Saitama, 351-01, Japan

  (Internet)      Ichihara at rik835.riken.go.jp
  (HEPnet/SPAN)   RIKEN::Ichihara (41911::Ichihara)
  (FAX [G3])      (81)-484-62-4641
-------------------------------------------------------------------------