lustre-iokit/obdfilter-survey/obdfilter-survey

   1 #!/bin/bash
   2 set -e
   3 ######################################################################
   4 # customize per survey
   5
   6 # specify obd instances to exercise
   7 # these can be either...
   8 # obdfilter instances (set 'ost_names')
   9 # ...or...
  10 # echo_client instances (set 'client_names')
  11 # ... use 'host:name' for obd instances on other nodes.
  12 # allow these to be passed in via string...
  13 # OR
  14 # one can specify only case=disk or case=network or case=netdisk through
  15 # command line.
  16
  17 # Perquisite: For "disk" case and "netdisk" case you need to have lustre setup
  18 #             with one or more ost's. For "network" case  you need to have all
  19 #             modules (those llmount.sh loades) loaded in kernel. And the
  20 #             'lctl dl' output must be blank.
  21
  22 # How to run test:
  23 # case 1 (local disk):
  24 #   $ nobjhi=2 thrhi=2 size=1024 case=disk sh obdfilter-survey
  25 #   one can also run test with user defined targets as follows,
  26 #   $ nobjhi=2 thrhi=2 size=1024 targets="lustre-OST0000 lustre-OST0001 ..." sh obdfilter-survey
  27 # case 2 (network):
  28 #   $ nobjhi=2 thrhi=2 size=1024 targets="<name/ip_of_server>" case=network sh obdfilter-survey
  29 #   where, targets is name or ip address of system, which you want to
  30 #   set as server.
  31 # case 3 (network and disk):
  32 #   $ nobjhi=2 thrhi=2 size=1024 case=netdisk sh obdfilter-survey
  33 #   one can also run test with user defined targets as follows,
  34 #   $ nobjhi=2 thrhi=2 size=1024 targets="<osc_name> ..." sh obdfilter-survey
  35 #[ NOTE: It is advised to have automated login (passwordless entry) between server and
  36 #  client systems on which this test runs.]
  37
  38 # include library
  39 source $(dirname $0)/iokit-libecho
  40
  41 # The following variables can be set in the environment, or on the
  42 # command line
  43 # result file prefix (date/time + hostname makes unique)
  44 # NB ensure path to it exists
  45 rslt_loc=${rslt_loc:-"/tmp"}
  46 rslt=${rslt:-"$rslt_loc/obdfilter_survey_$(date +%F@%R)_$(uname -n)"}
  47
  48 # Set this true to check file contents
  49 verify=${verify:-0}
  50
  51 # test targets
  52 targets=${targets:-""}
  53 # test case
  54 case=${case:-"disk"}
  55
  56 # total size (MBytes) per obd instance
  57 # large enough to avoid cache effects
  58 # and to make test startup/shutdown overhead insignificant
  59 size=${size:-16384}
  60
  61 # record size (KBytes) ( 7168 max)
  62 rszlo=${rszlo:-1024}
  63 rszhi=${rszhi:-1024}
  64 rszmax=${rszmax:-4096}
  65
  66 # number of objects per OST
  67 nobjlo=${nobjlo:-1}
  68 #was nobjhi=${nobjhi:-512}
  69 nobjhi=${nobjhi:-16}
  70
  71 # threads per OST (1024 max)
  72 thrlo=${thrlo:-1}
  73 thrhi=${thrhi:-16}
  74
  75 export LC_ALL=POSIX
  76
  77 # End of variables
  78
  79 # create a set of objects, check there are 'n' contiguous ones and
  80 # echo out the first or 'ERROR'
  81 # parameter: 1. hostname
  82 #            2. device number
  83 #            3. number of object to be created (specified by user)
  84 #            4. tempfile name
  85 create_objects () {
  86         local host=$1
  87         local devno=$2
  88         local nobj=$3
  89         local rfile=$4
  90
  91         remote_shell $host $lctl --device $devno create $nobj > $rfile 2>&1
  92         first=0
  93         prev=0
  94         count=0
  95         error=0
  96
  97         # Count number of objects (lines containing " is object id "), and
  98         # ensure that the objects numbers are sequential.
  99         #
 100         exec 3< $rfile
 101         while read -u3 line; do
 102                 case "$line" in
 103                 ( *' is object id '* )
 104                 set -- $line
 105                 if test $(( count += 1 )) -gt 1 ; then
 106                         (( $6 != prev + 1 )) && error=1
 107                 else
 108                         first=$(( $6 + 0 ))
 109                 fi
 110                 prev=$6
 111                 ;;
 112                 esac
 113         done
 114         exec 3<&-
 115
 116         if [ $nobj -ne $count ]; then
 117                 echo "ERROR: $nobj != $count" >&2
 118                 cat $rfile >&2
 119                 echo "ERROR"
 120         elif [ $error -ne 0 ]; then
 121                 echo "ERROR: non contiguous objs found" >&2
 122                 echo ERROR
 123         else
 124                 echo $first
 125         fi
 126         return $error
 127 }
 128
 129 # destroys all objects created in create_objects routine
 130 # parameter: 3. start obj id.
 131 destroy_objects () {
 132         local host=$1
 133         local devno=$2
 134         local obj0=$3
 135         local nobj=$4
 136         local rfile=$5
 137
 138         remote_shell $host $lctl --device $devno destroy $obj0 $nobj > $rfile 2>&1
 139 }
 140
 141 get_stats () {
 142         local rfile=$1
 143
 144         gawk < $rfile                                                   \
 145         '/^Selected device [0-9]+$/ {                                   \
 146                 n = 0;                                                  \
 147                 next;                                                   \
 148         }                                                               \
 149         /error/ {                                                       \
 150                 n = -1;                                                 \
 151                 exit;                                                   \
 152         }                                                               \
 153         /^Total/ {                                                      \
 154                 next;                                                   \
 155         }                                                               \
 156         /^[0-9]+\/[0-9]+ Total: [0-9]+\.[0-9]+\/second$/ {              \
 157                 n++;                                                    \
 158                 v=strtonum($3);                                         \
 159                 if (n == 1 || v < min)                                  \
 160                         min = v;                                        \
 161                 if (n == 1 || v > max)                                  \
 162                         max = v;                                        \
 163                 next;                                                   \
 164         }                                                               \
 165         {                                                               \
 166                 if (n != 0) {                                           \
 167                         n = -1;                                         \
 168                         exit;                                           \
 169                 }                                                       \
 170         }                                                               \
 171         END {                                                           \
 172                 printf "%d %f %f\n", n, min, max                        \
 173         }'
 174 }
 175
 176 get_global_stats () {
 177         local rfile=$1
 178
 179         awk < $rfile                                                    \
 180         'BEGIN {                                                        \
 181                 n = 0;                                                  \
 182         }                                                               \
 183         {                                                               \
 184                 n++;                                                    \
 185                 if (n == 1) {                                           \
 186                         err = $1;                                       \
 187                         min = $2;                                       \
 188                         max = $3;                                       \
 189                 } else {                                                \
 190                         if ($1 < err)                                   \
 191                                 err = $1;                               \
 192                         if ($2 < min)                                   \
 193                                 min = $2;                               \
 194                         if ($3 > max)                                   \
 195                                 max = $3;                               \
 196                 }                                                       \
 197         }                                                               \
 198         END {                                                           \
 199                 if (n == 0)                                             \
 200                         err = 0;                                        \
 201                 printf "%d %f %f\n", err, min, max                      \
 202         }'
 203 }
 204
 205 # enable or disable data check.
 206 # parameter: 1. read/write
 207 testname2type () {
 208         # 'x' disables data check
 209         if ((verify)); then
 210                 x=""
 211         else
 212                 x="x"
 213         fi
 214         case $1 in
 215         *write*)  echo "w$x";;
 216         *)        echo "r$x";;
 217         esac
 218 }
 219
 220 # for "echo_client + obdfilter" case, "prep + commit" mode should be used
 221 # for "echo_client + osc" case, "BRW" mode should be used
 222 testcase2mode() {
 223         case $case in
 224         disk)   echo "p";;
 225         *)      echo "g";;
 226         esac
 227 }
 228
 229 print_summary () {
 230         if [ "$1" = "-n" ]; then
 231                 minusn=$1; shift
 232         else
 233                 minusn=""
 234         fi
 235         echo $minusn "$*" >> $rsltf
 236         echo $minusn "$*"
 237 }
 238
 239 version_code() {
 240         # split arguments like "2.3.61" into "2", "3", "61"
 241         eval set -- $(tr "[:punct:]" " " <<< $*)
 242         echo -n "$((($1 << 16) | ($2 << 8) | $3))"
 243 }
 244
 245 get_lustre_version() {
 246         local host=${1:-${unique_hosts[0]}}
 247         remote_shell $host $lctl get_param -n version |
 248                 awk '/^lustre:/ {print $2}'
 249 }
 250
 251 # Check whether the record size (KBytes) exceeds the maximum bulk I/O RPC size
 252 # or not.
 253 check_record_size() {
 254         [ $(version_code $(get_lustre_version)) -lt $(version_code 2.3.61) ] &&
 255                 rszmax=1024
 256
 257         if [ "$rszhi" -gt "$rszmax" ]; then
 258                 echo "Test disk case support maximum ${rszmax}KB IO data" \
 259                      "(rszhi=$rszhi is too big), please use a smaller value."
 260                 return 1
 261         fi
 262         return 0
 263 }
 264
 265 # Customisation variables
 266 #####################################################################
 267 # One can change variable values in this section as per requirements
 268
 269 if [ -n "$targets" ]; then
 270         declare -a ost_names
 271         declare -a client_names
 272         count=0
 273         for name in $targets; do
 274                 if [ $case == "disk" ]; then
 275                 ost_names[$count]=$name
 276                 else
 277                 client_names[$count]=$name
 278                 fi
 279                 count=$((count + 1))
 280         done
 281 fi
 282
 283 # what tests to run (first must be write)
 284 tests_str=${tests_str:-""}
 285 if [ -n "$tests_str" ]; then
 286         declare -a tests
 287         count=0
 288         for name in $tests_str; do
 289                 tests[$count]=$name
 290                 count=$((count + 1))
 291         done
 292 else
 293         #tests=(write rewrite read reread rewrite_again)
 294         tests=(write rewrite read)
 295 fi
 296
 297 # restart from here iff all are defined
 298 restart_rsz=
 299 restart_thr=1
 300 restart_nobj=1
 301
 302 # machine's page size (KB)
 303 PAGE_SIZE=${PAGE_SIZE:-$(($(getconf PAGE_SIZE) / 1024))}
 304 PAGE_SIZE=${PAGE_SIZE:-4}
 305
 306 # max buffer_mem (total_threads * buffer size)
 307 # (to avoid lctl ENOMEM problems)
 308 max_buffer_mem=$((1024 * 1024))
 309 snap=1
 310 clean_srv_OSS=0
 311 # Customisation variables ends here.
 312 #####################################################################
 313 # leave the rest of this alone unless you know what you're doing...
 314
 315 # check and insert obdecho module
 316 if ! lsmod | grep obdecho > /dev/null; then
 317         modprobe obdecho
 318 fi
 319 if [ ${#tests[@]} -eq 0 -o "${tests[0]}" != "write" ]; then
 320         echo "tests: ${tests[@]}"
 321         echo "First test must be 'write'" 1>&2
 322         exit 1
 323 fi
 324
 325 rsltf="${rslt}.summary"
 326 workf="${rslt}.detail"
 327 cmdsf="${rslt}.script"
 328 vmstatf="${rslt}.vmstat"
 329 echo -n > $rsltf
 330 echo -n > $workf
 331
 332 # hide a little trick to unset this from the command line
 333 if [ "$lustre_root" == " " ]; then
 334         unset lustre_root
 335 fi
 336
 337 if [ -z "$lustre_root" ]; then
 338         lctl=lctl
 339 else
 340         lctl=${lustre_root}/utils/lctl
 341 fi
 342
 343 # split out hostnames from client/ost names
 344 ndevs=0
 345 for trgt in $targets; do
 346         str=($(split_hostname $trgt))
 347         host_names[$ndevs]=${str[0]}
 348         client_names[$ndevs]=${str[1]}
 349         ndevs=$((ndevs + 1))
 350 done
 351 if [ $case == "disk" ]; then
 352         for ((i = 0; i < $ndevs; i++)); do
 353                 ost_names[$i]=${client_names[$i]}
 354         done
 355 fi
 356 if [ $case == "netdisk" ]; then
 357         if [ "$targets" ]; then
 358                 for ((i = 0; i < $ndevs; i++)); do
 359                         setup_osc_for_remote_ost ${host_names[$i]} \
 360                                                  ${client_names[$i]} $i
 361                         osc_name=${client_names[$i]}_osc
 362                         ec_using_osc $osc_name
 363                         cleanup_oscs="$cleanup_oscs $osc_name"
 364                 done
 365         else
 366                 client_names_str=$($lctl dl | grep -v mdt |
 367                         awk '{if ($2 == "UP" && $3 == "osc") {print $4} }')
 368                 count=0;
 369                 for name in $client_names_str; do
 370                         client_names[$count]=$(echo $name | sed 's/-osc-.*$//')
 371                         count=$((count + 1))
 372                 done
 373
 374                 host_names_str=$($lctl dl -t | grep -v mdt |
 375                         awk '{if ($2 == "UP" && $3 == "osc") {print $7} }')
 376                 count=0;
 377                 for name in $host_names_str; do
 378                         host_names[$count]=$(echo $name | sed 's/@.*$//')
 379                         count=$((count + 1))
 380                 done
 381
 382                 for (( i = 0; i < $count; i++ )) do
 383                         setup_osc_for_remote_ost ${host_names[$i]} \
 384                                                  ${client_names[$i]} $i
 385                         osc_name=${client_names[$i]}_osc
 386                         ec_using_osc $osc_name
 387                         cleanup_oscs="$cleanup_oscs $osc_name"
 388                 done
 389         fi
 390
 391         echo_clients=$($lctl dl | grep echo_client |
 392                        awk "{if (\$2 == \"UP\" && \$3 == \"echo_client\") { \
 393                                 print \$4} }")
 394         cnt=0;
 395         for name in $echo_clients; do
 396                 client_names[$cnt]=$name
 397                 host_names[$cnt]=localhost
 398                 cnt=$((cnt + 1))
 399         done
 400         ndevs=${#client_names[@]}
 401 fi
 402 if [ $case == "network" ]; then
 403         server_nid=$targets
 404         if [ -z "$server_nid" ]; then
 405                 echo "Specify hostname or ip-address of server"
 406                 exit 1;
 407         fi
 408         # check for obdecho module on server
 409         if ! dsh $server_nid root "lsmod | grep obdecho > /dev/null"; then
 410                 dsh $server_nid root "modprobe obdecho"
 411         fi
 412         # Now do the server setup
 413         setup_srv_obd $server_nid "echo_srv"
 414         oss_on_srv=$(dsh $server_nid root "$lctl dl | grep OSS" |
 415                      awk '{ print $4 }')
 416         if [ -z $oss_on_srv ]; then
 417                 setup_OSS $server_nid
 418                 clean_srv_OSS=1
 419         fi
 420         if ! dsh $server_nid root "$lctl dl | grep obdecho > /dev/null 2>&1"; then
 421                 echo "obdecho not setup on server"
 422                 exit 1
 423         fi
 424         if ! dsh $server_nid root "$lctl dl | grep ost > /dev/null 2>&1"; then
 425                 echo "ost not setup on server"
 426                 exit 1
 427         fi
 428         # Now start client setup
 429         osc_names_str=$($lctl dl| grep osc | grep -v mdt | grep UP)
 430         if [ -n "$osc_names_str" ]; then
 431                 echo "The existing setup must be cleaned";
 432                 exit 0;
 433         fi
 434         ec_using_srv_nid $server_nid "echotmp" "echotmp_UUID"
 435         client_names[0]="echotmp_ecc"
 436 fi
 437 if [ -z "$targets" ]; then
 438         if [ $case == "disk" ]; then
 439                 get_targets
 440                 ndevs=${#ost_names[@]}
 441         fi
 442 fi
 443 # get vmstat started
 444 # disable portals debug and get obdecho loaded on all relevant hosts
 445 unique_hosts=($(unique ${host_names[@]}))
 446 load_obdechos
 447
 448 if [ $case == "disk" ]; then
 449         check_record_size || cleanup ${PIPESTATUS[0]}
 450 fi
 451
 452 pidcount=0
 453 for host in ${unique_hosts[@]}; do
 454         host_vmstatf=${vmstatf}_${host}
 455         echo -n > $host_vmstatf
 456         remote_shell $host "vmstat 5 >> $host_vmstatf" &> /dev/null &
 457         pid=$!
 458         vmstatpids[$pidcount]=$pid
 459         pidcount=$((pidcount + 1))
 460 done
 461 # get all the echo_client device numbers and names
 462 for ((i=0; i < $ndevs; i++)); do
 463         host=${host_names[$i]}
 464         devno=($(get_ec_devno $host "${client_names[$i]}" "${ost_names[$i]}"))
 465         if ((${#devno[@]} != 3)); then
 466                 exit 1
 467         fi
 468         devnos[$i]=${devno[0]}
 469         client_names[$i]=${devno[1]}
 470         do_teardown_ec[$i]=${devno[2]}
 471 done
 472 if (($ndevs <= 0 || ${#host_names[@]} <= 0)); then
 473         echo "no devices or hosts specified"
 474         cleanup 0 $clean_srv_OSS $cleanup_oscs
 475 fi
 476 # Buffers will be spread out among all hosts, so allow for that
 477 max_buffer_mem=$(( ${max_buffer_mem} * ${#unique_hosts[@]} ))
 478 print_summary "$(date) Obdfilter-survey for case=$case from $(hostname)"
 479 for ((rsz = $rszlo; rsz <= $rszhi; rsz*=2)); do
 480         for ((nobj = $nobjlo; nobj <= $nobjhi; nobj*=2)); do
 481                 for ((thr = $thrlo; thr <= $thrhi; thr*=2)); do
 482                         if ((thr % nobj)); then
 483                                 continue
 484                         fi
 485                         # restart?
 486                         if [ -n "$restart_rsz" -a \
 487                              -n "$restart_nobj" -a \
 488                              -n "$restart_thr" ]; then
 489                                 if ((rsz < restart_rsz ||
 490                                      (rsz == restart_rsz &&
 491                                       (nobj < restart_nobj ||
 492                                        (nobj == restart_nobj &&
 493                                         thr < restart_thr))))); then
 494                                         continue;
 495                                 fi
 496                         fi
 497
 498                         # compute parameters
 499                         total_thr=$((ndevs * thr))
 500                         total_nobj=$((ndevs * nobj))
 501                         pages=$((rsz / PAGE_SIZE))
 502                         actual_rsz=$((pages * PAGE_SIZE))
 503                         count=$((size * 1024 / (actual_rsz * thr)))
 504                         actual_size=$((actual_rsz * count * thr))
 505                         total_size=$((actual_size * ndevs))
 506
 507                         # show computed parameters
 508                         str=$(printf 'ost %2d sz %8dK rsz %4dK obj %4d thr %4d ' \
 509                               $ndevs $total_size $actual_rsz $total_nobj $total_thr)
 510                         echo "=======================> $str" >> $workf
 511                         print_summary -n "$str"
 512                         if ((total_thr * actual_rsz > max_buffer_mem)); then
 513                                 print_summary "Too much buffer space"
 514                                 continue
 515                         fi
 516
 517                         # create the objects
 518                         tmpf="${workf}_tmp"
 519                         for ((idx = 0; idx < $ndevs; idx++)); do
 520                                 host=${host_names[$idx]}
 521                                 devno=${devnos[$idx]}
 522                                 client_name="${host}:${client_names[$idx]}"
 523                                 echo "=============> Create $nobj on $client_name" >> $workf
 524                                 first_obj=$(create_objects $host $devno $nobj $tmpf)
 525                                 cat $tmpf >> $workf
 526                                 rm $tmpf
 527                                 if [ $first_obj = "ERROR" ]; then
 528                                         print_summary "created object #s on $client_name not contiguous"
 529                                         exit 1
 530                                 fi
 531                                 first_objs[$idx]=$first_obj
 532                         done # $ndevs
 533
 534                         # run tests
 535                         for test in ${tests[@]}; do
 536                                 declare -a pidarray
 537                                 for host in ${unique_hosts[@]}; do
 538                                         echo "starting run for test: $test rsz: $rsz " \
 539                                         "threads: $thr objects: $nobj" >> ${vmstatf}_${host}
 540                                 done
 541                                 print_summary -n "$test "
 542
 543                                 # create per-host script files
 544                                 for host in ${unique_hosts[@]}; do
 545                                         echo -n > ${cmdsf}_${host}
 546                                 done
 547                                 for ((idx = 0; idx < $ndevs; idx++)); do
 548                                         host=${host_names[$idx]}
 549                                         devno=${devnos[$idx]}
 550                                         tmpfi="${tmpf}_$idx"
 551                                         first_obj=${first_objs[$idx]}
 552                                         thr_per_obj=$((${thr}/${nobj}))
 553                                         echo >> ${cmdsf}_${host} \
 554                                         "$lctl > $tmpfi 2>&1 \\
 555                                         --threads $thr -$snap $devno \\
 556                                         test_brw $count $(testname2type $test) q $pages \\
 557                                         ${thr_per_obj}t${first_obj} $(testcase2mode)$pages &"
 558                                 done # $ndevs
 559                                 pidcount=0
 560                                 for host in ${unique_hosts[@]}; do
 561                                         echo "wait" >> ${cmdsf}_${host}
 562                                         pidarray[$pidcount]=0
 563                                         pidcount=$((pidcount + 1))
 564                                 done
 565                                 # timed run of all the per-host script files
 566                                 t0=$(date +%s.%N)
 567                                 pidcount=0
 568                                 for host in ${unique_hosts[@]}; do
 569                                         remote_shell $host bash < ${cmdsf}_${host} &
 570                                         pidarray[$pidcount]=$!
 571                                         pidcount=$((pidcount + 1))
 572                                 done
 573                                 pidcount=0
 574                                 for host in ${unique_hosts[@]}; do
 575                                         wait ${pidarray[$pidcount]}
 576                                         pidcount=$((pidcount + 1))
 577                                 done
 578                                 #wait
 579                                 t1=$(date +%s.%N)
 580                                 # clean up per-host script files
 581                                 for host in ${unique_hosts[@]}; do
 582                                         rm ${cmdsf}_${host}
 583                                 done
 584
 585                                 # compute bandwidth from total data / elapsed time
 586                                 str=$(awk "BEGIN {printf \"%7.2f \",\
 587                                 $total_size / (( $t1 - $t0 ) * 1024)}")
 588                                 print_summary -n "$str"
 589                                 # collect/check individual OST stats
 590                                 echo -n > $tmpf
 591                                 for ((idx = 0; idx < $ndevs; idx++)); do
 592                                         client_name="${host_names[$idx]}:${client_names[$idx]}"
 593                                         tmpfi="${tmpf}_$idx"
 594                                         echo "=============> $test $client_name" >> $workf
 595                                         host="${host_names[$idx]}"
 596                                         remote_shell $host cat $tmpfi > ${tmpfi}_local
 597                                         cat ${tmpfi}_local >> $workf
 598                                         get_stats ${tmpfi}_local >> $tmpf
 599                                         rm -f $tmpfi ${tmpfi}_local
 600                                 done # $ndevs
 601
 602                                 # compute/display global min/max stats
 603                                 echo "=============> $test global" >> $workf
 604                                 cat $tmpf >> $workf
 605                                 stats=($(get_global_stats $tmpf))
 606                                 rm $tmpf
 607                                 if ((stats[0] <= 0)); then
 608                                         if ((stats[0] < 0)); then
 609                                                 str=$(printf "%17s " ERROR)
 610                                         else
 611                                                 str=$(printf "%17s " SHORT)
 612                                         fi
 613                                 else
 614                                         str=$(awk "BEGIN {printf \"[%7.2f,%7.2f] \",\
 615                                         (${stats[1]} * $actual_rsz)/1024,\
 616                                         (${stats[2]} * $actual_rsz)/1024; exit}")
 617                                 fi
 618                                 print_summary -n "$str"
 619                         done # $tests[]
 620                         print_summary ""
 621
 622                         # destroy objects we created
 623                         for ((idx = 0; idx < $ndevs; idx++)); do
 624                                 host=${host_names[$idx]}
 625                                 devno=${devnos[$idx]}
 626                                 client_name="${host}:${client_names[$idx]}"
 627                                 first_obj=${first_objs[$idx]}
 628                                 echo "=============> Destroy $nobj on $client_name" >> $workf
 629                                 destroy_objects $host $devno $first_obj $nobj $tmpf
 630                                 cat $tmpf >> $workf
 631                                 rm $tmpf
 632                         done # $ndevs
 633                 done # $thr
 634         done # $nobj
 635 done # $rsz
 636 cleanup 0 $clean_srv_OSS $cleanup_oscs
 637 exit 0