# this is a quick and dirty migration of runemomniagg2.sh to the # --enable-demo mode of aggregate testing function kill_netperfs { pkill -ALRM netperf pgrep -P 1 -f netperf > /dev/null while [ $? -eq 0 ] do sleep 1 pgrep -P 1 -f netperf > /dev/null done } function run_cmd { NOW=`date +%s.%N` echo "Starting netperfs at $NOW for $TEST" | tee $TESTLOG i=0; # the starting point for our load level pauses PAUSE_AT=1 while [ $i -lt $MAX_INSTANCES ] do TARGET=${REMOTE_HOSTS[`expr $i % $NUM_REMOTE_HOSTS`]} echo "Starting netperfs on localhost targeting ${TARGET} for $TEST" | tee -a $TESTLOG id=`printf "%.5d" $i` $NETPERF -H $TARGET $NETPERF_CMD 2>&1 > netperf_${TEST}_${id}_to_${TARGET}.out & # give it a moment to get going sleep 1 i=`expr $i + 1` if [ $i -eq $PAUSE_AT ] && [ $i -ne $MAX_INSTANCES ] then NOW=`date +%s.%N` echo "Pausing for $DURATION seconds at $NOW with $i netperfs running for $TEST" | tee -a $TESTLOG sleep $DURATION PAUSE_AT=`expr $PAUSE_AT \* 2` NOW=`date +%s.%N` echo "Resuming at $NOW for $TEST" | tee -a $TESTLOG fi done NOW=`date +%s.%N` echo "Netperfs started by $NOW for $TEST" | tee -a $TESTLOG #wait for our test duration sleep $DURATION #kludgey but this sleep should mean that another interim result will be emitted sleep 3 # stop all the netperfs NOW=`date +%s.%N` echo "Netperfs stopping $NOW for $TEST" | tee -a $TESTLOG kill_netperfs NOW=`date +%s.%N` echo "Netperfs stopped $NOW for $TEST" | tee -a $TESTLOG } # very much like run_cmd, but it runs the tests one at a time rather # than in parallel. We keep the same logging strings to be compatible # (hopefully) with the post processing script, even though they don't # make all that much sense :) function run_cmd_serial { NOW=`date +%s.%N` echo "Starting netperfs at $NOW for $TEST" | tee $TESTLOG i=0; # the starting point for our load level pauses PAUSE_AT=1 while [ $i -lt $NUM_REMOTE_HOSTS ] do TARGET=${REMOTE_HOSTS[`expr $i % $NUM_REMOTE_HOSTS`]} echo "Starting netperfs on localhost targeting ${TARGET} for $TEST" | tee -a $TESTLOG id=`printf "%.5d" $i` $NETPERF -H $TARGET $NETPERF_CMD 2>&1 > netperf_${TEST}_${id}_to_${TARGET}.out & # give it a moment to get going sleep 1 i=`expr $i + 1` NOW=`date +%s.%N` echo "Pausing for $DURATION seconds at $NOW with $i netperfs running for $TEST" | tee -a $TESTLOG # the plus two is to make sure we have a full set of interim # results. probably not necessary here but we want to be # certain sleep `expr $DURATION + 1` kill_netperfs NOW=`date +%s.%N` THEN=`echo $NOW | awk -F "." '{printf("%d.%d",$1-1,$2)}'` echo "Resuming at $THEN for $TEST" | tee -a $TESTLOG done NOW=`date +%s.%N` echo "Netperfs started by $NOW for $TEST" | tee -a $TESTLOG # stop all the netperfs - of course actually they have all been # stopped already, we just want the log entries NOW=`date +%s.%N` echo "Netperfs stopping $NOW for $TEST" | tee -a $TESTLOG kill_netperfs NOW=`date +%s.%N` echo "Netperfs stopped $NOW for $TEST" | tee -a $TESTLOG } # here then is the "main" part if [ ! -f ./remote_hosts ] then echo "This script requires a remote_hosts file" exit -1 fi . ./remote_hosts # how many processors are there on this system NUM_CPUS=`grep processor /proc/cpuinfo | wc -l` # the number of netperf instances we will run will be up to 2x the # number of CPUs MAX_INSTANCES=`expr $NUM_CPUS \* 2` # but at least as many as there are entries in remote_hosts if [ $MAX_INSTANCES -lt $NUM_REMOTE_HOSTS ] then MAX_INSTANCES=$NUM_REMOTE_HOSTS fi # allow the netperf binary to be used to be overridden NETPERF=${NETPERF:="netperf"} if [ $NUM_REMOTE_HOSTS -lt 2 ] then echo "The list of remote hosts is too short. There must be at least 2." exit -1 fi # we assume that netservers are already running on all the load generators DURATION=120 # do not have a uuidgen? then use the one in netperf MY_UUID=`uuidgen` # with top-of-trunk we could make this 0 and run forever # but two hours is something of a failsafe if the signals # get lost LENGTH="-l 7200" OUTPUT="-o all" DO_STREAM=1; DO_MAERTS=1; # NOTE! The Bidir test depends on being able to set a socket buffer # size greater than 13 * 64KB or 832 KB or there is a risk of the test # hanging. If you are running linux, make certain that # net.core.[r|w]mem_max are sufficiently large DO_BIDIR=1; DO_RRAGG=1; DO_RR=1; DO_ANCILLARY=1; # UDP_RR for TPC/PPS using single-byte transactions. we do not use # TCP_RR any longer because any packet losses or other matters # affecting the congestion window will break our desire that there be # a one to one correspondence between requests/responses and packets. if [ $DO_RRAGG -eq 1 ]; then BURST=`find_max_burst.sh ${REMOTE_HOSTS[0]}` if [ $BURST -eq -1 ]; then # use a value that find_max_burst will not have picked BURST=9 echo "find_max_burst.sh returned -1 so picking a burst of $BURST" fi TEST="tps" TESTLOG="netperf_tps.log" NETPERF_CMD="-D 0.5 -c -C -f x -P 0 -t omni $LENGTH -v 2 -- -r 1 -b $BURST -e 1 -T udp -u $MY_UUID $OUTPUT" run_cmd fi # Bidirectional using burst-mode TCP_RR and large request/response size if [ $DO_BIDIR -eq 1 ]; then TEST="bidirectional" TESTLOG="netperf_bidirectional.log" NETPERF_CMD="-D 0.5 -c -C -f m -P 0 -t omni $LENGTH -v 2 -- -r 64K -s 1M -S 1M -b 12 -u $MY_UUID $OUTPUT" run_cmd fi # TCP_STREAM aka outbound with a 64K send size # the netperf command is everything but netperf -H mumble if [ $DO_STREAM -eq 1 ];then TEST="outbound" TESTLOG="netperf_outbound.log" NETPERF_CMD="-D 0.5 -c -C -f m -P 0 -t omni $LENGTH -v 2 -- -m 64K -u $MY_UUID $OUTPUT" run_cmd fi # TCP_MAERTS aka inbound with a 64K send size - why is this one last? # because presently when I pkill the netperf of a "MAERTS" test, the # netserver does not behave well and it may not be possible to get it # to behave well. but we will still have all the interim results even # if we don't get the final results, the useful parts of which will be # the same as the other tests anyway if [ $DO_MAERTS -eq 1 ]; then TEST="inbound" TESTLOG="netperf_inbound.log" NETPERF_CMD="-D 0.5 -c -C -f m -P 0 -t omni $LENGTH -v 2 -- -m ,64K -u $MY_UUID $OUTPUT" run_cmd fi # A single-stream of synchronous, no-burst TCP_RR in an "aggregate" # script? Yes, because the way the aggregate tests work, while there # is a way to see what the performance of a single bulk transfer was, # there is no way to see a basic latency - by the time # find_max_burst.sh has completed, we are past a burst size of 0 if [ $DO_RR -eq 1 ]; then if [ $DURATION -lt 60 ]; then DURATION=60 fi TEST="sync_tps" TESTLOG="netperf_sync_tps.log" NETPERF_CMD="-D 0.5 -c -C -f x -P 0 -t omni $LENGTH -v 2 -- -r 1 -u $MY_UUID $OUTPUT" run_cmd_serial fi # now some ancillary things which may nor may not work on your platform if [ $DO_ANCILLARY -eq 1 ];then dmidecode 2>&1 > dmidecode.txt uname -a 2>&1 > uname.txt cat /proc/cpuinfo 2>&1 > cpuinfo.txt cat /proc/meminfo 2>&1 > meminfo.txt ifconfig -a 2>&1 > ifconfig.txt netstat -rn 2>&1 > netstat.txt dpkg -l 2>&1 > dpkg.txt rpm -qa 2>&1 > rpm.txt cat /proc/interrupts 2>&1 > interrupts.txt i=0 while [ $i -lt `expr $NUM_REMOTE_HOSTS - 1` ] do traceroute ${REMOTE_HOSTS[$i]} > traceroute_${REMOTE_HOSTS[$i]}.txt i=`expr $i + 1` done fi