lustre-iokit/sgpdd-survey/sgpdd-survey

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