| 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} + ${filled} - 1 ))"
|
|---|
| 83 | fi
|
|---|
| 84 | fi
|
|---|
| 85 | }
|
|---|
| 86 |
|
|---|
| 87 | # numtimes=5
|
|---|
| 88 | numtimes=1
|
|---|
| 89 |
|
|---|
| 90 | # num_threads='2 4 8 16 24 32'
|
|---|
| 91 | # side_chan_threads='6 12 18 24 30' # must be mults of 6
|
|---|
| 92 | num_threads='2'
|
|---|
| 93 | side_chan_threads='6'
|
|---|
| 94 |
|
|---|
| 95 | chan_size='10'
|
|---|
| 96 | future_time='10'
|
|---|
| 97 |
|
|---|
| 98 | # toggle benchmarks
|
|---|
| 99 | spin=${true}
|
|---|
| 100 | contend=${true}
|
|---|
| 101 | sidechan=${true}
|
|---|
| 102 | future=${true}
|
|---|
| 103 | spin=${false}
|
|---|
| 104 | contend=${false}
|
|---|
| 105 | sidechan=${false}
|
|---|
| 106 | # future=${false}
|
|---|
| 107 |
|
|---|
| 108 | runCFA=${true}
|
|---|
| 109 | runGO=${true}
|
|---|
| 110 | runUCPP=${true}
|
|---|
| 111 | # runCFA=${false}
|
|---|
| 112 | runGO=${false}
|
|---|
| 113 | # runUCPP=${false}
|
|---|
| 114 |
|
|---|
| 115 | cfa=~/cfa-cc/driver/cfa
|
|---|
| 116 |
|
|---|
| 117 | # Helpers to minimize code duplication
|
|---|
| 118 |
|
|---|
| 119 | # repeats a command ${numtimes}
|
|---|
| 120 | preprint=''
|
|---|
| 121 | repeat_command() {
|
|---|
| 122 | t=1
|
|---|
| 123 | while [ ${t} -le ${numtimes} ] ; do
|
|---|
| 124 | echo -n -e ${preprint}
|
|---|
| 125 | "${@}"
|
|---|
| 126 | t=`expr ${t} + 1`
|
|---|
| 127 | done
|
|---|
| 128 | }
|
|---|
| 129 |
|
|---|
| 130 | # prints the leading info for a given run of a variant
|
|---|
| 131 | print_header() {
|
|---|
| 132 | echo ${1}':'
|
|---|
| 133 | echo -e "cores\tthroughput (entries)"
|
|---|
| 134 | }
|
|---|
| 135 |
|
|---|
| 136 | # runs the current benchmark with provided args
|
|---|
| 137 | # only works for standard-run benchmarks (not Akka)
|
|---|
| 138 | # must split into pre and post args to be able to supply val of p
|
|---|
| 139 | pre_args=''
|
|---|
| 140 | post_args=''
|
|---|
| 141 | single_run() {
|
|---|
| 142 | affinity ${1}
|
|---|
| 143 | preprint="${1}\t"
|
|---|
| 144 | repeat_command taskset -c ${taskset} ./a.${hostname} ${pre_args} ${1} ${post_args}
|
|---|
| 145 | }
|
|---|
| 146 |
|
|---|
| 147 | # runs the current bench for all processor vals
|
|---|
| 148 | # works for standard benchs that dont need to set a config file (not Akka or CAF)
|
|---|
| 149 | run_bench() {
|
|---|
| 150 | for p in ${num_threads} ; do
|
|---|
| 151 | single_run ${p}
|
|---|
| 152 | done
|
|---|
| 153 | }
|
|---|
| 154 |
|
|---|
| 155 | run_side_chan() {
|
|---|
| 156 | i=1
|
|---|
| 157 | for p in ${side_chan_threads} ; do
|
|---|
| 158 | affinity ${p}
|
|---|
| 159 | preprint="${p}\t"
|
|---|
| 160 | repeat_command taskset -c ${taskset} ./a.${hostname} ${pre_args} ${i} ${post_args}
|
|---|
| 161 | i=`expr ${i} + 1`
|
|---|
| 162 | done
|
|---|
| 163 | }
|
|---|
| 164 |
|
|---|
| 165 | run_future() {
|
|---|
| 166 | for p in ${num_threads} ; do
|
|---|
| 167 | pre_args=$(( ${p} - 1 ))
|
|---|
| 168 | affinity ${p}
|
|---|
| 169 | preprint="${p}\t"
|
|---|
| 170 | repeat_command taskset -c ${taskset} ./a.${hostname} ${pre_args} ${post_args}
|
|---|
| 171 | done
|
|---|
| 172 | }
|
|---|
| 173 |
|
|---|
| 174 | arch # get hostname
|
|---|
| 175 |
|
|---|
| 176 | # set up leading info for python script
|
|---|
| 177 | echo $numtimes
|
|---|
| 178 | echo $num_threads
|
|---|
| 179 | echo $side_chan_threads
|
|---|
| 180 |
|
|---|
| 181 | if [ ${runCFA} -eq ${true} ]; then
|
|---|
| 182 | echo -n 'CFA '
|
|---|
| 183 | fi
|
|---|
| 184 | if [ ${runGO} -eq ${true} ]; then
|
|---|
| 185 | echo -n 'Go '
|
|---|
| 186 | fi
|
|---|
| 187 | echo ""
|
|---|
| 188 |
|
|---|
| 189 | # done printing header info for output
|
|---|
| 190 |
|
|---|
| 191 | # cfa flags
|
|---|
| 192 | cfa_flags='-quiet -O3 -nodebug -DNDEBUG'
|
|---|
| 193 |
|
|---|
| 194 | # UCPP flags
|
|---|
| 195 | UCPPflags="-quiet -g -Wall -Wextra -O3 -nodebug -DNDEBUG -multi"
|
|---|
| 196 | UCPP=~/ucpp/u++-7.0.0/bin/u++
|
|---|
| 197 |
|
|---|
| 198 | # run the benchmarks
|
|---|
| 199 |
|
|---|
| 200 | run_contend() {
|
|---|
| 201 | post_args=${1}
|
|---|
| 202 |
|
|---|
| 203 | if [ ${runCFA} -eq ${true} ] ; then
|
|---|
| 204 | cd cfa # CFA RUN
|
|---|
| 205 | print_header 'CFA'
|
|---|
| 206 | ${cfa} ${cfa_flags} '-DNUM_CHANS='${3} ${2}.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 207 | run_bench
|
|---|
| 208 | rm a.${hostname}
|
|---|
| 209 | cd - > /dev/null
|
|---|
| 210 | fi # done CFA
|
|---|
| 211 |
|
|---|
| 212 | if [ ${runGO} -eq ${true} ] ; then
|
|---|
| 213 | cd go/${2}${3} # Go RUN
|
|---|
| 214 | print_header 'Go'
|
|---|
| 215 | go build -o a.${hostname} > /dev/null 2>&1
|
|---|
| 216 | run_bench
|
|---|
| 217 | rm a.${hostname}
|
|---|
| 218 | cd - > /dev/null
|
|---|
| 219 | fi # done Go
|
|---|
| 220 | }
|
|---|
| 221 |
|
|---|
| 222 | # /usr/bin/time -f "%Uu %Ss %Er %Mkb"
|
|---|
| 223 | if [ ${contend} -eq ${true} ] ; then
|
|---|
| 224 | echo "contend2: "
|
|---|
| 225 | run_contend ${chan_size} 'contend' '2'
|
|---|
| 226 | echo "contend4: "
|
|---|
| 227 | run_contend ${chan_size} 'contend' '4'
|
|---|
| 228 | echo "contend8: "
|
|---|
| 229 | run_contend ${chan_size} 'contend' '8'
|
|---|
| 230 | fi
|
|---|
| 231 |
|
|---|
| 232 | if [ ${spin} -eq ${true} ] ; then
|
|---|
| 233 | echo "spin2: "
|
|---|
| 234 | run_contend ${chan_size} 'spin' '2'
|
|---|
| 235 | echo "spin4: "
|
|---|
| 236 | run_contend ${chan_size} 'spin' '4'
|
|---|
| 237 | echo "spin8: "
|
|---|
| 238 | run_contend ${chan_size} 'spin' '8'
|
|---|
| 239 | fi
|
|---|
| 240 |
|
|---|
| 241 | if [ ${sidechan} -eq ${true} ] ; then
|
|---|
| 242 | echo "sidechan: "
|
|---|
| 243 | post_args=${chan_size}
|
|---|
| 244 | if [ ${runCFA} -eq ${true} ] ; then
|
|---|
| 245 | cd cfa # CFA RUN
|
|---|
| 246 | print_header 'CFA'
|
|---|
| 247 | ${cfa} ${cfa_flags} sidechan.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 248 | run_side_chan
|
|---|
| 249 | rm a.${hostname}
|
|---|
| 250 | cd - > /dev/null
|
|---|
| 251 | fi # done CFA
|
|---|
| 252 |
|
|---|
| 253 | if [ ${runGO} -eq ${true} ] ; then
|
|---|
| 254 | cd go/sidechan
|
|---|
| 255 | print_header 'Go'
|
|---|
| 256 | go build -o a.${hostname} > /dev/null 2>&1
|
|---|
| 257 | run_side_chan
|
|---|
| 258 | rm a.${hostname}
|
|---|
| 259 | cd - > /dev/null
|
|---|
| 260 | fi # done Go
|
|---|
| 261 | fi
|
|---|
| 262 |
|
|---|
| 263 | if [ ${future} -eq ${true} ] ; then
|
|---|
| 264 | echo "future: "
|
|---|
| 265 | post_args=${future_time}
|
|---|
| 266 | if [ ${runCFA} -eq ${true} ] ; then
|
|---|
| 267 | cd cfa # CFA RUN
|
|---|
| 268 | print_header 'CFA'
|
|---|
| 269 | ${cfa} ${cfa_flags} future.cfa -o a.${hostname} > /dev/null 2>&1
|
|---|
| 270 | run_future
|
|---|
| 271 | rm a.${hostname}
|
|---|
| 272 | cd - > /dev/null
|
|---|
| 273 | fi # done CFA
|
|---|
| 274 |
|
|---|
| 275 | if [ ${runUCPP} -eq ${true} ] ; then
|
|---|
| 276 | cd ucpp
|
|---|
| 277 | print_header 'uC++'
|
|---|
| 278 | ${UCPP} ${UCPPflags} future.cc -o a.${hostname} > /dev/null 2>&1
|
|---|
| 279 | run_future
|
|---|
| 280 | rm a.${hostname}
|
|---|
| 281 | cd - > /dev/null
|
|---|
| 282 | fi # done Go
|
|---|
| 283 | fi
|
|---|
| 284 |
|
|---|
