| 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} + ${filled} - 1 ))"
|
|---|
| 83 | fi
|
|---|
| 84 | fi
|
|---|
| 85 | }
|
|---|
| 86 |
|
|---|
| 87 | #used for output formatting
|
|---|
| 88 | column_headers="proc\ttime (s)"
|
|---|
| 89 |
|
|---|
| 90 | # executor config
|
|---|
| 91 | batch='100'
|
|---|
| 92 | nRounds='400' #500
|
|---|
| 93 |
|
|---|
| 94 | # matrix config
|
|---|
| 95 | size='3072'
|
|---|
| 96 |
|
|---|
| 97 | # repeat config
|
|---|
| 98 | messages='100000'
|
|---|
| 99 | n_repeats='200' #200
|
|---|
| 100 |
|
|---|
| 101 | # balance config
|
|---|
| 102 | nOneRounds='2000'
|
|---|
| 103 | nMultiRounds='800'
|
|---|
| 104 |
|
|---|
| 105 | # static config
|
|---|
| 106 | n_static_sends='100000000'
|
|---|
| 107 | n_static_sends_slow='10000000'
|
|---|
| 108 |
|
|---|
| 109 | # dynamic config
|
|---|
| 110 | n_dynamic_sends='20000000'
|
|---|
| 111 | n_dynamic_sends_slow='2000000'
|
|---|
| 112 |
|
|---|
| 113 | numtimes=5
|
|---|
| 114 |
|
|---|
| 115 | # bench_cores='1 2 4 8 16 24 32'
|
|---|
| 116 | bench_cores='1 2 4 8 16 24 32 48'
|
|---|
| 117 | # bench_cores='48'
|
|---|
| 118 |
|
|---|
| 119 | # toggle missed gulp tracking config (overrides specified config)
|
|---|
| 120 | missed_gulps=${true}
|
|---|
| 121 | # missed_gulps=${false}
|
|---|
| 122 |
|
|---|
| 123 | # toggle mem collection config (overrides specified config)
|
|---|
| 124 | # outputs mem instead of time
|
|---|
| 125 | mem=${true}
|
|---|
| 126 | mem=${false}
|
|---|
| 127 |
|
|---|
| 128 | # toggle benchmarks
|
|---|
| 129 | executor=${true}
|
|---|
| 130 | matrix=${true}
|
|---|
| 131 | repeat=${true}
|
|---|
| 132 | balance=${true}
|
|---|
| 133 | static=${true}
|
|---|
| 134 | dynamic=${true}
|
|---|
| 135 | # executor=${false}
|
|---|
| 136 | # matrix=${false}
|
|---|
| 137 | # repeat=${false}
|
|---|
| 138 | # balance=${false}
|
|---|
| 139 | # static=${false}
|
|---|
| 140 | # dynamic=${false}
|
|---|
| 141 |
|
|---|
| 142 | # names=('Longest-Victim' 'No-Stealing' 'Random')
|
|---|
| 143 | # var_flags=('-D__STEAL=1 -DSEARCH=1' '' '-D__STEAL=1 -DRAND=1')
|
|---|
| 144 |
|
|---|
| 145 | # names=('CFA' 'CFA-EMPTY')
|
|---|
| 146 | # var_flags=('' '-DEMPTY')
|
|---|
| 147 |
|
|---|
| 148 | # names=('CFA')
|
|---|
| 149 | # var_flags=('-D__STEAL=1 -DSEARCH=1')
|
|---|
| 150 |
|
|---|
| 151 | # names=('CFA' 'CFA-STAT')
|
|---|
| 152 | # var_flags=('-D__STEAL=1 -DSEARCH=1' '-D__STEAL=1 -DSTATS')
|
|---|
| 153 |
|
|---|
| 154 | # names=()
|
|---|
| 155 | # var_flags=()
|
|---|
| 156 |
|
|---|
| 157 | names=('CFA')
|
|---|
| 158 | var_flags=('')
|
|---|
| 159 |
|
|---|
| 160 | runCAF=${true}
|
|---|
| 161 | runUCPP=${true}
|
|---|
| 162 | runPROTO=${true}
|
|---|
| 163 | runAKKA=${true}
|
|---|
| 164 | runCAF=${false}
|
|---|
| 165 | runUCPP=${false}
|
|---|
| 166 | runPROTO=${false}
|
|---|
| 167 | runAKKA=${false}
|
|---|
| 168 |
|
|---|
| 169 | if [ ${missed_gulps} -eq ${true} ] ; then
|
|---|
| 170 | bench_cores='2 4 8 16 24 32 48'
|
|---|
| 171 | column_headers="proc\tmissed\ttime (s)"
|
|---|
| 172 | names=('CFA')
|
|---|
| 173 | var_flags=('-D__STEAL=1 -DSEARCH=1 -DACTOR_STATS_QUEUE_MISSED')
|
|---|
| 174 | runCAF=${false}
|
|---|
| 175 | runUCPP=${false}
|
|---|
| 176 | runPROTO=${false}
|
|---|
| 177 | runAKKA=${false}
|
|---|
| 178 | executor=${true}
|
|---|
| 179 | matrix=${true}
|
|---|
| 180 | repeat=${true}
|
|---|
| 181 | fi
|
|---|
| 182 |
|
|---|
| 183 | if [ ${mem} -eq ${true} ] ; then
|
|---|
| 184 | bench_cores='48'
|
|---|
| 185 | column_headers="proc\tmem (kB)"
|
|---|
| 186 | names=('CFA')
|
|---|
| 187 | var_flags=('-D__STEAL=1 -DSEARCH=1')
|
|---|
| 188 | runCAF=${true}
|
|---|
| 189 | runUCPP=${true}
|
|---|
| 190 | runPROTO=${true}
|
|---|
| 191 | runAKKA=${true}
|
|---|
| 192 | executor=${true}
|
|---|
| 193 | matrix=${false}
|
|---|
| 194 | repeat=${false}
|
|---|
| 195 | balance=${false}
|
|---|
| 196 | static=${false}
|
|---|
| 197 | dynamic=${false}
|
|---|
| 198 | fi
|
|---|
| 199 |
|
|---|
| 200 | cfa=~/cfa-cc/driver/cfa
|
|---|
| 201 |
|
|---|
| 202 | # Helpers to minimize code duplication
|
|---|
| 203 |
|
|---|
| 204 | # repeats a command ${numtimes}
|
|---|
| 205 | preprint=''
|
|---|
| 206 | repeat_command() {
|
|---|
| 207 | t=1
|
|---|
| 208 | while [ ${t} -le ${numtimes} ] ; do
|
|---|
| 209 | echo -n -e ${preprint}
|
|---|
| 210 | "${@}"
|
|---|
| 211 | t=`expr ${t} + 1`
|
|---|
| 212 | done
|
|---|
| 213 | }
|
|---|
| 214 |
|
|---|
| 215 | # prints the leading info for a given run of a variant
|
|---|
| 216 | print_header() {
|
|---|
| 217 | echo ${1}':'
|
|---|
| 218 | if [ ${mem} -eq ${false} ] ; then
|
|---|
| 219 | echo -e $column_headers
|
|---|
| 220 | fi
|
|---|
| 221 | }
|
|---|
| 222 |
|
|---|
| 223 | # runs the current benchmark with provided args
|
|---|
| 224 | # only works for standard-run benchmarks (not Akka)
|
|---|
| 225 | # must split into pre and post args to be able to supply val of p
|
|---|
| 226 | pre_args=''
|
|---|
| 227 | post_args=''
|
|---|
| 228 | single_run() {
|
|---|
| 229 | affinity ${1}
|
|---|
| 230 | preprint="${1}\t"
|
|---|
| 231 | if [ ${mem} -eq ${true} ] ; then
|
|---|
| 232 | repeat_command /usr/bin/time -f "%M" taskset -c ${taskset} ./a.${hostname} ${pre_args} ${1} ${post_args} > /dev/null
|
|---|
| 233 | else
|
|---|
| 234 | repeat_command taskset -c ${taskset} ./a.${hostname} ${pre_args} ${1} ${post_args}
|
|---|
| 235 | fi
|
|---|
| 236 | }
|
|---|
| 237 |
|
|---|
| 238 | # runs the current bench for all processor vals
|
|---|
| 239 | # works for standard benchs that dont need to set a config file (not Akka or CAF)
|
|---|
| 240 | run_bench() {
|
|---|
| 241 | for p in ${bench_cores} ; do
|
|---|
| 242 | single_run ${p}
|
|---|
| 243 | done
|
|---|
| 244 | }
|
|---|
| 245 |
|
|---|
| 246 | set_akka_parallelism() {
|
|---|
| 247 | sed -i "s/parallelism-min = .*/parallelism-min = ${1}/g" application.conf
|
|---|
| 248 | sed -i "s/parallelism-max = .*/parallelism-max = ${1}/g" application.conf
|
|---|
| 249 | }
|
|---|
| 250 |
|
|---|
| 251 | arch # get hostname
|
|---|
| 252 |
|
|---|
| 253 | # set up leading info for python script
|
|---|
| 254 | echo $numtimes
|
|---|
| 255 | echo $bench_cores
|
|---|
| 256 |
|
|---|
| 257 | for i in ${!names[@]}; do
|
|---|
| 258 | echo -n ${names[$i]}" "
|
|---|
| 259 | done
|
|---|
| 260 | if [ ${runCAF} -eq ${true} ] ; then
|
|---|
| 261 | echo -n 'CAF '
|
|---|
| 262 | fi # done CAF
|
|---|
| 263 | if [ ${runAKKA} -eq ${true} ] ; then
|
|---|
| 264 | echo -n 'Akka '
|
|---|
| 265 | fi # done AKKA
|
|---|
| 266 | if [ ${runUCPP} -eq ${true} ] ; then
|
|---|
| 267 | echo -n 'uC++ '
|
|---|
| 268 | fi # done UCPP
|
|---|
| 269 | if [ ${runPROTO} -eq ${true} ] ; then
|
|---|
| 270 | echo -n 'ProtoActor '
|
|---|
| 271 | fi
|
|---|
| 272 | echo ""
|
|---|
| 273 |
|
|---|
| 274 | # done printing header info for output
|
|---|
| 275 |
|
|---|
| 276 | # cfa flags
|
|---|
| 277 | cfa_flags='-quiet -O3 -nodebug -DNDEBUG'
|
|---|
| 278 |
|
|---|
| 279 | # CAF flags
|
|---|
| 280 | prefixCAF='-O3 -Wall -std=c++17 -I/home/pabuhr/software/nasus/Actors/experiments/CAF/actor-framework/libcaf_core -I/home/pabuhr/software/nasus/Actors/experiments/CAF/actor-framework/libcaf_core/caf -I/home/pabuhr/software/nasus/Actors/experiments/CAF/actor-framework/build/libcaf_core -L/home/pabuhr/software/nasus/Actors/experiments/CAF/actor-framework/build/libcaf_core -L/home/pabuhr/software/nasus/Actors/experiments/CAF/actor-framework/build/libcaf_io'
|
|---|
| 281 | suffixCAF='-lcaf_io -lcaf_core -Wl,-rpath=CAF/actor-framework/build/libcaf_core'
|
|---|
| 282 |
|
|---|
| 283 | # AKKA flags
|
|---|
| 284 | sbtflags="--warn -J-Xmx32g"
|
|---|
| 285 |
|
|---|
| 286 | # UCPP flags
|
|---|
| 287 | UCPPflags="-quiet -g -Wall -Wextra -O3 -nodebug -DNDEBUG -multi"
|
|---|
| 288 | UCPP=~/ucpp/u++-7.0.0/bin/u++
|
|---|
| 289 |
|
|---|
| 290 | # run the benchmarks
|
|---|
| 291 |
|
|---|
| 292 | # /usr/bin/time -f "%Uu %Ss %Er %Mkb"
|
|---|
| 293 | if [ ${executor} -eq ${true} ] ; then
|
|---|
| 294 | if [ ${mem} -eq ${false} ] ; then
|
|---|
| 295 | echo "executor"
|
|---|
| 296 | else
|
|---|
| 297 | echo "mem"
|
|---|
| 298 | fi
|
|---|
| 299 | pre_args="40000 100 ${nRounds}"
|
|---|
| 300 | post_args="${batch}"
|
|---|
| 301 | cd cfa # CFA RUN
|
|---|
| 302 | for i in ${!names[@]}; do
|
|---|
| 303 | print_header ${names[$i]}
|
|---|
| 304 | ${cfa} ${cfa_flags} ${var_flags[$i]} executor.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 305 | run_bench
|
|---|
| 306 | rm a.${hostname}
|
|---|
| 307 | done
|
|---|
| 308 | cd - > /dev/null # done CFA
|
|---|
| 309 | if [ ${runCAF} -eq ${true} ] ; then # CAF RUN
|
|---|
| 310 | cd caf
|
|---|
| 311 | print_header 'CAF'
|
|---|
| 312 | g++ ${prefixCAF} CAFExecutor.cpp ${suffixCAF} -o a.${hostname} > /dev/null 2>&1
|
|---|
| 313 | for p in ${bench_cores} ; do
|
|---|
| 314 | sed -i "s/max-threads = .*/max-threads = ${p}/g" caf-application.conf # set CAF parallelism
|
|---|
| 315 | single_run ${p}
|
|---|
| 316 | done
|
|---|
| 317 | rm a.${hostname}
|
|---|
| 318 | # set back parallelism
|
|---|
| 319 | sed -i "s/max-threads = .*/max-threads = 1/g" caf-application.conf
|
|---|
| 320 | cd - > /dev/null
|
|---|
| 321 | fi # done CAF
|
|---|
| 322 | if [ ${runAKKA} -eq ${true} ] ; then # AKKA RUN
|
|---|
| 323 | cd akka/Executor
|
|---|
| 324 | rm -rf project target # random out of memory errors without this
|
|---|
| 325 | print_header 'Akka'
|
|---|
| 326 | for p in ${bench_cores} ; do
|
|---|
| 327 | set_akka_parallelism ${p}
|
|---|
| 328 | affinity ${p}
|
|---|
| 329 | if [ ${mem} -eq ${true} ] ; then
|
|---|
| 330 | t=1
|
|---|
| 331 | while [ ${t} -le ${numtimes} ] ; do
|
|---|
| 332 | /usr/bin/time -f "%M" taskset -c ${taskset} sbt ${sbtflags} "run ${pre_args} ${p} ${post_args} 1" 2>&1 | grep -v "SLF4J:" | grep -v "Slf4jLogger started" | sed -e "s/\x1b\[0J//" | grep -v "\."
|
|---|
| 333 | t=`expr ${t} + 1`
|
|---|
| 334 | done
|
|---|
| 335 | else
|
|---|
| 336 | taskset -c ${taskset} sbt ${sbtflags} "run ${pre_args} ${p} ${post_args} ${numtimes}" 2>&1 | grep -v "SLF4J:" | grep -v "Slf4jLogger started" | sed -e "s/\x1b\[0J//"
|
|---|
| 337 | fi
|
|---|
| 338 |
|
|---|
| 339 | sbt clean > /dev/null
|
|---|
| 340 | done
|
|---|
| 341 | # set back parallelism
|
|---|
| 342 | set_akka_parallelism 1
|
|---|
| 343 | cd - > /dev/null
|
|---|
| 344 | fi # done AKKA
|
|---|
| 345 | if [ ${runUCPP} -eq ${true} ] ; then # UCPP RUN
|
|---|
| 346 | cd ucpp
|
|---|
| 347 | print_header 'uC++'
|
|---|
| 348 | ${UCPP} ${UCPPflags} uC++Executor.cc -o a.${hostname} > /dev/null 2>&1
|
|---|
| 349 | run_bench
|
|---|
| 350 | rm a.${hostname}
|
|---|
| 351 | cd - > /dev/null
|
|---|
| 352 | fi # done UCPP
|
|---|
| 353 | if [ ${runPROTO} -eq ${true} ] ; then # PROTO RUN
|
|---|
| 354 | print_header 'ProtoActor'
|
|---|
| 355 | cd proto/Executor
|
|---|
| 356 | go build -o a.${hostname} > /dev/null 2>&1
|
|---|
| 357 | run_bench
|
|---|
| 358 | rm a.${hostname}
|
|---|
| 359 | cd - > /dev/null
|
|---|
| 360 | fi # done PROTO
|
|---|
| 361 | echo ""
|
|---|
| 362 | fi
|
|---|
| 363 |
|
|---|
| 364 | if [ ${matrix} -eq ${true} ] ; then
|
|---|
| 365 | echo "matrix"
|
|---|
| 366 | pre_args="${size} ${size} ${size}"
|
|---|
| 367 | post_args=""
|
|---|
| 368 | cd cfa
|
|---|
| 369 | for i in ${!names[@]}; do
|
|---|
| 370 | print_header ${names[$i]}
|
|---|
| 371 | ${cfa} ${cfa_flags} ${var_flags[$i]} matrix.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 372 | run_bench
|
|---|
| 373 | rm a.${hostname}
|
|---|
| 374 | done
|
|---|
| 375 | cd - > /dev/null
|
|---|
| 376 | if [ ${runCAF} -eq ${true} ] ; then # CAF RUN
|
|---|
| 377 | cd caf
|
|---|
| 378 | print_header 'CAF'
|
|---|
| 379 | g++ ${prefixCAF} CAFMatrix.cpp ${suffixCAF} -o a.${hostname} > /dev/null 2>&1
|
|---|
| 380 | for p in ${bench_cores} ; do
|
|---|
| 381 | sed -i "s/max-threads = .*/max-threads = ${p}/g" caf-application.conf # set CAF parallelism
|
|---|
| 382 | single_run ${p}
|
|---|
| 383 | done
|
|---|
| 384 | rm a.${hostname}
|
|---|
| 385 |
|
|---|
| 386 | # set back parallelism
|
|---|
| 387 | sed -i "s/max-threads = .*/max-threads = 1/g" caf-application.conf
|
|---|
| 388 | cd - > /dev/null
|
|---|
| 389 | fi # done CAF
|
|---|
| 390 | if [ ${runAKKA} -eq ${true} ] ; then # AKKA RUN
|
|---|
| 391 | cd akka/Matrix
|
|---|
| 392 | rm -rf project target # random out of memory errors without this
|
|---|
| 393 | print_header 'Akka'
|
|---|
| 394 | for p in ${bench_cores} ; do
|
|---|
| 395 | set_akka_parallelism ${p}
|
|---|
| 396 | affinity ${p}
|
|---|
| 397 | repeat_command taskset -c ${taskset} sbt ${sbtflags} "run ${pre_args} ${p}" 2>&1 | grep -v "SLF4J:" | grep -v "Slf4jLogger started" | sed -e "s/\x1b\[0J//"
|
|---|
| 398 | sbt clean > /dev/null
|
|---|
| 399 | done
|
|---|
| 400 | # set back parallelism
|
|---|
| 401 | set_akka_parallelism 1
|
|---|
| 402 | cd - > /dev/null
|
|---|
| 403 | fi # done AKKA
|
|---|
| 404 | if [ ${runUCPP} -eq ${true} ] ; then # UCPP RUN
|
|---|
| 405 | cd ucpp
|
|---|
| 406 | print_header 'uC++'
|
|---|
| 407 | ${UCPP} ${UCPPflags} uC++Matrix.cc -o a.${hostname} > /dev/null 2>&1
|
|---|
| 408 | run_bench
|
|---|
| 409 | rm a.${hostname}
|
|---|
| 410 | cd - > /dev/null
|
|---|
| 411 | fi # done UCPP
|
|---|
| 412 | if [ ${runPROTO} -eq ${true} ] ; then # PROTO RUN
|
|---|
| 413 | cd proto/Matrix
|
|---|
| 414 | print_header 'ProtoActor'
|
|---|
| 415 | go build -o a.${hostname} > /dev/null 2>&1
|
|---|
| 416 | run_bench
|
|---|
| 417 | rm a.${hostname}
|
|---|
| 418 | cd - > /dev/null
|
|---|
| 419 | fi # done PROTO
|
|---|
| 420 | echo ""
|
|---|
| 421 | fi
|
|---|
| 422 |
|
|---|
| 423 | if [ ${repeat} -eq ${true} ] ; then
|
|---|
| 424 | echo "repeat"
|
|---|
| 425 | pre_args="${messages}"
|
|---|
| 426 | post_args="${n_repeats}"
|
|---|
| 427 | cd cfa
|
|---|
| 428 | for i in ${!names[@]}; do
|
|---|
| 429 | print_header ${names[$i]}
|
|---|
| 430 | ${cfa} ${cfa_flags} ${var_flags[$i]} repeat.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 431 | run_bench
|
|---|
| 432 | rm a.${hostname}
|
|---|
| 433 | done
|
|---|
| 434 | cd - > /dev/null
|
|---|
| 435 | if [ ${runCAF} -eq ${true} ] ; then # CAF RUN
|
|---|
| 436 | cd caf
|
|---|
| 437 | print_header 'CAF'
|
|---|
| 438 | g++ ${prefixCAF} CAFRepeat.cpp ${suffixCAF} -o a.${hostname} > /dev/null 2>&1
|
|---|
| 439 | for p in ${bench_cores} ; do
|
|---|
| 440 | sed -i "s/max-threads = .*/max-threads = ${p}/g" caf-application.conf # set CAF parallelism
|
|---|
| 441 | single_run ${p}
|
|---|
| 442 | done
|
|---|
| 443 | rm a.${hostname}
|
|---|
| 444 |
|
|---|
| 445 | # set back parallelism
|
|---|
| 446 | sed -i "s/max-threads = .*/max-threads = 1/g" caf-application.conf
|
|---|
| 447 | cd - > /dev/null
|
|---|
| 448 | fi # done CAF
|
|---|
| 449 | if [ ${runAKKA} -eq ${true} ] ; then # AKKA RUN
|
|---|
| 450 | cd akka/Repeat
|
|---|
| 451 | rm -rf project target # random out of memory errors without this
|
|---|
| 452 | print_header 'Akka'
|
|---|
| 453 | for p in ${bench_cores} ; do
|
|---|
| 454 | set_akka_parallelism ${p}
|
|---|
| 455 | affinity ${p}
|
|---|
| 456 | repeat_command taskset -c ${taskset} sbt ${sbtflags} "run ${pre_args} ${p} ${post_args}" 2>&1 | grep -v "SLF4J:" | grep -v "Slf4jLogger started" | sed -e "s/\x1b\[0J//"
|
|---|
| 457 | sbt clean > /dev/null
|
|---|
| 458 | done
|
|---|
| 459 | # set back parallelism
|
|---|
| 460 | set_akka_parallelism 1
|
|---|
| 461 | cd - > /dev/null
|
|---|
| 462 | fi # done AKKA
|
|---|
| 463 | if [ ${runUCPP} -eq ${true} ] ; then # UCPP RUN
|
|---|
| 464 | cd ucpp
|
|---|
| 465 | print_header 'uC++'
|
|---|
| 466 | ${UCPP} ${UCPPflags} uC++Repeat.cc -o a.${hostname} > /dev/null 2>&1
|
|---|
| 467 | run_bench
|
|---|
| 468 | rm a.${hostname}
|
|---|
| 469 | cd - > /dev/null
|
|---|
| 470 | fi # done UCPP
|
|---|
| 471 | if [ ${runPROTO} -eq ${true} ] ; then # PROTO RUN
|
|---|
| 472 | print_header 'ProtoActor'
|
|---|
| 473 | cd proto/Repeat
|
|---|
| 474 | go build -o a.${hostname} > /dev/null 2>&1
|
|---|
| 475 | run_bench
|
|---|
| 476 | rm a.${hostname}
|
|---|
| 477 | cd - > /dev/null
|
|---|
| 478 | fi # done PROTO
|
|---|
| 479 | echo ""
|
|---|
| 480 | fi
|
|---|
| 481 |
|
|---|
| 482 |
|
|---|
| 483 | if [ ${static} -eq ${true} ] ; then
|
|---|
| 484 | echo "static"
|
|---|
| 485 | preprint=''
|
|---|
| 486 | cd cfa
|
|---|
| 487 | for i in ${!names[@]}; do
|
|---|
| 488 | echo ${names[$i]}
|
|---|
| 489 | ${cfa} ${cfa_flags} ${var_flags[$i]} static.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 490 | repeat_command taskset -c ${startcore} ./a.${hostname} ${n_static_sends}
|
|---|
| 491 | rm a.${hostname}
|
|---|
| 492 | done
|
|---|
| 493 | cd - > /dev/null
|
|---|
| 494 | if [ ${runCAF} -eq ${true} ] ; then # CAF RUN
|
|---|
| 495 | cd caf
|
|---|
| 496 | echo 'CAF:'
|
|---|
| 497 | g++ ${prefixCAF} CAFSendStatic.cpp ${suffixCAF} -o a.${hostname} > /dev/null 2>&1
|
|---|
| 498 | sed -i "s/max-threads = .*/max-threads = 1/g" caf-application.conf # set CAF parallelism
|
|---|
| 499 | repeat_command taskset -c ${startcore} ./a.${hostname} ${n_static_sends_slow}
|
|---|
| 500 | rm a.${hostname}
|
|---|
| 501 | cd - > /dev/null
|
|---|
| 502 | fi # done CAF
|
|---|
| 503 | if [ ${runAKKA} -eq ${true} ] ; then # AKKA RUN
|
|---|
| 504 | cd akka/SendStatic
|
|---|
| 505 | rm -rf project target # random out of memory errors without this
|
|---|
| 506 | echo 'Akka:'
|
|---|
| 507 | set_akka_parallelism 1
|
|---|
| 508 | taskset -c ${startcore} sbt ${sbtflags} "run ${n_static_sends} ${numtimes}" 2>&1 | grep -v "SLF4J:" | grep -v "Slf4jLogger started" | sed -e "s/\x1b\[0J//"
|
|---|
| 509 | cd - > /dev/null
|
|---|
| 510 | fi # done AKKA
|
|---|
| 511 | if [ ${runUCPP} -eq ${true} ] ; then # UCPP RUN
|
|---|
| 512 | cd ucpp
|
|---|
| 513 | echo 'uC++:'
|
|---|
| 514 | ${UCPP} ${UCPPflags} uC++SendStatic.cc -o a.${hostname} > /dev/null 2>&1
|
|---|
| 515 | repeat_command taskset -c ${startcore} ./a.${hostname} ${n_static_sends}
|
|---|
| 516 | rm a.${hostname}
|
|---|
| 517 | cd - > /dev/null
|
|---|
| 518 | fi # done UCPP
|
|---|
| 519 | if [ ${runPROTO} -eq ${true} ] ; then # PROTO RUN
|
|---|
| 520 | cd proto/SendStatic
|
|---|
| 521 | echo 'ProtoActor:'
|
|---|
| 522 | go build -o a.${hostname} > /dev/null 2>&1
|
|---|
| 523 | repeat_command taskset -c ${startcore} ./a.${hostname} ${n_static_sends}
|
|---|
| 524 | rm a.${hostname}
|
|---|
| 525 | cd - > /dev/null
|
|---|
| 526 | fi # done PROTO
|
|---|
| 527 | echo ""
|
|---|
| 528 | fi
|
|---|
| 529 |
|
|---|
| 530 | if [ ${dynamic} -eq ${true} ] ; then
|
|---|
| 531 | echo "dynamic"
|
|---|
| 532 | cd cfa
|
|---|
| 533 | for i in ${!names[@]}; do
|
|---|
| 534 | echo ${names[$i]}
|
|---|
| 535 | ${cfa} ${cfa_flags} ${var_flags[$i]} dynamic.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 536 | repeat_command taskset -c ${startcore} ./a.${hostname} ${n_dynamic_sends}
|
|---|
| 537 | rm a.${hostname}
|
|---|
| 538 | done
|
|---|
| 539 | cd - > /dev/null
|
|---|
| 540 | if [ ${runCAF} -eq ${true} ] ; then # CAF RUN
|
|---|
| 541 | cd caf
|
|---|
| 542 | echo 'CAF:'
|
|---|
| 543 | g++ ${prefixCAF} CAFSendDynamic.cpp ${suffixCAF} -o a.${hostname} > /dev/null 2>&1
|
|---|
| 544 | sed -i "s/max-threads = .*/max-threads = 1/g" caf-application.conf # set CAF parallelism
|
|---|
| 545 | repeat_command taskset -c ${startcore} ./a.${hostname} ${n_dynamic_sends_slow}
|
|---|
| 546 | rm a.${hostname}
|
|---|
| 547 | cd - > /dev/null
|
|---|
| 548 | fi # done CAF
|
|---|
| 549 | if [ ${runAKKA} -eq ${true} ] ; then # AKKA RUN
|
|---|
| 550 | cd akka/SendDynamic
|
|---|
| 551 | rm -rf project target # random out of memory errors without this
|
|---|
| 552 | echo 'Akka:'
|
|---|
| 553 | set_akka_parallelism 1
|
|---|
| 554 | taskset -c ${startcore} sbt ${sbtflags} "run ${n_dynamic_sends_slow} ${numtimes}" 2>&1 | grep -v "SLF4J:" | grep -v "Slf4jLogger started" | sed -e "s/\x1b\[0J//"
|
|---|
| 555 | cd - > /dev/null
|
|---|
| 556 | fi # done AKKA
|
|---|
| 557 | if [ ${runUCPP} -eq ${true} ] ; then # UCPP RUN
|
|---|
| 558 | cd ucpp
|
|---|
| 559 | echo 'uC++:'
|
|---|
| 560 | ${UCPP} ${UCPPflags} uC++SendDynamic.cc -o a.${hostname} > /dev/null 2>&1
|
|---|
| 561 | repeat_command taskset -c ${startcore} ./a.${hostname} ${n_dynamic_sends}
|
|---|
| 562 | rm a.${hostname}
|
|---|
| 563 | cd - > /dev/null
|
|---|
| 564 | fi # done UCPP
|
|---|
| 565 | if [ ${runPROTO} -eq ${true} ] ; then # PROTO RUN
|
|---|
| 566 | cd proto/SendDynamic
|
|---|
| 567 | echo 'ProtoActor:'
|
|---|
| 568 | go build -o a.${hostname} > /dev/null 2>&1
|
|---|
| 569 | repeat_command taskset -c ${startcore} ./a.${hostname} ${n_dynamic_sends_slow}
|
|---|
| 570 | rm a.${hostname}
|
|---|
| 571 | cd - > /dev/null
|
|---|
| 572 | fi # done PROTO
|
|---|
| 573 | echo ""
|
|---|
| 574 | fi
|
|---|
| 575 |
|
|---|
| 576 |
|
|---|
| 577 | if [ ${balance} -eq ${true} ] ; then
|
|---|
| 578 | cd cfa
|
|---|
| 579 | echo "balance_one"
|
|---|
| 580 | for i in ${!names[@]}; do
|
|---|
| 581 | echo ${names[$i]}':'
|
|---|
| 582 | echo -e $column_headers
|
|---|
| 583 | ${cfa} ${cfa_flags} ${var_flags[$i]} balance.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 584 | for p in ${bench_cores} ; do
|
|---|
| 585 | affinity ${p}
|
|---|
| 586 | preprint="${p}\t"
|
|---|
| 587 | repeat_command taskset -c ${taskset} ./a.${hostname} 32 32 ${nOneRounds} ${p}
|
|---|
| 588 | done
|
|---|
| 589 | rm a.${hostname}
|
|---|
| 590 | done
|
|---|
| 591 | echo ""
|
|---|
| 592 | echo "balance_multi"
|
|---|
| 593 | for i in ${!names[@]}; do
|
|---|
| 594 | echo ${names[$i]}':'
|
|---|
| 595 | echo -e $column_headers
|
|---|
| 596 | ${cfa} ${cfa_flags} ${var_flags[$i]} -DMULTI balance.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 597 | for p in ${bench_cores} ; do
|
|---|
| 598 | affinity ${p}
|
|---|
| 599 | preprint="${p}\t"
|
|---|
| 600 | repeat_command taskset -c ${taskset} ./a.${hostname} 32 32 ${nMultiRounds} ${p}
|
|---|
| 601 | done
|
|---|
| 602 | rm a.${hostname}
|
|---|
| 603 | done
|
|---|
| 604 | echo ""
|
|---|
| 605 | cd - > /dev/null
|
|---|
| 606 | fi
|
|---|
| 607 |
|
|---|
