[601bd9e] | 1 | #!/bin/bash - |
---|
| 2 | |
---|
| 3 | false=0; true=1 |
---|
| 4 | |
---|
| 5 | # Usage: arch [ hostname ] returns hostname, cores, startcore |
---|
| 6 | # |
---|
| 7 | # Define machine architecture based on starting socket, CPUs (cores) per socket, number of |
---|
| 8 | # sockets, has hyperthreading. |
---|
| 9 | |
---|
| 10 | start=0 |
---|
| 11 | |
---|
| 12 | arch() { |
---|
| 13 | hostname=${1:-`hostname`} # return value |
---|
| 14 | hashyper=${true} # assume machine has hyperthreads |
---|
| 15 | if [ "${hostname}" = "plg2" ] ; then |
---|
| 16 | startsocket=${start} |
---|
| 17 | cps=16 # coresPerSocket |
---|
| 18 | sockets=2 |
---|
| 19 | hashyper=${false} # has no hyperthreads |
---|
| 20 | elif [ "${hostname}" = "nasus" ] ; then |
---|
| 21 | startsocket=${start} |
---|
| 22 | cps=64 # coresPerSocket |
---|
| 23 | sockets=2 |
---|
| 24 | elif [ "${hostname}" = "pyke" ] ; then |
---|
| 25 | startsocket=${start} |
---|
| 26 | cps=24 # coresPerSocket |
---|
| 27 | sockets=2 |
---|
| 28 | elif [ "${hostname}" = "jax" ] ; then |
---|
| 29 | startsocket=${start} |
---|
| 30 | cps=24 # coresPerSocket |
---|
| 31 | sockets=4 |
---|
| 32 | else |
---|
| 33 | echo "unsupported host" ${hostname} |
---|
| 34 | exit 1 |
---|
| 35 | fi |
---|
| 36 | cores=$(( ${cps} * ${sockets} )) |
---|
| 37 | startcore=$(( ${startsocket} * ${cps} )) |
---|
| 38 | } |
---|
| 39 | |
---|
| 40 | # Usage: affinity (global cps, sockets, startsocket, hashyper, cores, startcore, wrap) |
---|
| 41 | # returns taskset argument |
---|
| 42 | # |
---|
| 43 | # This routine assumes hyperthreading has only 2 hyperthreads per core. |
---|
| 44 | # |
---|
| 45 | # If hyperthread scanning is used: processor units are assigned across the low-number hyperthreads |
---|
| 46 | # of the socket's cores. When the low-number hyperthreads are filled, the high-number hyperhtreads |
---|
| 47 | # are assigned across the socket's cores. Then the next socket is assigned. |
---|
| 48 | # |
---|
| 49 | # If hyperthread wrapping is used: processor units are assigned in low/high-number pairs of |
---|
| 50 | # hyperthreads across the socket's cores. Then the next socket is assigned. |
---|
| 51 | |
---|
| 52 | wrap=${false} # set to control hyperthread assignment across socket cores |
---|
| 53 | |
---|
| 54 | affinity() { |
---|
| 55 | if [ ${wrap} -eq ${true} -a ${hashyper} -eq ${false} ] ; then |
---|
| 56 | echo "architecture does not support hyperthreading for wrapping" |
---|
| 57 | exit 1 |
---|
| 58 | fi |
---|
| 59 | taskset="" # return value |
---|
| 60 | set -- $(( ${1} - 1 )) # decrement $1 |
---|
| 61 | if [ ${1} -eq 0 ] ; then taskset="${startcore}-${startcore}"; return; fi |
---|
| 62 | if [ ${1} -ge $(( ${cps} * ( ${sockets} - ${startsocket} ) * ( ${hashyper} + 1 ) )) ] ; then # error |
---|
| 63 | echo "not enough cores $(( ${cores} * ${sockets} )) for $(( ${1} + 1 )) starting at ${startcore}" |
---|
| 64 | exit 1 |
---|
| 65 | fi |
---|
| 66 | if [ ${hashyper} -eq ${false} ] ; then taskset="${startcore}-$(( ${1} + ${startcore} ))"; return; fi # no hyperthreads |
---|
| 67 | start2=$(( ${startcore} + ${cores} )) |
---|
| 68 | if [ ${wrap} -eq ${true} ] ; then # hyperthread wrapping |
---|
| 69 | end1=$(( ${1} / 2 + ${startcore} )) |
---|
| 70 | end2=$(( ${end1} + ${cores} )) |
---|
| 71 | if [ $(( ${1} % 2 )) -eq 0 ] ; then |
---|
| 72 | end2=$(( ${end2} - 1 )) |
---|
| 73 | fi |
---|
| 74 | taskset="${startcore}-${end1},${start2}-${end2}" |
---|
| 75 | else # hyperthread scanning |
---|
| 76 | if [ ${1} -lt ${cps} ] ; then taskset="${startcore}-$(( ${1} + ${startcore} ))"; return; fi |
---|
| 77 | filled=$(( ${1} / ( ${cps} * 2 ) * ${cps} )) |
---|
| 78 | modulus=$(( ${1} % ( ${cps} * 2 ) )) # leftover cores added to saturated sockets |
---|
| 79 | if [ ${modulus} -gt ${cps} ] ; then |
---|
| 80 | taskset="${startcore}-$(( ${startcore} + ${filled} + ${cps} - 1 )),${start2}-$(( ${start2} + ${filled} + ${modulus} % ${cps} ))" |
---|
| 81 | else |
---|
| 82 | taskset="${startcore}-$(( ${startcore} + ${filled} + ${modulus} )),${start2}-$(( ${start2} + ${filled} - 1 ))" |
---|
| 83 | fi |
---|
| 84 | fi |
---|
| 85 | } |
---|
| 86 | |
---|
| 87 | numtimes=5 |
---|
| 88 | |
---|
| 89 | # bench_cores='1 2 4 8 16 24 32' |
---|
| 90 | bench_cores='8 16 24 32 48' |
---|
| 91 | # bench_cores='32' |
---|
| 92 | |
---|
| 93 | # toggle benchmarks |
---|
| 94 | executor=${true} |
---|
| 95 | matrix=${true} |
---|
| 96 | repeat=${true} |
---|
| 97 | balance=${true} |
---|
| 98 | static=${false} |
---|
| 99 | dynamic=${false} |
---|
| 100 | |
---|
| 101 | # executor config |
---|
| 102 | batch='1' |
---|
| 103 | nRounds='500' |
---|
| 104 | |
---|
| 105 | # matrix config |
---|
| 106 | size='3000' |
---|
| 107 | |
---|
| 108 | # repeat config |
---|
| 109 | messages='100000' |
---|
| 110 | n_repeats='100' |
---|
| 111 | |
---|
| 112 | # balance config |
---|
| 113 | nOneRounds='2000' |
---|
| 114 | nMultiRounds='800' |
---|
| 115 | |
---|
| 116 | # static config |
---|
| 117 | n_static_sends='10000000' |
---|
| 118 | |
---|
| 119 | # dynamic config |
---|
| 120 | n_dynamic_sends='10000000' |
---|
| 121 | |
---|
| 122 | # names=('NOSTEAL' 'CLOSE' 'SEARCH') |
---|
| 123 | # var_flags=('-D__STEAL=1 -DRAND=1' '-D__STEAL=1 -DCLOSE=1' '-D__STEAL=1 -DSEARCH=1') |
---|
| 124 | |
---|
| 125 | names=('SEARCH') |
---|
| 126 | var_flags=('-D__STEAL=1 -DSEARCH=1') |
---|
| 127 | |
---|
| 128 | |
---|
| 129 | # names=('NOSTEAL' 'SEARCH' 'RAND' 'SEARCHC' 'RANDC') |
---|
| 130 | # var_flags=('' '-D__STEAL=1 -DSEARCH=1' '-D__STEAL=1 -DRAND=1' '-D__STEAL=1 -DSEARCH=1 -DCYCLE' '-D__STEAL=1 -DRAND=1 -DCYCLE') |
---|
| 131 | |
---|
| 132 | # names=('NOSTEAL' 'RAND' 'CLOSE' 'SEARCH') |
---|
| 133 | # var_flags=('' '-D__STEAL=1 -DRAND=1' '-D__STEAL=1 -DCLOSE=1' '-D__STEAL=1 -DSEARCH=1') |
---|
| 134 | |
---|
| 135 | cfa_flags='-quiet -O3 -nodebug -DNDEBUG' |
---|
| 136 | |
---|
| 137 | cfa=~/cfa-cc/driver/cfa |
---|
| 138 | |
---|
| 139 | preprint='' |
---|
| 140 | repeat_command() { |
---|
| 141 | t=1 |
---|
| 142 | while [ ${t} -le ${numtimes} ] ; do |
---|
| 143 | echo -n -e ${preprint} |
---|
| 144 | "${@}" |
---|
| 145 | t=`expr ${t} + 1` |
---|
| 146 | done |
---|
| 147 | } |
---|
| 148 | |
---|
| 149 | arch nasus |
---|
| 150 | |
---|
| 151 | echo $numtimes |
---|
| 152 | echo $bench_cores |
---|
| 153 | |
---|
| 154 | for i in ${!names[@]}; do |
---|
| 155 | echo -n ${names[$i]}" " |
---|
| 156 | done |
---|
| 157 | echo "" |
---|
| 158 | |
---|
| 159 | # /usr/bin/time -f "%Uu %Ss %Er %Mkb" |
---|
| 160 | |
---|
| 161 | if [ ${executor} -eq ${true} ] ; then |
---|
| 162 | echo "executor" |
---|
| 163 | for i in ${!names[@]}; do |
---|
| 164 | echo ${names[$i]}':' |
---|
| 165 | echo -e "proc\ttime (s)" |
---|
| 166 | ${cfa} ${cfa_flags} ${var_flags[$i]} executor.cfa > /dev/null 2>&1 |
---|
| 167 | for p in ${bench_cores} ; do |
---|
| 168 | affinity ${p} |
---|
| 169 | # echo "taskset -c ${startcore}-`expr ${startcore} + ${p} - 1` ./a.out 40000 100 ${nRounds} ${p} ${batch}" |
---|
| 170 | preprint="${p}\t" |
---|
| 171 | repeat_command taskset -c ${taskset} ./a.out 40000 100 ${nRounds} ${p} ${batch} |
---|
| 172 | done |
---|
| 173 | rm a.out |
---|
| 174 | done |
---|
| 175 | echo "" |
---|
| 176 | fi |
---|
| 177 | |
---|
| 178 | if [ ${matrix} -eq ${true} ] ; then |
---|
| 179 | echo "matrix" |
---|
| 180 | for i in ${!names[@]}; do |
---|
| 181 | echo ${names[$i]} |
---|
| 182 | echo -e "proc\ttime (s)" |
---|
| 183 | ${cfa} ${cfa_flags} ${var_flags[$i]} matrix.cfa > /dev/null 2>&1 |
---|
| 184 | for p in ${bench_cores} ; do |
---|
| 185 | affinity ${p} |
---|
| 186 | preprint="${p}\t" |
---|
| 187 | repeat_command taskset -c ${taskset} ./a.out ${size} ${size} ${size} ${p} |
---|
| 188 | done |
---|
| 189 | rm a.out |
---|
| 190 | done |
---|
| 191 | echo "" |
---|
| 192 | fi |
---|
| 193 | |
---|
| 194 | if [ ${repeat} -eq ${true} ] ; then |
---|
| 195 | echo "repeat" |
---|
| 196 | echo -e "proc\ttime (s)" |
---|
| 197 | for i in ${!names[@]}; do |
---|
| 198 | echo ${names[$i]} |
---|
| 199 | ${cfa} ${cfa_flags} ${var_flags[$i]} repeat.cfa > /dev/null 2>&1 |
---|
| 200 | for p in ${bench_cores} ; do |
---|
| 201 | affinity ${p} |
---|
| 202 | preprint="${p}\t" |
---|
| 203 | repeat_command taskset -c ${taskset} ./a.out ${messages} ${p} 256 ${n_repeats} |
---|
| 204 | done |
---|
| 205 | rm a.out |
---|
| 206 | done |
---|
| 207 | echo "" |
---|
| 208 | fi |
---|
| 209 | |
---|
| 210 | |
---|
| 211 | if [ ${static} -eq ${true} ] ; then |
---|
| 212 | echo "static" |
---|
| 213 | for i in ${!names[@]}; do |
---|
| 214 | echo ${names[$i]} |
---|
| 215 | ${cfa} ${cfa_flags} ${var_flags[$i]} static.cfa > /dev/null 2>&1 |
---|
| 216 | repeat_command taskset -c ${startcore} ./a.out ${n_static_sends} |
---|
| 217 | rm a.out |
---|
| 218 | done |
---|
| 219 | echo "" |
---|
| 220 | fi |
---|
| 221 | |
---|
| 222 | if [ ${dynamic} -eq ${true} ] ; then |
---|
| 223 | echo "dynamic" |
---|
| 224 | for i in ${!names[@]}; do |
---|
| 225 | echo ${names[$i]} |
---|
| 226 | ${cfa} ${cfa_flags} ${var_flags[$i]} dynamic.cfa > /dev/null 2>&1 |
---|
| 227 | repeat_command taskset -c ${startcore} ./a.out ${n_dynamic_sends} |
---|
| 228 | done |
---|
| 229 | echo "" |
---|
| 230 | fi |
---|
| 231 | |
---|
| 232 | |
---|
| 233 | if [ ${balance} -eq ${true} ] ; then |
---|
| 234 | echo "balance_one" |
---|
| 235 | for i in ${!names[@]}; do |
---|
| 236 | echo ${names[$i]}':' |
---|
| 237 | echo -e "proc\ttime (s)" |
---|
| 238 | ${cfa} ${cfa_flags} ${var_flags[$i]} balance.cfa > /dev/null 2>&1 |
---|
| 239 | for p in ${bench_cores} ; do |
---|
| 240 | affinity ${p} |
---|
| 241 | preprint="${p}\t" |
---|
| 242 | repeat_command taskset -c ${taskset} ./a.out 32 32 ${nOneRounds} ${p} |
---|
| 243 | done |
---|
| 244 | rm a.out |
---|
| 245 | done |
---|
| 246 | echo "" |
---|
| 247 | echo "balance_multi" |
---|
| 248 | for i in ${!names[@]}; do |
---|
| 249 | echo ${names[$i]}':' |
---|
| 250 | echo -e "proc\ttime (s)" |
---|
| 251 | ${cfa} ${cfa_flags} ${var_flags[$i]} -DMULTI balance.cfa > /dev/null 2>&1 |
---|
| 252 | for p in ${bench_cores} ; do |
---|
| 253 | affinity ${p} |
---|
| 254 | preprint="${p}\t" |
---|
| 255 | repeat_command taskset -c ${taskset} ./a.out 32 32 ${nMultiRounds} ${p} |
---|
| 256 | done |
---|
| 257 | rm a.out |
---|
| 258 | done |
---|
| 259 | echo "" |
---|
| 260 | fi |
---|
| 261 | |
---|