;
; See also ~/timegcm/tgcm23/save/speedup.pro
;--------------------------------------------------------------------
pro addtext,text,xstart,ystart
;xstart = .5
;ystart = .9
spacedown = .03
ntext = n_elements(text)
;print,'addtext: ntext=',ntext
if (ntext eq 0) then return
for i=1,ntext do begin
  xyouts,xstart,ystart-(i-1)*spacedown,text(i-1),/normal,$
    charsize=1.2
; print,'addtext: i=',i,' text(i-1)=',text(i-1)
endfor
end
;--------------------------------------------------------------------
;
; Main:
pro perf,ps=ps
;
; Set device (ps or x)
;
if (keyword_set(ps)) then begin
  set_plot,'ps'
  device,file='perf.ps'
  ips = 1
endif else begin
  set_plot,'x'
  ips = 0
endelse
;
; Plot nprocs on x-axis, wall-clock time on y-axis:
;
;- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 
;
 spval = 9999.
 niter = 480
 envvars = [                             $ 
  "  _DSM_PLACEMENT=ROUND_ROBIN",$
  "  _DSM_VERBOSE=ON           ",$
  "  _DSM_MIGRATION=OFF        ",$
  "  _DSM_BARRIER=FOP          ",$
  "  _DSM_WAIT=SPIN            ",$
  "  MP_SUGNUMTHD=OFF          ",$
  "  MPC_GANG=OFF              ",$
  "  TRAP_FPE=UNDERFL=FLUSH_ZERO;OVERFL=MAX;",$
  "           DIVZERO=ABORT,TRACE;INVALID=ABORT,TRACE"]
;
; x-axis explanation:
 xdesc = [             $
  "OpenMP: n_threads ",$
  "MPI   : n_tasks   ",$
  "Hybrid: ntasks (nodes) x ncpus/node (threads)"]
 pmethod = [                             $ 
  "CRAY/J90   : OpenMP only",$
  "Compaq/ES40: Hybrid MPI+OpenMP",$
  "IBM/WH2    : MPI only",$
  "SGI/O2K    : OpenMP only"]
 nruns_sgi = 7
 nruns_ibm = 7
 nruns_cray = 7
 nruns_compaq = 3
;xtitle='Number of threads (SGI) or tasks (IBM)'
 xtitle='Total number of processes'
;xrange = [0,25]
 xrange = [0,30]
;
; x = number of cpus, y = wall-clock (secs)
;x_sgi = [1,     2,     4,     6,     9,     12,    18]
;y_sgi = [3971., 2441., 1332., 1071., 895.,  803.,  791.]
;x_ibm = [1,     2,     4,     6,     9,     12,    18]
;y_ibm = [3662., 2116., 1590., 1494., 1240., 1077., 962.]
;ytitle='Wall clock time (seconds)'
;yrange = [0,4000]
;
; x = number of cpus, y = average wallclock secs per step
 x_sgi  = [1,     2,     4,     6,     9,     12,    18]
 y_sgi  = [8.3,   5.0,   2.8,   2.2,   1.8,   1.7,   1.6]
 x_ibm  = [1,     2,     4,     6,     9,     12,    18]
 y_ibm  = [7.6,   4.4,   3.3,   3.1,   2.6,   2.2,   2.0]
 x_cray = [1,     2,     4,     6,     9,     12,    18]
 y_cray = [21.,   11.4,  8.2,   7.3,   3.5,   3.2,   3.1]
 x_compaq = [8,   12,  24]
 y_compaq = [4.1, 2.7, 2.0]
 ytitle='Average wallclock seconds per simulated time step'
 yrange = [0,10]
;- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 
;
chsize = 1.2
if (keyword_set(ps)) then device,font_size=10.,xsize=19.,ysize=20.
plot,x_sgi,y_sgi,yrange=yrange,xrange=xrange,title=$
  'TIEGCM (tgcm14a): (' + $
  string(format="(i3)",niter) + ' 3-minute timesteps)',charsize=chsize,$
  xcharsize=chsize, ycharsize=chsize,xtitle=xtitle,ytitle=ytitle,$
  psym=-4,max_value=spval/10.,$
  position=[.1,.25,.95,.95],/norm 
oplot,x_ibm,y_ibm,psym=-6,max_value=spval/10.
oplot,x_cray,y_cray,psym=-5,max_value=spval/10.
oplot,x_compaq,y_compaq,psym=-2,max_value=spval/10.
;oplot,[0,20],[0,20],linestyle=2
xyouts,x_sgi(nruns_sgi-1)+0.5,y_sgi(nruns_sgi-1),'SGI/O2K'
xyouts,x_ibm(nruns_ibm-1)+0.5,y_ibm(nruns_ibm-1),'IBM/WH2'
xyouts,x_cray(nruns_cray-1)+0.5,y_cray(nruns_cray-1),'CRAY/J90'
xyouts,x_compaq(nruns_compaq-1)+0.5,y_compaq(nruns_compaq-1),'Compaq/ES40'
addtext,pmethod,.5,.9
addtext,xdesc,.36,.15
;
; Close ps file:
;
if (keyword_set(ps)) then begin
  device,/close
endif
end