IRIX certification, optimization

Compiling and building our code

What is displayed here are results obtained before June 3rd 1997. Code was compiled against the development libraries.

From now on code will be built against the release libraries.

Very soon the results in this page will be moved to a different page.

As for now, only a simple comparison has been made between the 2 libraries.
A release executable has been run and compared with a development executable run on the same data file. As could be expected (hoped?) there is NO DIFFERENCE (in terms of a Unix diff) between the oddpack outputs.
There is a slight increase in the execution time (15 ms per event), but this needs more investigation.

Where and how the code is run

As for now, the code is run in batch on any IRIX node of the batch system.
Those nodes are single processor 150 MHz SGI R4400Cr's , running IRIX 5.3.
Now, the entire file 10106.02 is analyzed (same as for the AIX certification), for a total of 38212 events.
The same namelist as for the AIX certification is used, i.e.:

 rawall = -1
 recall = -2

Furthermore, the code is compiled using the -O option.

Code stability

As for now, I'm still playing with the profiles...

Optimizations

Baseline profile (50 routines) as obtained from the prof analyzer run in pc-sampling (i.e. the program is interupted every 10ms, recording the value of the program counter)




   time     (%)       cum time          procedure (file) 
   
 2.4e+03s  (  9.6)   2.4e+03s	         spacer_ 
 2.1e+03s  (  8.6)   4.5e+03s        	 lsqall_ 
 2.1e+03s  (  8.4)   6.6e+03s    	mat2prj_ 
 1.9e+03s  (  7.8)   8.6e+03s    	 __sqrtf 
 1.4e+03s  (  5.7)     1e+04s    	   lsq1_ 
 1.3e+03s  (  5.2)   1.1e+04s	          fitsp_ 
 1.2e+03s  (  4.9)   1.2e+04s	         duesta_ 
 9.3e+02s  (  3.7)   1.3e+04s	        mat3shr_ 
 7.2e+02s  (  2.9)   1.4e+04s	 	  nuova_ 
 5.8e+02s  (  2.3)   1.5e+04s		 ycorr2_ 
 5.2e+02s  (  2.1)   1.5e+04s		 tsvmat_ 
 5.2e+02s  (  2.1)   1.6e+04s		   p0p1_ 
 5.1e+02s  (  2.0)   1.6e+04s		  ykick_ 
 3.3e+02s  (  1.3)   1.7e+04s		   eion_ 
 2.8e+02s  (  1.1)   1.7e+04s	 	   saga_ 
 2.8e+02s  (  1.1)   1.7e+04s		  vzero_ 
 2.7e+02s  (  1.1)   1.7e+04s	 	   __pow 
 2.3e+02s  (  0.9)   1.8e+04s		 ftrace_ 
 2.3e+02s  (  0.9)   1.8e+04s		  fitvw_ 
 2.2e+02s  (  0.9)   1.8e+04s		  zeroe_ 
   2e+02s  (  0.8)   1.8e+04s	 	 xcorr2_ 
 1.8e+02s  (  0.7)   1.8e+04s		  __sqrt 
 1.6e+02s  (  0.7)   1.9e+04s		 ie_pwc_ 
 1.6e+02s  (  0.6)   1.9e+04s		  ycorr_ 
 1.4e+02s  (  0.6)   1.9e+04s          upkpwc_c_ 
 1.4e+02s  (  0.6)   1.9e+04s	  	   newt_ 
 1.3e+02s  (  0.5)   1.9e+04s     omu_intersect_ 
 1.2e+02s  (  0.5)   1.9e+04s		   hist_ 
 1.2e+02s  (  0.5)   1.9e+04s		put2prj_ 
 1.2e+02s  (  0.5)     2e+04s		  pkick_ 
 1.1e+02s  (  0.4)     2e+04s		 matinv_ 
 1.1e+02s  (  0.4)     2e+04s		 linchk_ 
 1.1e+02s  (  0.4)     2e+04s		 xyprim_ 
   1e+02s  (  0.4)     2e+04s	 	 dvdcov_ 
   1e+02s  (  0.4)     2e+04s    	mat3prj_ 
   1e+02s  (  0.4)     2e+04s            emunpk_ 
      97s  (  0.4)     2e+04s	           __log 
      92s  (  0.4)     2e+04s             __logf 
      91s  (  0.4)     2e+04s	         lsqmag_ 
      87s  (  0.3)   2.1e+04s	 	   gexy_ 
      83s  (  0.3)   2.1e+04s	 	 scruta_ 
      81s  (  0.3)   2.1e+04s	 	   plot_ 
      78s  (  0.3)   2.1e+04s	 	  xcorr_ 
      72s  (  0.3)   2.1e+04s	 	   flvp_ 
      71s  (  0.3)   2.1e+04s	 	  xkick_ 
      71s  (  0.3)   2.1e+04s		 cospdb_ 
      70s  (  0.3)   2.1e+04s		  fit3s_ 
      69s  (  0.3)   2.1e+04s		   fit5_ 
      68s  (  0.3)   2.1e+04s	 	 cluste_ 
      67s  (  0.3)   2.1e+04s	 	 pbgcls_

***					       ***
***	TOTAL EVENTS  =     38212	       ***
***	  TOTAL TIME  =   24343.5     secs     ***
***   TIME PER EVENT  =   637.064     msecs    ***
***					       ***

The same executable has been run on a different data file (9954.02) and execution time is approximately the same:


***					       ***
***	TOTAL EVENTS  =     38095	       ***
***	  TOTAL TIME  =   22111.9     secs     ***
***   TIME PER EVENT  =   580.441     msecs    ***
***					       ***

When comparing the profiler output, altough the order is different, one finds that approximately the same 10 routines are using most of the CPU:


 	2.1e+03s  (  9.6)    2.1e+03s	     lsqall_ 
 	2.1e+03s  (  9.3)    4.2e+03s	    mat2prj_ 
 	1.7e+03s  (  7.6)    5.8e+03s	     __sqrtf 
 	1.5e+03s  (  6.9)    7.4e+03s	     spacer_ 
 	1.4e+03s  (  6.2)    8.7e+03s	       lsq1_ 
	1.2e+03s  (  5.4)    9.9e+03s	     duesta_ 
	1.1e+03s  (  4.9)    1.1e+04s	      nuova_ 
	9.8e+02s  (  4.4)    1.2e+04s	      fitsp_ 
	9.7e+02s  (  4.4)    1.3e+04s	    mat3shr_ 
 	5.1e+02s  (  2.3)    1.3e+04s	     ycorr2_

Just for fun you can compare the profile from a different namelist

Here you can find the talk I gave about MINI-PASS1

O2 versus O1 studies fro MICRORICO

Send comments to: Matteo Boschini.
Last update: 04-June-1997