lustre-iokit/ior-survey/ior-survey

   1 #!/bin/bash
   2
   3 # cluster name (all node names are this followed by the node number)
   4 cluster=mdev
   5
   6 # client node numbers (individual numbers or inclusive ranges)
   7 clients=(7-8)
   8
   9 # numbers of clients to survey
  10 clients_lo=1
  11 clients_hi=2
  12 clients_iterator="+=1"
  13
  14 # numbers of tasks per client to survey
  15 tasks_per_client_lo=1
  16 tasks_per_client_hi=8
  17 tasks_per_client_iterator="*=2"
  18
  19 # record sizes to survey
  20 rsize_lo=1M
  21 rsize_hi=1M
  22 rsize_iterator="*=2"
  23
  24 ## which tests to run (first must be write)
  25 # clear_cache)   not really a test; just uncache everything
  26 # *write*)   write
  27 # *)         read
  28 #tests=(write rewrite read reread rewrite_again)
  29 tests=(write rewrite clear_cache read reread)
  30
  31 # total # bytes written/read by any client node
  32 min_per_client_size=4G
  33 min_total_size=8G
  34
  35 # should each task do I/O to its own file?
  36 file_per_task=1
  37
  38 # the binaries
  39 IOR=/usr/local/sbin/IOR
  40 llmount=llmount
  41 pdsh=pdsh
  42
  43 # the result file prefix (date/time + hostname makes unique)
  44 rslt=/tmp/ior_survey_`date +%F@%R`_`uname -n`
  45
  46 # where lustre is mounted on the clients
  47 lustre=/mnt/lustre
  48
  49 # basename of the test file(s)
  50 testfile=${lustre}/ior_survey_testfile
  51
  52 # pdsh args required to instantiate all instances of IOR in parallel
  53 # the chosen module must support '-n <procs-per-node>'
  54 # -R<module>, -f<fanout> etc
  55 pdsh_mpiargs="-Rmqsh"
  56
  57 #don't spin for MPI completions
  58 export LIBELAN_WAITTYPE=0
  59
  60 ################################################################################
  61 # dont change stuff below here unless you know what you're doing...
  62
  63 count_range() {
  64     echo $1 | awk '{ nvals=split($1, vals, "-");\
  65                      if (nvals == 1) print 1;\
  66                      else if (nvals == 2) printf "%d\n", vals[2] - vals[1] + 1;}'
  67 }
  68
  69 base_range() {
  70     echo $1 | awk '{ split($1, vals, "-"); print vals[1]; }'
  71 }
  72
  73 idx2nodenum() {
  74     local n=$1; shift
  75     while ((1)); do
  76         local range=$1; shift
  77         if [ -z "$range" ]; then
  78             return
  79         fi
  80         chunk=`count_range $range`
  81         if ((chunk > n)); then
  82             base=`base_range $range`
  83             echo $((base + n))
  84             return
  85         fi
  86         n=$((n-chunk))
  87     done
  88 }
  89
  90 n2noderange() {
  91     local n=$1; shift
  92     sep=""
  93     nodes="["
  94     while ((n > 0)); do
  95         local range=$1; shift
  96         if [ -z "$range" ]; then
  97             return
  98         fi
  99         local base=`base_range $range`
 100         local chunk=`count_range $range`
 101         if ((chunk > n)); then chunk=n; fi
 102         local nodes="${nodes}${sep}${base}"; sep=","
 103         if ((chunk > 1)); then nodes="${nodes}-$((base+chunk-1))"; fi
 104         n=$((n-chunk))
 105     done
 106     echo "${nodes}]"
 107 }
 108
 109 countnodes() {
 110     local radix=16384
 111     local n=0
 112     while ((radix > 0)); do
 113         local nodes=`n2noderange $((n+radix)) $@`
 114         if [ -n "$nodes" ]; then
 115             n=$((n+radix))
 116         fi
 117         radix=$((radix/2))
 118     done
 119     echo $n
 120 }
 121
 122 parse_number() {
 123     local str=$1
 124     case $str in
 125         *G|*g) n=`echo $str | sed 's/[gG]//'`; echo $((n*1024*1024*1024));;
 126         *M|*m) n=`echo $str | sed 's/[Mm]//'`; echo $((n*1024*1024));;
 127         *K|*k) n=`echo $str | sed 's/[Kk]//'`; echo $((n*1024));;
 128         *)     echo $1;;
 129     esac
 130 }
 131
 132 pp_number() {
 133     local n=$1
 134     local G=$((1024*1024*1024))
 135     local M=$((1024*1024))
 136     local K=$((1024))
 137     if ((n%G == 0 && n >= G)); then
 138         echo "$((n/G))G"
 139     elif ((n%M == 0 && n >= M)); then
 140         echo "$((n/M))M"
 141     elif ((n%K == 0 && n >= K)); then
 142         echo "$((n/K))K"
 143     else
 144         echo $n
 145     fi
 146 }
 147
 148 if [ ${#tests[@]} -eq 0 -o "${tests[0]}" != "write" ]; then
 149     echo "First test must be 'write'" 1>&2
 150     exit 1
 151 fi
 152
 153 rsltf="${rslt}.summary"
 154 workf="${rslt}.detail"
 155 echo -n > $rsltf
 156 echo -n > $workf
 157
 158 print_summary () {
 159     if [ "$1" = "-n" ]; then
 160         minusn=$1; shift
 161     else
 162         minusn=""
 163     fi
 164     echo $minusn "$*" >> $rsltf
 165     echo $minusn "$*"
 166 }
 167
 168 # convert params to actual numbers
 169 min_per_client_size=`parse_number $min_per_client_size`
 170 min_total_size=`parse_number $min_total_size`
 171
 172 rsize_lo=`parse_number $rsize_lo`
 173 rsize_hi=`parse_number $rsize_hi`
 174
 175 # check on actual numbers of client nodes
 176 nclients=`countnodes ${clients[@]}`
 177 if ((clients_hi > nclients)); then clients_hi=$nclients; fi
 178
 179 for ((rsize=rsize_lo; rsize<=rsize_hi; rsize$rsize_iterator)); do
 180     pp_rsize=`pp_number $rsize`
 181
 182     for ((nclnt=clients_lo; nclnt<=clients_hi; nclnt$clients_iterator)); do
 183         test_clients="${cluster}`n2noderange $nclnt ${clients[@]}`"
 184
 185         per_client_size=$((min_total_size/nclnt))
 186         if ((per_client_size < min_per_client_size)); then
 187             per_client_size=$min_per_client_size
 188         fi
 189
 190         for ((ntask=tasks_per_client_lo; ntask <= tasks_per_client_hi; ntask$tasks_per_client_iterator)); do
 191             per_task_size=$((per_client_size/ntask))
 192             if ((per_task_size%rsize != 0)); then
 193                 per_task_size=$(((per_task_size/rsize + 1)*rsize))
 194             fi
 195             total_size=`pp_number $((per_task_size*nclnt*ntask))`
 196
 197             hdrstr=`printf "Total: %5sB rsize: %4sB clients: %4d tasks: %3d: " \
 198                 $total_size $pp_rsize $nclnt $ntask`
 199             print_summary -n "$hdrstr"
 200
 201             for ((test_idx=0; test_idx < ${#tests[@]}; test_idx++)); do
 202                 test=${tests[$test_idx]}
 203
 204                 print_summary -n "$test "
 205                 echo "===========> ${hdrstr} on $test_clients doing $test" >> $workf
 206                 tmpf=${workf}_tmp
 207                 echo -n > $tmpf
 208
 209                 if [ "$test" = "clear_cache" ]; then
 210                     clear_cache='for LRU in /proc/fs/lustre/ldlm/namespaces/*/lru_size; do; echo clear > $LRU; done'
 211                     echo "=> $clear_cache" >> $tmpf
 212                     $pdsh -S -b -w "$test_clients" >> $tmpf 2>&1 \
 213                         "$clear_cache"
 214                     status=$?
 215                     echo "Completion Status: $status" >> $tmpf
 216
 217                     if ((status)); then
 218                         result="ERROR"
 219                     else
 220                         result="OK"
 221                     fi
 222                 else
 223                     # check lustre is mounted everywhere it's needed
 224                     cmd="(mount -t lustre; mount -t lustre_lite) | grep $lustre"
 225                     echo "=> Mount Check: $cmd" >> $tmpf
 226                     $pdsh -S -b -w "$test_clients" >> $tmpf 2>&1 \
 227                         "$cmd"
 228                     status=$?
 229                     echo "Completion Status: $status" >> $tmpf
 230                     if ((status)); then
 231                         cat $tmpf >> $workf
 232                         rm $tmpf
 233                         print_summary "Lustre NOT mounted on $lustre somewhere"
 234                         exit 1
 235                     fi
 236
 237                     cmdline=(
 238                     $IOR                     # the command
 239                     -o${testfile}            # test file prefix
 240                     -b${per_task_size}       # bytes per task
 241                     -t${rsize}               # record size
 242                     -e                       # fsync before close
 243                     -q                       # quit on error
 244                     )
 245
 246                     idx=${#cmdline[@]}
 247
 248                     # keep the test file(s) unless this is the last test
 249                     ((test_idx < ${#tests[@]}-1)) && cmdline[$((idx++))]="-k"
 250
 251                     # use the existing test file(s) unless this is the first test
 252                     ((test_idx > 0)) && cmdline[$((idx++))]="-E"
 253
 254                     # file-per-task
 255                     ((file_per_task)) && cmdline[$((idx++))]="-F"
 256
 257                     case "$test" in
 258                     *write*) cmdline[$((idx++))]="-w"
 259                              awkstr="Max Write";;
 260                     *)       cmdline[$((idx++))]="-r"
 261                              awkstr="Max Read";;
 262                     esac
 263
 264                     echo "=> ${cmdline[@]}" >> $tmpf
 265
 266                     $pdsh -S -b $pdsh_mpiargs -w "$test_clients" -n $ntask >> $tmpf 2>&1 \
 267                         "${cmdline[@]}"
 268                     status=$?
 269
 270                     echo "Completion Status: $status" >> $tmpf
 271
 272                     if ((status)); then
 273                         result="ERROR"
 274                     else
 275                         result=`awk < $tmpf "/$awkstr/ {print $ 3; found=1; exit}\
 276                                              END       {if (!found) print \"ERROR\"}"`
 277                     fi
 278                 fi
 279
 280                 cat $tmpf >> $workf
 281                 rm $tmpf
 282
 283                 str=`printf "%8s" "$result"`
 284                 print_summary -n "$str "
 285             done
 286             print_summary ""
 287         done
 288     done
 289 done
 290