-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun.heimdallr
executable file
·114 lines (101 loc) · 3.73 KB
/
run.heimdallr
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
#!/bin/tcsh -x
#PBS -N heim
##PBS -q bw44-sm
##PBS -q bw44-lg
##PBS -l walltime=10:00
##PBS -l walltime=120:00
##PBS -l place=scatter,select=512:nodetype=BW36
##PBS -l place=scatter,select=4
##PBS -l place=scatter,select=128:clockmhz=2300:nodetype=BW36
##PBS -l place=scatter,select=32:clockmhz=2300:nodetype=BW36
#PBS -l place=scatter,select=16:clockmhz=2300:nodetype=BW36
##PBS -l place=scatter,select=4:clockmhz=2300:nodetype=BW36
##PBS -l place=scatter,select=1
#PBS -j oe
cd $PBS_O_WORKDIR
module unload craype-hugepages8M
# module swap craype-hugepages8M craype-hugepages2M
#module swap craype-hugepages8M craype-hugepages16M
#module swap pmi pmi/5.0.15
#module unload atp totalview perftools-base
#module load cray-shmem
module list
#module load cray-openshmemx
#setenv LD_LIBRARY_PATH $CRAY_LD_LIBRARY_PATH\:$LD_LIBRARY_PATH
# This benchmark is meant to be run most often on a single cpu,
# half of the cpus of a numa domain, all of the cpus
# of a numa domain, and then all of the cpus across
# the full shared domain.
# Other configurations may be interesting depending on
# the processor.
#
# It is very common for one to place 1 MPI rank on every cpu
# of a numa domain. If one runs with pure MPI then one would expect
# to not get any "wrong" answers even without using atomics.
#foreach ncore (1 2 4 8 16)
#foreach ncore (1 )
#foreach ncore (72 )
#foreach ncore (144 )
#foreach ncore (144 143 142 140 136 )
#foreach ncore (16 )
#foreach ncore (45 )
#foreach ncore (4608 )
foreach ncore (576 )
#foreach ncore ( 18432 )
# foreach ncore (72)
# foreach ncore (4 )
# foreach ncore (2 )
setenv file output.$ncore
echo $file
rm $file
@ nt = 1
foreach nthreads (1 2 )
#foreach nthreads ( 2 )
#foreach nthreads ( 1 )
#foreach nthreads (1 2 9 18 )
@ nt = $nt + 1
echo nt
#@ nthreads = ( 1 << $nt )
echo $nthreads
@ nr = $ncore / $nthreads
echo $nr
@ symsize = $nthreads * ( 2 << 29 )
echo $symsize
#@ tsize = $nt + 15
@ tsize = $nt + 20
#@ tsize = $nt + 10
#@ nu = $tsize - 2
@ nu = $tsize
setenv OMP_NUM_THREADS $nthreads
#setenv OMP_NUM_THREADS 9
setenv SHMEM_MAX_NUM_THREADS $nthreads
setenv SHMEM_SYMMETRIC_SIZE $symsize
echo $SHMEM_SYMMETRIC_SIZE
echo $tsize
#foreach hugepage (craype-hugepages8M craype-hugepages2M)
#foreach hugepage (craype-hugepages8M )
foreach hugepage (craype-hugepages2M )
module load $hugepage
# The arguments control the size of the table and the number of updates
# For example, if the first argument is 26 the size of the table will be 2^26*8 = 0.5 GiB
# and if the second argument is 25 each rank will execute 2^25 updates distribute across all threads.
# The total memory footprint will be the table size plus the number of updates times 8 bytes.
# The program simple replicates the full table and updates on each mpi rank.
# The spirit of the benchmark is that when running using all of the CPUs in a NUMA domain
# that the size of the table consume significantly more space that is available
# in the last level cache.
# aprun -n$nr -d$nthreads ./heimdallr 15 6 >>! $file
# aprun -n$nr -d$nthreads ./heimdallr 15 2 >>! $file
#aprun -n8 -d$nthreads -cc depth ./heimdallr 15 2 >>! $file
# aprun -n1 -d$nthreads -cc depth ./heimdallr 15 10 >>! $file
#aprun -n$nr -d$nthreads -cc depth ./heimdallr -t $tsize -n $nu -G -128 -W 1024 -T ATOMIC_INC -T HABU_INC >>! $file
aprun -n$nr -d$nthreads -cc depth ./heimdallr -t $tsize -n $nu -G -128 -W 2048 -T ALL >>! $file
# aprun -n$nr -d$nthreads -cc depth ./heimdallr -t $tsize -n $nu -G -128 -W 1024 -T HABU_INC -T ATOMIC_INC >>! $file
# aprun -n$nr -d$nthreads -cc depth ./heimdallr $tsize 15 >>! $file
#aprun -n$nrank -d$nthreads ./heimdallr 28 26 > $file
module unload $hugepage
end
end
end
cat $file
#qsub run.heimdallr