Preliminary Look

1
Preliminary Look

• Steven Piepers
• Quantum Monte Carlo Code

2
What is the code?As far as I can tell

• Quantum Monte Carlo
• Generate input file
• Chomp (all the time)
• Solving wave equation/minimizing potentials
• Send minimized nuclei to master
• Possible redist work
• Synchronize
• Lots of repetition

3
Help from the source

• Profiling pre-exists
• Time, FLOPS, memory
• Memory is just IBM/Mac
• FLOPS counts in code
• MPI
• Tracks the amount communicated in messages

4
How does it perform?

• Computation takes from 80-95 of wall clock time
• MPI is very simple
• One whole nucleus per proc
• Memory Limitation
• Keeps MPI costs very low
• i.e. can run on ethernet
• So, it scales quite well.

5
Basic Algorithm

• Init fbn wave function
• Init some positions (randomly)
• Init wave functions and probability density
• Propogate time/synchronize procs

6

• Number Total msec/
  Flop/ MFLOPS
• of cases min. case
  case
• Wavefunctions 670250 1.8
  0.163 165499 1014.881
• 2-body Prop. 0 0.0
  0.000 0 0.000
• 3-body Prop. Vijk 0 0.0
  0.000 0 0.000
• Propagation step 0 0.0
  0.000 0 0.000
• Other Propagation 0 0.0
  0.000 0 0.000
• Other vij 0 0.0
  0.000 0 0.000
• Total accounted for time 110926. MFLOP
  1.8 Min 1014.881 MFLOPS
• MFLOP/Wall-second 2857.285
• 96.1 of total compute time is accounted for
• ..
• Master got 53.45 Mbytes 1.3767
  Mbytes/sec
• Master sent 0.00 Mbytes 0.0000
  Mbytes/sec
• Master got 30005 messages 772.8822
  messages/sec

7
Totals for all slaves Num
Cases Total Time Time/case
Min.
Sec. Propagation 4059060
116.9 0.002 Branching 379
0.0 0.003 Energies
223050 59.2 0.016
Kinetic
22.2 0.006 2-b Potential
31.9 0.009 3-b
Potential 4.4
0.001 Densities
0.7 0.000 Other compute
223050 0.3 0.000 All compute
223050 176.4
0.047 Config. write (wall) 223050 6.7
0.002 Above include Number
Total msec/ Flop/ MFLOPS
of cases min.
case case Wavefunctions 22612208
62.2 0.165 165031 1000.523 2-body
Prop. 4059060 39.8 0.588
324854 552.399 3-body Prop. Vijk 4282110
60.1 0.842 805563 957.102 Propagation
step 4059060 7.9 0.117 34435
293.892 Other Propagation 4059060 0.6
0.009 2115 237.629 Other vij
223050 4.7 1.251 1210116
967.236
8
Total accounted for time 8918108. MFLOP
175.2 Min
848.397
MFLOPS MFLOP/Wall-second 15321.884 99.3
of total compute time is accounted for
Master got 505.62 Mbytes 0.8687
Mbytes/sec Master sent 5.21 Mbytes
0.0089 Mbytes/sec Master got 223639 messages
384.2262 messages/sec Master wall min. in
loop 9.7 Master idle wall min.
9.4 Total available wall min.
184.3 Total compute min.
176.4 Efficency
95.7 Speed up 18.2
9
Data Structures

• Pretty straight forward arrays
• Wave function solution on the grid
• Grows 2 with number of particle
• Quickly moves from FLOPS to memory bound