Yorick Performance

The table below reports results of a serie of tests that compare the performance, in term of speed, of Yorick (1.6) versus IDL (6.0). The tests were performed on 2 computers, a G5 (dual 2GHz G5 processors, 3.5GB RAM, running MacOS 10.3.5) and a Linux Box (dual 2.8GHz Ahtlon processor, 2GB RAM, running Redhat 9 (the compilation flags are listed here).

These tests are rather arbitrary. I had to keep to relatively simple tests (basically single commands) to avoid comparing implementation rather than intrinsic speed. The selection of subject/command certainly reflects my own interests, but it's a start. This kind of tests are always difficult to make, so just take these with a grain of salt. I have not reported the results of graphic tests in this table (I did some, which also showed comparable performance). All these tests were performed while there was no significant other load on the computers.

Overall, Yorick does well compared to IDL. On the Linux box, Yorick (again, for this limited and selective serie of test) was 21% than IDL. On the G5, Yorick was 35% faster than IDL! Last data point: Yorick runs 35% faster on the G5 than on the Linux Box onto which these tests were made.

All results reported in the table below are in seconds. Results were averaged over 10 instances of each operation. The average and rms is given. Click on the test label on the left to see the code. Most of these operations were made on 1024x1024 arrays (see code for details).
 

	G5		Linux Box
Test	Yorick	IDL	Yorick	IDL
Memory allocation	1.24+/-0.05	1.27+/-0.06	1.338+/-0.005	1.33+/-0.01
Array Operations	0.0207+/-0.0030	0.0183+/-0.0025	0.0265+/-0.0001	0.0191+/-0.0001
FFT, stock	0.65+/-0.02	0.90+/-0.04	0.75+/-0.07	1.21+/-0.01
Optimized FFT	0.245+/-0.027	-	-	-
Matrix multiply	2.47+/-0.04	2.48+/-0.05	9.47+/-0.02	5.15+/-0.01
SVD decomposition	0.62+/-0.04	1.42+/-0.05	1.137+/-0.001	1.971+/-0.003
Matrix Transpose	0.170+/-0.018	0.131+/-0.005	0.0730+/-0.0001	0.245+/-0.002
Poidev	1.23+/-0.03	4.25+/-0.17	1.208+/-0.002	3.47+/-0.09

Compilation:

Compilation flags for yorick (obviously, I don't have access to the IDL compilation flags as only the binaries are distributed):

• Linux box: -O3
• G5: -O3 -fomit-frame-pointer -fstrict-aliasing -mcpu=970 -mtune=970 -mpowerpc-gpopt

Full tests commands

Memory allocation

Just a malloc() here, so no real wonder the 2 codes are so similar in performance.

• Yorick

  etime=[]; 
  for (i=1;i<=10;i++) {
    tic;
    im=array(float,[3,1024,1024,100]);
    grow,etime,tac();
  }
  etime(avg); etime(rms)
  

• IDL

  etime=fltarr(10)
  for i=0,9 do begin 
    tic=systime(1)
    im=fltarr(1024,1024,100)
    etime(i)=systime(1)-tic
  endfor
  print,avg(etime),stdev(etime) 
  

Array Operations

• Yorick

  im1=im2=array(float,[2,1024,1024]); 
  etime=[]; 
  for (i=1;i<=10;i++) {
    tic;
    im=im1*im2;
    grow,etime,tac();
  }
  etime(avg); etime(rms)
  

• IDL

  im1=fltarr(1024,1024)
  im2=im1
  etime=fltarr(10)
  for i=0,9 do begin 
    tic=systime(1)
    im =im1*im2
    etime(i)=systime(1)-tic
  endfor
  print,avg(etime),stdev(etime)
  

Stock FFTs

Note that the yorick FFT operator only processed double (the yorick complex type is made of 2 double).

• Yorick

  im=array(complex,[2,1024,1024]) // each plan is of type double
  etime=[]
  for (i=1;i<=10;i++) {
    tic;
    out=fft(im,1); 
    grow,etime,tac();
  }
  etime(avg); etime(rms)
  

• IDL

  im1=fltarr(1024,1024)
  im=complex(im1,im1)
  etime=fltarr(10)
  for i=0,9 do begin 
    tic=systime(1)
    out=fft(im,1)
    etime(i)=systime(1)-tic
  endfor
  print,avg(etime),stdev(etime)
  

Optimized FFT

This one use a special yorick plugin that uses the apple vectorial libraries (thus only available on apple computers). Listed for information

• Yorick

  im1=im2=array(float,[2,1024,1024]); 
  etime=[]; 
  for (i=1;i<=10;i++) {
    tic;
    out=fftVE(im1,im2,1); 
    grow,etime,tac();}
  etime(avg); etime(rms)
  

Matrix multiply

I don't know why yorick is so much behind on the linux box for this test.

• Yorick

  ar1=ar2=array(float,[2,1024,1024]); 
  etime=[]; 
  for (i=1;i<=10;i++) {
    tic;
    ar=ar1(,+)*ar2(+,); 
    grow,etime,tac();
  }
  etime(avg); etime(rms)
  

• IDL

  ar1=fltarr(1024,1024)
  ar2=ar1
  etime=fltarr(10)
  for i=0,9 do begin 
    tic=systime(1)
    ar=ar1#ar2
    etime(i)=systime(1)-tic
  endfor
  print,avg(etime),stdev(etime)
  

SVD decomposition

• Yorick

  a=unit(500); 
  etime=[]; 
  for (i=1;i<=10;i++) {
    tic; 
    w=SVdec(a,u,vt); 
    grow,etime,tac();
  }
  etime(avg); etime(rms)
  

• IDL

  a=identity(500)
  etime=fltarr(10)
  for i=0,9 do begin 
    tic=systime(1)
    svdc,a,w,u,v
    etime(i)=systime(1)-tic
  endfor
  print,avg(etime),stdev(etime)
  

Matrix Transpose

• Yorick

  ar=array(float,[2,1024,1024]); 
  etime=[]; 
  for (i=1;i<=10;i++) {
    tic;
    ar1=transpose(ar); 
    grow,etime,tac();
  }
  etime(avg); etime(rms)
  

• IDL

  ar=fltarr(1024,1024)
  for i=0,9 do begin 
    tic=systime(1)
    ar1=transpose(ar)
    etime(i)=systime(1)-tic
  endfor
  print,avg(etime),stdev(etime)
  

Poidev

Poidev is not in the standard yorick distribution. It belongs to a plugin (yao).

• Yorick:
  

  etime=[]; 
  for (i=1;i<=10;i++) {
    tic;
    im=poidev(array(100,1024*1024));
    grow,etime,tac();
  }
  etime(avg); etime(rms)
  

• IDL:
  

  etime=fltarr(10)
  for i=0,9 do begin 
    tic=systime(1)
    im=poidev(replicate(100,1024*1024l))
    etime(i)=systime(1)-tic
  endfor
  print,avg(etime),stdev(etime) 
  

Page written by Francois Rigaut, 2004sep15.

Last modified: Mon Aug 01, 2005, 00:46:51 HST