Improve HMC on small lattices with imaginary chemical potential (Benchmark #347)


Added by Christopher Pinke almost 7 years ago. Updated almost 7 years ago.


Status:In Progress Start date:31 Oct 2012
Priority:Normal Due date:
Assignee:Matthias Bach % Done:

30%

Category:-
Target version:-

Description

Currently, I obtain ~ 5GFLOPS in the HMC using this inputfile:

#global settings
NS=8
NT=4
use_rec12=0
use_gpu=1
prec=64

#fermion-parameters
kappa=.165
cgmax=1000
beta=5.260
#startcondition=cold
#startcondition=hot
startcondition=continue
sourcefile=conf.save
savefrequency=10
fermact=WILSON
theta_fermion_temporal=1

# enable imag chem pot
use_chem_pot_im=true
# the RW values are a mu_I = (k+1) * pi / NC / NT
# for NT = 4:
# k = 0 -> a mu_I = pi / 12
# chem_pot_im = 0.261799387799149
# k = 1 -> a mu_I = pi / 6
#chem_pot_im = 0.523598775598299

# simulate deep within the second region
chem_pot_im = 0.698131700797732

#solver-parameters
use_evenodd=yes
solver=cg

#HMC-parameters
hmcsteps=100
integrationsteps0=200
num_timescales=1
tau=1.

This is strange, especially since the dslash has been tuned for small lattices recently.

I use the following settings for the GPU (from Matthias):

export DISPLAY=:0
export GPU_MAX_HEAP_SIZE=75

# report where we are
srun hostname

# check gpu
srun aticonfig --odgc --odgt --adapter=all

# blablabla
srun aticonfig --od-enable
# modify gpu clock to factory defaults
srun aticonfig --odsc 850,1200
# check gpu again
srun aticonfig --odgc --odgt --adapter=all


conf.save - Gauge field configuration file (1.1 MB) Christopher Pinke, 31 Oct 2012 03:27 pm


Related issues

blocked by CL2QCD - Defect #350: Heatbath test fails on CPU Done 31 Oct 2012
blocked by CL2QCD - Defect #351: Opencl_Module_Hmc test fails on GPU Done 01 Nov 2012

Associated revisions

Revision e3b91606
Added by Matthias Bach almost 7 years ago

Fix error in non-eo code

refs #336
refs #347

History

Updated by Matthias Bach almost 7 years ago

  • Status changed from New to In Progress

Updated by Matthias Bach almost 7 years ago

As of 356239e6a55a57a9748067a05a7f201278df6be5 the performance of the inverter for the given problem is about 8 Gflops.

  • % Done changed from 0 to 30

Also available in: Atom PDF