计算化学公社

标题: orca在双路计算机上不同的cpu耗时相差很大 [打印本页]

作者
Author: carl0206 时间: 2022-5-4 14:45
标题: orca在双路计算机上不同的cpu耗时相差很大
有个很奇怪的问题
用一台2*8179M的机器做orca(5.0.1版本，openmpi版本411）运算，在cpu0上的计算时间显著低于cpu1，不知道有没有人知道为什么会出现这样的情况，已经怎样提高cpu1的算力？

命令如下：
/xx/orca501/orca 1.inp "-rf /xx/cores/1.txt" > 1.out

1.txt是按照Sobereva老师建议的内核绑定文件：
rank 0=xw-desktop slot=0:0
rank 1=xw-desktop slot=0:1
rank 2=xw-desktop slot=0:2
rank 3=xw-desktop slot=0:3

或者
rank 0=xw-desktop slot=1:0
rank 1=xw-desktop slot=1:1
rank 2=xw-desktop slot=1:2
rank 3=xw-desktop slot=1:3

分别对应CPU0和CPU1

1.inp:
! M062X def2-TZVP def2/J TightSCF Opt Freq CPCM(chloroform) DEFGRID3 D3zero nopop noautostart miniprint

%maxcore 3000
%pal
nprocs 4 end

* xyz 0 1
xxxxxx

*

用CPU0：
Timings for individual modules:

Sum of individual times       ...    461.427 sec (= 7.690 min)
GTO integral calculation       ...       2.398 sec (= 0.040 min) 0.5 %
SCF iterations                ...    168.533 sec (= 2.809 min)  36.5 %
SCF Gradient evaluation       ...    40.387 sec (= 0.673 min) 8.8 %
Geometry relaxation          ...       0.240 sec (= 0.004 min) 0.1 %
Analytical frequency calculation...    249.870 sec (= 4.164 min)  54.2 %
                           ****ORCA TERMINATED NORMALLY****
TOTAL RUN TIME: 0 days 0 hours 7 minutes 43 seconds 819 msec

如果用CPU1：

Timings for individual modules:

Sum of individual times       ...    856.495 sec (=  14.275 min)
GTO integral calculation       ...       2.815 sec (= 0.047 min) 0.3 %
SCF iterations                ...    178.850 sec (= 2.981 min)  20.9 %
SCF Gradient evaluation       ...    42.603 sec (= 0.710 min) 5.0 %
Geometry relaxation          ...       0.281 sec (= 0.005 min) 0.0 %
Analytical frequency calculation...    631.947 sec (=  10.532 min)  73.8 %
                           ****ORCA TERMINATED NORMALLY****
TOTAL RUN TIME: 0 days 0 hours 14 minutes 18 seconds 854 msec

作者
Author: sobereva 时间: 2022-5-5 08:21
监控一下计算过程中CPU运行频率，看看两个U有没有显著差别。并且也关注一下CPU温度，看看是否其中一个过高

作者
Author: carl0206 时间: 2022-5-6 09:53

sobereva 发表于 2022-5-5 08:21
监控一下计算过程中CPU运行频率，看看两个U有没有显著差别。并且也关注一下CPU温度，看看是否其中一个过高

谢谢回复，频率和温度都没问题，我看了一眼输出文件，主要是在算频率的时候耗时差别很大：
CPU0：
----------------------------------------------
Forming right-hand sides of CP-SCF equations    ...
----------------------------------------------
One electron integral derivatives  (SHARK)    ... done (    12.1 sec)
Transforming the overlap derivative matrices    ... done (    0.8 sec)
Adding tr(S(x)F(y) C-PCM term to the hessian    ... done (    0.0 sec)
Making the Q(x) pseudodensities                ... done (    0.0 sec)
Adding the E*S(x)*S(y) terms to the Hessian    ... done (    0.0 sec)
Calculating energy weighted overlap derivatives  ... done (    0.0 sec)
Two electron integral derivatives (SHARK-RI)    ... done (    15.7 sec)
Exchange-correlation integral derivatives       ... done (    61.2 sec)
tr(F(y)Q(x)) contribution to the Hessian       ... done (    0.0 sec)
Response fock operator R(S(x)) (SHARK-RIJCOSX) ... done (    11.3 sec)
XC Response fock operator R(S(x))             ... done (    19.2 sec)
tr(F(y)S(x)) contribution to the Hessian       ... done (    0.0 sec)
Transforming and finalizing RHSs                ... done (    0.0 sec)

----------------------------------------------
Solving the CP-SCF equations (RIJCOSX)          ...
----------------------------------------------
   CP-SCF ITERATION 0: 1.3414e-01 ( 6.1 sec 0/ 24 done)
   CP-SCF ITERATION 1: 2.1732e-02 ( 6.4 sec 0/ 24 done)
   CP-SCF ITERATION 2: 7.7381e-03 ( 6.5 sec 0/ 24 done)
   CP-SCF ITERATION 3: 1.7551e-03 ( 6.5 sec 0/ 24 done)
   CP-SCF ITERATION 4: 5.0784e-04 ( 6.5 sec 0/ 24 done)
   CP-SCF ITERATION 5: 1.5690e-04 ( 6.5 sec  17/ 24 done)
   CP-SCF ITERATION 6: 3.2301e-05 ( 2.5 sec  24/ 24 done)
                  *** THE CP-SCF HAS CONVERGED ***

                                             ... done (    41.2 sec)
Forming perturbed density Hessian contributions  ... done (    0.0 sec)
Calculating [Q*(d^2 A/dXdY)*Q  ] C-PCM term    ... done (    0.0 sec)
Calculating [d^2(V_N)/(dXdY)*Q ] C-PCM term    ... done (    0.0 sec)
Calculating [d^2(V_el)/(dXdY)*Q] C-PCM term    ... done (    0.4 sec)
Calculating C-PCM term that depends on dQ/dX    ... done (    0.0 sec)
2nd integral derivative contribs (SHARK-RI)    ... done (    62.5 sec)
Exchange-correlation Hessian                   ... done (    16.3 sec)
Dispersion correction in the Hessian          ... done (    0.5 sec)
Dipole derivatives                            ... done (    0.3 sec)

Total SCF Hessian time: 0 days 0 hours 4 min 9 sec

CPU1：
----------------------------------------------
Forming right-hand sides of CP-SCF equations    ...
----------------------------------------------
One electron integral derivatives  (SHARK)    ... done (    12.5 sec)
Transforming the overlap derivative matrices    ... done (    0.8 sec)
Adding tr(S(x)F(y) C-PCM term to the hessian    ... done (    0.0 sec)
Making the Q(x) pseudodensities                ... done (    0.0 sec)
Adding the E*S(x)*S(y) terms to the Hessian    ... done (    0.0 sec)
Calculating energy weighted overlap derivatives  ... done (    0.0 sec)
Two electron integral derivatives (SHARK-RI)    ... done (    15.9 sec)
Exchange-correlation integral derivatives       ... done (    86.5 sec)
tr(F(y)Q(x)) contribution to the Hessian       ... done (    0.1 sec)
Response fock operator R(S(x)) (SHARK-RIJCOSX) ... done ( 242.6 sec)
XC Response fock operator R(S(x))             ... done (    25.6 sec)
tr(F(y)S(x)) contribution to the Hessian       ... done (    0.0 sec)
Transforming and finalizing RHSs                ... done (    0.1 sec)

----------------------------------------------
Solving the CP-SCF equations (RIJCOSX)          ...
----------------------------------------------
   CP-SCF ITERATION 0: 1.3414e-01 (  17.2 sec 0/ 24 done)
   CP-SCF ITERATION 1: 2.1732e-02 (  16.9 sec 0/ 24 done)
   CP-SCF ITERATION 2: 7.7381e-03 (  16.9 sec 0/ 24 done)
   CP-SCF ITERATION 3: 1.7551e-03 (  16.7 sec 0/ 24 done)
   CP-SCF ITERATION 4: 5.0784e-04 (  16.8 sec 0/ 24 done)
   CP-SCF ITERATION 5: 1.5690e-04 (  16.8 sec  17/ 24 done)
   CP-SCF ITERATION 6: 3.2301e-05 ( 4.2 sec  24/ 24 done)
                  *** THE CP-SCF HAS CONVERGED ***

                                             ... done ( 106.1 sec)
Forming perturbed density Hessian contributions  ... done (    0.0 sec)
Calculating [Q*(d^2 A/dXdY)*Q  ] C-PCM term    ... done (    0.0 sec)
Calculating [d^2(V_N)/(dXdY)*Q ] C-PCM term    ... done (    0.0 sec)
Calculating [d^2(V_el)/(dXdY)*Q] C-PCM term    ... done (    0.4 sec)
Calculating C-PCM term that depends on dQ/dX    ... done (    0.1 sec)
2nd integral derivative contribs (SHARK-RI)    ... done (    66.4 sec)
Exchange-correlation Hessian                   ... done (    44.9 sec)
Dispersion correction in the Hessian          ... done (    14.5 sec)
Dipole derivatives                            ... done (    1.5 sec)

Total SCF Hessian time: 0 days 0 hours 10 min 31 sec

作者
Author: sobereva 时间: 2022-5-6 10:09

carl0206 发表于 2022-5-6 09:53
谢谢回复，频率和温度都没问题，我看了一眼输出文件，主要是在算频率的时候耗时差别很大：
CPU0：
---- ...

看看机子上内存插的方式对不对，是否两个CPU的内存通道都等效地利用了

作者
Author: carl0206 时间: 2022-5-6 11:10

sobereva 发表于 2022-5-6 10:09
看看机子上内存插的方式对不对，是否两个CPU的内存通道都等效地利用了

一共12个槽，都插满了，应该没问题吧

欢迎光临计算化学公社 (http://bbs.keinsci.com/)