Changeset 313 for trunk/LMDZ.MARS/README

Timestamp:

Oct 10, 2011, 12:03:17 PM (13 years ago)

Author:

acolaitis

Message:

* AC 10/10/2011 *

*
This commit aims at increasing the thermals speed. Using these corrections, gcm performances in 64x48x32 using 1 tracer goes from 27.9% elapsed time in thermals to 18.76%.

*
Additional work needs to be done in tracer advection to gain speed in high tracer number configuration. (tracer advection (but not momentum nor temperature) could be decoupled from sub-timestep, as they do not act on the thermals scheme (water vapor is neglected as we use theta and not theta_v, and radiative effect of dust is not computed in the thermals.))

*
=> TOP 5 of routine contributions to gcm runtime :

Each sample counts as 0.01 seconds.

% cumulative self self total

time seconds seconds calls s/call s/call name
18.76 6.33 6.33 960 0.01 0.01 thermcell_main_mars_
17.19 12.13 5.80 svml_powf4.A
13.72 16.76 4.63 10369 0.00 0.00 filtreg_

3.94 18.09 1.33 intel_new_memset
3.73 19.35 1.26 2880 0.00 0.00 thermcell_dqupdown_

note: thermcell_main_mars_ does call quite a lot power computations (see svml_powf4.A), but this number will not increase with tracer numbers.

*
=> LOG:

M 312 libf/phymars/thermcell_main_mars.F90
^{------------------- removed (commented) computations on buoyancy which is purely diagnostic}

tuned internal convergence loop and added convergence criterion

M 312 libf/phymars/thermcell_dqupdown.F90
^{------------------- removed (commented) downdraft-related if-loops (as we do not advect tracers and momentum in downdrafts for now)}

M 312 libf/phymars/calltherm_mars.F90
^{------------------- removed (commented) diagnostic-related computations}

changed default thermals spliting and aspect ratio
corrected a bug where maximum height was not correctly computed and could result in convective adjustment used in place of thermals
when using certains sets of nsplit and r_aspect (was not happening with the baseline version, so that this correction is transparent to
users)

File:

• trunk/LMDZ.MARS/README (modified) (1 diff)

trunk/LMDZ.MARS/README

@@ -1034 +1034 @@
 >> Bug correction in lect_start_archive.F ; in some cases layer(:) was not
    initialized.
+
+
+== 10/10/2011 == AC
+
+***********
+This commit aims at increasing the thermals speed. Using these corrections, gcm performances in 64x48x32 using 1 tracer goes from 27.9% elapsed time in thermals to 18.76%.
+
+***********
+Additional work needs to be done in tracer advection to gain speed in high tracer number configuration. (tracer advection (but not momentum nor temperature) could be decoupled from sub-timestep, as they do not act on the thermals scheme (water vapor is neglected as we use theta and not theta_v, and radiative effect of dust is not computed in the thermals.))
+
+***********
+=> TOP 5 of routine contributions to gcm runtime :
+
+Each sample counts as 0.01 seconds.
+  %   cumulative   self              self     total
+ time   seconds   seconds    calls   s/call   s/call  name
+ 18.76      6.33     6.33      960     0.01     0.01  thermcell_main_mars_
+ 17.19     12.13     5.80                             __svml_powf4.A
+ 13.72     16.76     4.63    10369     0.00     0.00  filtreg_
+  3.94     18.09     1.33                             __intel_new_memset
+  3.73     19.35     1.26     2880     0.00     0.00  thermcell_dqupdown_
+
+note: thermcell_main_mars_ does call quite a lot power computations (see __svml_powf4.A), but this number will not increase with tracer numbers.
+
+***********
+=> LOG:
+
+M             312   libf/phymars/thermcell_main_mars.F90
+^------------------- removed (commented) computations on buoyancy which is purely diagnostic
+                     tuned internal convergence loop and added convergence criterion
+
+M             312   libf/phymars/thermcell_dqupdown.F90
+^------------------- removed (commented) downdraft-related if-loops (as we do not advect tracers and momentum in downdrafts for now)
+
+M             312   libf/phymars/calltherm_mars.F90
+^------------------- removed (commented) diagnostic-related computations
+                     changed default thermals spliting and aspect ratio
+                     corrected a bug where maximum height was not correctly computed and could result in convective adjustment used in place of thermals
+                     when using certains sets of nsplit and r_aspect (was not happening with the baseline version, so that this correction is transparent to
+                     users)
+********************
+