parallel load difference under linux (long)

From: Tru Huynh <tru_at_pasteur.fr>
Date: Thu 22 Mar 2001 21:01:47 +0100

Hello,

Please bear with my ignorance on parallel sander (amber6) :)
On our toy cluster (4x 1GHz K7, fast ethernet) the breakout of
a benchmark run is as follow:
(logfile and tail of the output file)

x4
               Parallel Profiling Results

                          A
                  N n
          F o g G
          X n B S G B R
O T
          d b n h B d u
t o
          i o d a r i n
h t
  P s n D k a s m
e a
  E t d i e d t d
r l
---------------------------------------------------------------------
  0 11.3 154.8 16.0 1.6 599.8 5.0 1.7 1.3
791.6
  1 6.3 153.6 15.5 1.6 723.3 2.9 0.7 0.5
904.5
  2 6.4 154.2 14.9 1.5 918.9 3.2 0.9 0.4
1100.5
  3 6.5 153.3 14.8 1.5 838.6 2.9 0.8 0.5
1018.9
 av 7.6 154.0 15.3 1.6 770.2 3.5 1.0
std 2.1 0.6 0.6 0.0 120.5 0.9 0.4
min 6.3 153.3 14.4 1.5 599.8 2.9 0.7
max 11.3 154.8 16.1 1.6 918.9 5.0 1.7
---------------------------------------------------------------------
|
| ELAPSED TIME = 791.800 TOTAL TIME = 791.800
|
|
| Routine Sec %
| ----------------------------
| Nonbond 154.83 19.55
| Bond 0.30 0.04
| Angle 3.31 0.42
| Dihedral 12.44 1.57
| Shake 1.59 0.20
| GBrad 599.81 75.75
| Force 0.00 0.00
| GBraddist 4.98 0.63
| F,Xdist 11.29 1.43
| Other 1.54 0.19
| ----------------------------
| Total 791.80 0.22 Hours

| Nonsetup 791.60 99.97%

| Highest rstack allocated: 4815
| MAX_RSTACK = 1600000

| Highest istack allocated: 0
| MAX_ISTACK = 100000

| Setup wallclock 0 seconds
| Nonsetup wallclock 0 seconds


x2

               Parallel Profiling Results

                          A
                  N n
          F o g G
          X n B S G B R
O T
          d b n h B d u
t o
          i o d a r i n
h t
  P s n D k a s m
e a
  E t d i e d t d
r l
---------------------------------------------------------------------
  0 7.9 308.1 28.9 2.0 1318.8 3.3 1.7 1.4
1672.0
  1 6.3 305.5 27.9 2.1 1754.7 3.0 1.5 0.5
2101.5
 av 7.1 306.8 28.4 2.0 1536.8 3.2 1.6
std 0.8 1.3 0.5 0.1 217.9 0.1 0.1
min 6.3 305.5 27.9 2.0 1318.8 3.0 1.5
max 7.9 308.1 29.0 2.1 1754.7 3.3 1.7
---------------------------------------------------------------------
|
| ELAPSED TIME = 1672.200 TOTAL TIME = 1672.200
|
|
| Routine Sec %
| ----------------------------
| Nonbond 308.11 18.43
| Bond 0.60 0.04
| Angle 6.84 0.41
| Dihedral 21.45 1.28
| Shake 1.95 0.12
| GBrad 1318.83 78.87
| Force 0.00 0.00
| GBraddist 3.28 0.20
| F,Xdist 7.89 0.47
| Other 1.54 0.09
| ----------------------------
| Total 1672.20 0.46 Hours

| Nonsetup 1672.02 99.99%

| Highest rstack allocated: 4815
| MAX_RSTACK = 1600000

| Highest istack allocated: 0
| MAX_ISTACK = 100000

| Setup wallclock 0 seconds
| Nonsetup wallclock 0 seconds

x1

               Parallel Profiling Results

                          A
                  N n
          F o g G
          X n B S G B R
O T
          d b n h B d u
t o
          i o d a r i n
h t
  P s n D k a s m
e a
  E t d i e d t d
r l
---------------------------------------------------------------------
  0 0.0 612.7 55.9 2.7 3070.5 0.0 2.1 1.0
3745.0
 av 0.0 612.7 55.9 2.7 3070.5 0.0 2.1
std 0.0 0.0 0.0 0.0 0.0 0.0 0.0
min 0.0 612.7 55.9 2.7 3070.5 0.0 2.1
max 0.0 612.7 55.9 2.7 3070.5 0.0 2.1
---------------------------------------------------------------------
|
| ELAPSED TIME = 3745.130 TOTAL TIME = 3745.130
|
|
| Routine Sec %
| ----------------------------
| Nonbond 612.73 16.36
| Bond 1.32 0.04
| Angle 11.19 0.30
| Dihedral 43.44 1.16
| Shake 2.71 0.07
| GBrad 3070.51 81.99
| Force 0.00 0.00
| GBraddist 0.01 0.00
| F,Xdist 0.01 0.00
| Other 1.16 0.03
| ----------------------------
| Total 3745.13 1.04 Hours
 
| Nonsetup 3744.99 100.00%
 
| Highest rstack allocated: 0
| MAX_RSTACK = 1600000
 
| Highest istack allocated: 0
| MAX_ISTACK = 100000
 
| Setup wallclock 0 seconds
| Nonsetup wallclock 0 seconds

This was run with lam-mpi (current) but we observe the
same behaviour with mpich, and if we randomize the nodes.

Is there any one who can explain the load repartition when
running a parallel sander?

Regards,

Tru

-- 
mailto:tru_at_pasteur.fr | 
Institut Pasteur - Bioinformatique Structurale
25-28 rue du Docteur Roux, 75724 Paris CEDEX 15 France
Received on Thu Mar 22 2001 - 12:01:47 PST
Custom Search