[docs]def reorganize_resources(mpi_new, nodes, mpi_per_node, threads):
nodes = mpi_new/mpi_per_node
resources = {'num_machines': int(nodes),
'num_mpiprocs_per_machine': int(mpi_per_node),
'num_cores_per_mpiproc': int(threads)}
return resources
[docs]def find_commensurate(a,b):
for j in range(a,0,-1):
if a%b==0:
return a
else:
a -= 1
[docs]def balance(tasks,a,b,rec=0):
r = a/b
#if r > 1: r=1/r
t = int(tasks*r)
if t == 0: t = 1
for i in range(t,0,-1):
if tasks%i == 0:
t = i
break
k = tasks//t
if rec > 0: a, b=a*2, b*2
if t > a:
rec +=1
t,k = balance(t,a/2,b,rec)
elif k > b:
rec +=1
t,k = balance(t,a,b/2,rec)
return t,k
[docs]def distribute(tasks=10, what='DIP', **ROLEs):
ROLEs_DIP = {'k','c','v'}
ROLEs_X = {'k','c','v','g','q'}
ROLEs_SE = {'q','qp','b','g'}
#for role in ROLEs.keys():
# print(role,ROLEs[role],ROLEs[role]%tasks,ROLEs[role]//tasks)
if what=='DIP':
if tasks > ROLEs['c']*ROLEs['k']*ROLEs['v']:
print('set')
tasks = ROLEs['c']*ROLEs['k']*ROLEs['v']
b,k = balance(tasks=tasks,a=ROLEs['c']+ROLEs['v'],b=ROLEs['k'])
c,v = balance(tasks=b,a=ROLEs['c'],b=ROLEs['v'])
print('mpi k c v')
return k*c*v,k,c,v
elif what == 'X':
if tasks > ROLEs['c']*ROLEs['k']*ROLEs['v']*ROLEs['g']*ROLEs['q']:
tasks = ROLEs['c']*ROLEs['k']*ROLEs['v']*ROLEs['g']*ROLEs['q']
b,k = balance(tasks=tasks,a=ROLEs['c']+ROLEs['v'],b=ROLEs['k'])
c,v = balance(tasks=b,a=ROLEs['c'],b=ROLEs['v'])
if ROLEs['g']> 1:
c,g = balance(tasks=c,a=ROLEs['c'],b=ROLEs['g'])
else:
g = 1
q = 1
print('mpi q k g c v')
return q*k*g*c*v,q,k,g,c,v
elif what == 'SE':
if tasks > ROLEs['q']*ROLEs['qp']*ROLEs['b']*ROLEs['g']:
tasks = ROLEs['q']*ROLEs['qp']*ROLEs['b']*ROLEs['g']
b,qp = balance(tasks=tasks,a=ROLEs['b'],b=ROLEs['qp'])
print(b,qp)
if ROLEs['g']> 1:
b,g = balance(tasks=b,a=ROLEs['b'],b=ROLEs['g'])
else:
g = 1
q = 1
print('mpi q qp b g')
return q*qp*b*g,q,qp,b,g
[docs]def find_parallelism_qp(nodes, mpi_per_node, threads, bands, occupied=2, qp_corrected=2, kpoints = 1, \
last_qp = 2, namelist = {}):
mpi = nodes*mpi_per_node
# GW #
q = 1
g = 1
k = 1
c = 1
v = 1
b = 1
qp = 1
g_vecs = 1
k_dip, c_dip, v_dip = 1 , 1 , 1
k_X, c_X, v_X, g_X, q_X = 1 , 1 , 1 , 1, 1
qp_SE, b_SE, g_SE, q_SE =1 , 1 , 1 , 1
if 'HF_issue' in []:
if last_qp <= occupied:
print('last_qp lower than occupied, setting to occupied')
last_qp = occupied+1
bands = last_qp*qp_corrected
for i in range(10):
mpi1, k_dip, c_dip, v_dip = distribute(tasks=mpi, what='DIP', c=bands-occupied, v=occupied, k=kpoints, g=g_vecs)
mpi2, k_X, c_X, v_X, g_X, q_X = distribute(tasks=mpi1, what='X', c=bands-occupied, v=occupied, k=kpoints, g=g_vecs,q=q)
mpi3, qp_SE, b_SE, g_SE, q_SE = distribute(tasks=mpi1, what='SE', b=bands, qp=qp, g=g_vecs,q=q)
if mpi1 == mpi2 and mpi2 == mpi3:
if mpi1%mpi_per_node==0:
mpi = mpi1
break
else:
mpi=find_commensurate(mpi1,mpi_per_node)
else:
mpi = mpi3
for i in [mpi, k, c, v, g, q, qp, b]:
if i == 0 :
i = 1
mpi_DIP = {'k':k_dip,'c':c_dip,'v':v_dip}
mpi_X = {'q':q_X,'k':k_X,'g':g_X,'c':c_X,'v':v_X}
mpi_SE = {'q':q_SE,'qp':qp_SE,'b':b_SE,'g':g_SE}
parallelism = {}
parallelism['X_CPU'] = ''
parallelism['X_ROLEs'] = ''
parallelism['SE_CPU'] = ''
parallelism['SE_ROLEs'] = ''
parallelism['DIP_CPU'] = ''
parallelism['DIP_ROLEs'] = ''
for key in mpi_X:
parallelism['X_CPU'] = parallelism['X_CPU']+str(int(mpi_X[key]))+' '
parallelism['X_ROLEs'] = parallelism['X_ROLEs']+key+' '
for key in mpi_SE:
parallelism['SE_CPU'] = parallelism['SE_CPU']+str(int(mpi_SE[key]))+' '
parallelism['SE_ROLEs'] = parallelism['SE_ROLEs']+key+' '
for key in mpi_DIP:
parallelism['DIP_CPU'] = parallelism['DIP_CPU']+str(int(mpi_DIP[key]))+' '
parallelism['DIP_ROLEs'] = parallelism['DIP_ROLEs']+key+' '
resources = reorganize_resources(mpi, nodes, mpi_per_node, threads)
return parallelism, resources
