Hi Åsmund,
The reason is that these two have very different implementations:
1. nmropt: the restraint is done on the restraint working unit running on GPU, similar to regular bonded terms.
2. nfe: the coordinates are downloaded from CPU first. CPU does the restraint part and send the updated forces back to GPU. This leads to not only slow calculations on CPU but also, more critical, extra expensive communication between CPU and GPU.
If you just need simple umbrella sampling, I would suggest nmrop.
Best,
Taisung
-----Original Message-----
From: Åsmund Røhr Kjendseth [mailto:asmund.rohr.nmbu.no]
Sent: Tuesday, November 5, 2019 4:13 AM
To: AMBER Mailing List <amber.ambermd.org>
Subject: [AMBER] Large GPU performance difference for umbrella sampling methods: (infe=1 and &pmd) vs (nmropt=1 and disang-rst)
Dear developers,
I have compared the GPU performance of two different umbrella sampling methods in amber using AMBER18, fully updated. I observed low performance when turning on US via infe=1 and the &pmd - the output decreased form 20 ns/day to 6 ns/day (compared to infe=0). When applying the same restraint using nmropt=1, no decrease in performance was observed. Is this behavior correct/to be expected?
Cheers,
Åsmund
###### pmd #####
&cntrl
ig=-1
iwrap =1
imin=0,irest=1,ntx=5,
nstlim=250000,dt=0.002,
ntc=2,ntf=2,
cut=10, ntb=1, ntp=0, tautp=10.0,
ntpr=500, ntwx=500, ntwr=1000,
ntt=1,
temp0=300.0,
tol=1.0e-8,jfastw=0,
infe=1,
/
&ewald
dsum_tol=1.0e-6,
order=4, skinnb=2.0, vdwmeth=1,
/
&pmd
output_file='pmd.dat'
output_freq=50
cv_file='cv.in'
/
NSTEP = 25500 TIME(PS) = 591.000 TEMP(K) = 300.71 PRESS = 0.0
Etot = -657462.4218 EKtot = 116350.1172 EPtot = -773812.5389
BOND = 1010.6757 ANGLE = 2714.6333 DIHED = 1961.8964
UB = 0.0000 IMP = 0.0000 CMAP = 263.6470
1-4 NB = 1232.4826 1-4 EEL = 15833.8706 VDWAALS = 113422.0802
EELEC = -910251.8367 EHBOND = 0.0000 RESTRAINT = 0.0119
EAMBER (non-restraint) = -773812.5509
------------------------------------------------------------------------------
| Current Timing Info
| -------------------
| Total steps : 250000 | Completed : 25500 | Remaining : 224500
|
| Average timings for last 2500 steps:
| Elapsed(s) = 66.90 Per Step(ms) = 26.76
| ns/day = 6.46 seconds/ns = 13379.64
|
| Average timings for all steps:
| Elapsed(s) = 681.80 Per Step(ms) = 26.74
| ns/day = 6.46 seconds/ns = 13368.59
##### nmropt ####
&cntrl
ig=-1
iwrap =1
imin=0,irest=1,ntx=5,
nstlim=250000,dt=0.002,
ntc=2,ntf=2,
cut=10, ntb=1, ntp=0, tautp=10.0,
ntpr=500, ntwx=500, ntwr=1000,
ntt=1,
temp0=300.0,
tol=1.0e-8,jfastw=0,
nmropt=1,
/
&ewald
dsum_tol=1.0e-6,
order=4, skinnb=2.0, vdwmeth=1,
/
&wt
type='DUMPFREQ', istep1=50,
/
&wt type='END',
/
DISANG=disang.dat
DUMPAVE=distance.dat
NSTEP = 219000 TIME(PS) = 978.000 TEMP(K) = 299.62 PRESS = 0.0
Etot = -657369.7698 EKtot = 115927.8359 EPtot = -773297.6057
BOND = 1036.2415 ANGLE = 2707.2816 DIHED = 1990.8650
UB = 0.0000 IMP = 0.0000 CMAP = 255.2401
1-4 NB = 1235.2865 1-4 EEL = 15804.5869 VDWAALS = 112549.9441
EELEC = -908877.2922 EHBOND = 0.0000 RESTRAINT = 0.2408
EAMBER (non-restraint) = -773297.8465
NMR restraints: Bond = 0.241 Angle = 0.000 Torsion = 0.000
===============================================================================
------------------------------------------------------------------------------
| Current Timing Info
| -------------------
| Total steps : 250000 | Completed : 219000 | Remaining : 31000
|
| Average timings for last 9000 steps:
| Elapsed(s) = 60.40 Per Step(ms) = 6.71
| ns/day = 25.75 seconds/ns = 3355.59
|
| Average timings for all steps:
| Elapsed(s) = 1463.42 Per Step(ms) = 6.68
| ns/day = 25.86 seconds/ns = 3341.14
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Nov 05 2019 - 05:00:02 PST
