lustre-iokit/sgpdd-survey/sgpdd-survey

   1 #!/bin/bash
   2
   3 ######################################################################
   4 # customize per survey
   5
   6 # the SCSI devices to measure
   7 scsidevs=${scsidevs:-"/dev/sde /dev/sdh"}
   8
   9 # result file prefix.  date/time+hostname makes unique
  10 # NB ensure the path exists if it includes subdirs
  11 rslt=${rslt:-"/tmp/sgpdd_survey_`date +%F@%R`_`uname -n`"}
  12
  13 # what to do (read or write)
  14 actions="write read"
  15
  16 # total size per device (MBytes)
  17 # NB bigger than device cache is good
  18 size=8192
  19
  20 # record size (KBytes)
  21 rszlo=1024
  22 rszhi=1024
  23
  24 # Concurrent regions per device
  25 crglo=${crglo:-1}
  26 crghi=${crghi:-256}
  27
  28 # threads to share between concurrent regions per device
  29 # NB survey skips over #thr < #regions and #thr/#regions > SG_MAX_QUEUE
  30 thrlo=${thrlo:-1}
  31 thrhi=${thrhi:-4096}
  32
  33 #####################################################################
  34 # leave the rest of this alone unless you know what you're doing...
  35
  36 # sgp_dd's idea of disk sector size (Bytes)
  37 bs=512
  38 # and max # threads one instance will spawn
  39 SG_MAX_QUEUE=16
  40
  41 # map given device names into SG device names
  42 i=0
  43 devs=()
  44 for d in $scsidevs; do
  45     devs[$i]=`sg_map | awk "{if ($ 2 == \"$d\") print $ 1}"`
  46     if [ -z "$devs[$i]" ]; then
  47         echo "Can't find SG device for $d"
  48         exit 1
  49     fi
  50     i=$((i+1))
  51 done
  52 ndevs=${#devs[@]}
  53
  54 rsltf=${rslt}.summary
  55 workf=${rslt}.detail
  56 echo -n > $rsltf
  57 echo -n > $workf
  58
  59 print_summary () {
  60     if [ "$1" = "-n" ]; then
  61         minusn=$1; shift
  62     else
  63         minusn=""
  64     fi
  65     echo $minusn "$*" >> $rsltf
  66     echo $minusn "$*"
  67 }
  68
  69 for ((rsz=$rszlo;rsz<=$rszhi;rsz*=2)); do
  70     for ((crg=$crglo;crg<=$crghi;crg*=2)); do
  71         for ((thr=$thrlo;thr<=$thrhi;thr*=2)); do
  72             if ((thr < crg || thr/crg > SG_MAX_QUEUE)); then
  73                 continue
  74             fi
  75             # compute parameters
  76             bpt=$((rsz*1024/bs))
  77             blocks=$((size*((1024*1024)/bs)/crg))
  78             count=$blocks
  79             # show computed parameters
  80             actual_rsz=$((bpt*bs/1024))
  81             actual_size=$((bs*count*crg/1024))
  82             str=`printf 'total_size %8dK rsz %4d crg %5d thr %5d ' \
  83                          $((actual_size*ndevs)) $actual_rsz $((crg*ndevs)) $((thr*ndevs))`
  84             echo "==============> $str" >> $workf
  85             print_summary -n "$str"
  86             freemem=`awk < /proc/meminfo '/^MemTotal:/ {printf "%d\n", $2}'`
  87             if (((actual_rsz*thr/crg + 64)*crg*ndevs > freemem)); then
  88                 print_summary "ENOMEM"
  89                 continue
  90             fi
  91             # run tests
  92             for action in $actions; do
  93                 print_summary -n "$action "
  94                 echo "=====> $action" >> $workf
  95                 tmpf=${workf}_tmp
  96                 # start test
  97                 t0=`date +%s.%N`
  98                 for ((i=0;i<ndevs;i++)); do
  99                     dev=${devs[i]}
 100                     if [ $action = read ]; then
 101                         inf="if=$dev"
 102                         outf="of=/dev/null"
 103                         skip=skip
 104                     else
 105                         inf="if=/dev/zero"
 106                         outf="of=$dev"
 107                         skip=seek
 108                     fi
 109                     for ((j=0;j<crg;j++)); do
 110                         sgp_dd 2> ${tmpf}_${i}_${j} \
 111                             $inf $outf ${skip}=$((1024+j*blocks)) \
 112                             thr=$((thr/crg)) count=$count bs=$bs bpt=$bpt time=1&
 113                     done
 114                 done
 115                 wait
 116                 t1=`date +%s.%N`
 117                 # collect/check individual stats
 118                 echo > $tmpf
 119                 ok=0
 120                 for ((i=0;i<ndevs;i++)); do
 121                     for ((j=0;j<crg;j++)); do
 122                         rtmp=${tmpf}_${i}_${j}
 123                         if grep 'time to transfer data' $rtmp > /dev/null 2>&1; then
 124                             ok=$((ok + 1))
 125                         fi
 126                         cat ${rtmp} >> $tmpf
 127                         cat ${rtmp} >> $workf
 128                         rm  ${rtmp}
 129                     done
 130                 done
 131                 if ((ok != ndevs*crg)); then
 132                     print_summary -n "$((ndevs*crg - ok)) failed "
 133                 else
 134                     # compute MB/sec from elapsed
 135                     bw=`awk "BEGIN {printf \"%7.2f MB/s\", $actual_size * $ndevs / (( $t1 - $t0 ) * 1024); exit}"`
 136                     # compute MB/sec from nregions*slowest
 137                     check=`awk < $tmpf \
 138                         '/time to transfer data/ {mb=$8/1.048576; if (n == 0 || mb < min) min = mb; n++}\
 139                         END {printf "%5d x %6.2f = %7.2f MB/s", n, min, min * n}'`
 140                     print_summary -n "$bw $check "
 141                 fi
 142                 rm $tmpf
 143             done
 144             print_summary ""
 145         done
 146     done
 147 done