Just wanted to revive this thread... I recently started using a software raid on my Ubuntu server and came across this thread to help tune an mdadm raid array.
Anyway.... I've rewritten the tuning script to be a bit more automated. I've only tested on my own system, so there might be bugs for others for situations that I haven't accounted for.
So... here are some differences:
1. auto detect the mdadm configured raid arrays
2. auto detect NCQ capable disks
3. does not currently tune the filesystems, but does echo out the command for tuning the filesystem
4. auto detect chunk size of array
5. auto detect raid level
6. Minor modification of stripe cache size calculation
Here is the script (obviously, same caveats as before, run at your own risk, modify as required) -
Code:
#!/bin/sh
# NEED FOLLOWING UTILS
# -- hdparm
# -- lvm
# Add VM tuning stuff?
#vm.swappiness = 1 # set low to limit swapping
#vm.vfs_cache_pressure = 50 # set lower to cache more inodes / dir entries
#vm.dirty_background_ratio = 5 # set low on systems with lots of memory
# Too HIGH on systems with lots of memory
# means huge page flushes which will hurt IO performance
#vm.dirty_ratio = 10 # set low on systems with lots of memory
# DEFAULTS
BLOCKSIZE=4 # of filesystem in KB (should I determine?)
FORCECHUNKSIZE=true # force max sectors KB to chunk size > 512
TUNEFS=true # run tune2fs on filesystem if ext[3|4]
SCHEDULER=deadline # cfq / noop / anticipatory / deadline
NR_REQUESTS=64 # NR REQUESTS
NCQDEPTH=31 # NCQ DEPTH
MDSPEEDLIMIT=200000 # Array speed_limit_max in KB/s
# ----------------------------------------------------------------------
#
# BODY
#
# ----------------------------------------------------------------------
# determine list of arrays
mdadm -Es | grep ARRAY | while read x1 x2 x3 x4 x5
do
# INIT VARIABLES
RAIDLEVEL=0
NDEVICES=0
CHUNKSIZE=0
ARRAYSTATUS=0
DISKS=""
SPARES=""
NUMDISKS=0
NUMSPARES=0
NUMPARITY=0
NCQ=0
NUMNCQDISKS=0
RASIZE=0
MDRASIZE=0
STRIPECACHESIZE=0
MINMAXHWSECKB=999999999
STRIDE=0
STRIPEWIDTH=0
# GET DETAILS OF ARRAY
ARRAY=`basename $x2`
RAIDLEVEL=`echo $x3 | cut -d'=' -f2`
case $RAIDLEVEL in
"raid6") NUMPARITY=2 ;;
"raid5") NUMPARITY=1 ;;
"raid4") NUMPARITY=1 ;;
"raid3") NUMPARITY=1 ;;
"raid1") NUMPARITY=1 ;;
"raid0") NUMPARITY=0 ;;
*)
echo "Unknown RAID level"
esac
echo ""
echo "======================================================================"
echo "FOUND ARRAY - $ARRAY / $RAIDLEVEL"
CHUNKSIZE=`mdadm --detail /dev/$ARRAY | grep 'Chunk Size' | tr -d A-Za-z':'[:blank:]`
echo "-- Chunk Size = $CHUNKSIZE KB"
FOO1=`grep "$ARRAY : " /proc/mdstat`
ARRAYSTATUS=`echo $FOO1 | cut -f 3`
# GET LIST OF DISKS IN ARRAY
echo ""
echo "Getting active devices and spares list"
for DISK in `echo $FOO1 | cut -f 5- -d \ `
do
LETTER=`echo $DISK | cut -c 1-3`
echo $DISK | grep '(S)'
RC=$?
if [ $RC -gt 0 ]
then
echo "-- $DISK - Active"
DISKS="$DISKS $LETTER"
NUMDISKS=$((NUMDISKS+1))
else
echo "-- $DISK - Spare"
SPARES="$SPARES $LETTER"
NUMSPARES=$((NUMDISKS+1))
fi
done
echo ""
echo "Active Disks ($NUMDISKS) - $DISKS"
echo "Spares Disks ($NUMSPARES) - $SPARES"
# DETERMINE SETTINGS
RASIZE=$(($NUMDISKS*($NUMDISKS-$NUMPARITY)*2*$CHUNKSIZE)) # Disk read ahead in 512byte blocks
MDRASIZE=$(($RASIZE*$NUMDISKS)) # Array read ahead in 512byte blocks
STRIPECACHESIZE=$(($RASIZE*2/8)) # in pages per device
for DISK in $DISKS $SPARES
do
# check max_hw_sectors_kb
FOO1=`cat /sys/block/$DISK/queue/max_hw_sectors_kb | awk '{print $1}'`
if [ $FOO1 -lt $MINMAXHWSECKB ]
then
MINMAXHWSECKB=$FOO1
fi
# check NCQ
hdparm -I /dev/$DISK | grep NCQ >> /dev/null
if [ $? -eq 0 ]
then
NUMNCQDISKS=$((NUMNCQDISKS+1))
fi
done
if [ $CHUNKSIZE -le $MINMAXHWSECKB ]
then
MINMAXHWSECKB=$CHUNKSIZE
fi
if [ $NUMNCQDISKS -lt $NUMDISKS ]
then
NCQDEPTH=1
echo "WARNING! ONLY $NUMNCQDISKS DISKS ARE NCQ CAPABLE!"
fi
echo ""
echo "TUNED SETTINGS"
echo "-- DISK READ AHEAD = $RASIZE blocks"
echo "-- ARRAY READ AHEAD = $MDRASIZE blocks"
echo "-- STRIPE CACHE = $STRIPECACHESIZE pages"
echo "-- MAX SECTORS KB = $MINMAXHWSECKB KB"
echo "-- NCQ DEPTH = $NCQDEPTH"
# TUNE ARRAY
echo ""
echo "TUNING ARRAY"
blockdev --setra $MDRASIZE /dev/$ARRAY
echo "-- $ARRAY read ahead set to $MDRASIZE blocks"
echo "$STRIPECACHESIZE" > /sys/block/$ARRAY/md/stripe_cache_size
echo "-- $ARRAY stripe_cache_size set to $STRIPECACHESIZE pages"
echo $MDSPEEDLIMIT > /proc/sys/dev/raid/speed_limit_max
echo "-- $ARRAY speed_limit_max set to $MDSPEEDLIMIT"
# TUNE DISKS
echo ""
echo "TUNING DISKS"
echo "Settings : "
echo " read ahead = $RASIZE blocks"
echo " max_sectors_kb = $MINMAXHWSECKB KB"
echo " scheduler = $SCHEDULER"
echo " nr_requests = $NR_REQUESTS"
echo " queue_depth = $NCQDEPTH"
for DISK in $DISKS $SPARES
do
echo "-- Tuning $DISK"
blockdev --setra $RASIZE /dev/$DISK
echo $MINMAXHWSECKB > /sys/block/$DISK/queue/max_sectors_kb
echo $SCHEDULER > /sys/block/$DISK/queue/scheduler
echo $NR_REQUESTS > /sys/block/$DISK/queue/nr_requests
echo $NCQDEPTH > /sys/block/$DISK/device/queue_depth
done
# TUNE ext3/exti4 FILESYSTEMS
STRIDE=$(($CHUNKSIZE/$BLOCKSIZE))
STRIPEWIDTH=$(($CHUNKSIZE/$BLOCKSIZE*($NUMDISKS-$NUMPARITY)))
echo ""
echo "TUNING FILESYSTEMS"
echo "For each filesystem on this array, run the following command:"
echo " tune2fs -E stride=$STRIDE,stripe-width=$STRIPEWIDTH <filesystem>"
echo ""
done
Something to note... with systems with lots of memory, you'll want to set vm.dirty_background_ratio and vm.dirty_ratio to a fairly low value, otherwise the page flushes will kill your write IO performance. I had mine set fairly high and my write performance would hover around 100MB/s (using dd to write a 20GB file). Setting the values to 5 and 10 respectively on my server (6GB RAM), I now get 180 MB/s with the same command.
My array -
Code:
Personalities : [linear] [multipath] [raid0] [raid1] [raid6] [raid5] [raid4] [raid10]
md0 : active raid6 sdj[7] sdc[0] sde[2] sdf[3] sdd[1] sdg[4] sdi[6] sdh[5]
5860574208 blocks level 6, 256k chunk, algorithm 2 [8/8] [UUUUUUUU]
All disks are 1TB Samsung disks.
Using dd to write a 20GB file -
Code:
5120+0 records in
5120+0 records out
21474836480 bytes (21 GB) copied, 117.421 s, 183 MB/s
Using dd to read in the same 20GB file -
Code:
5120+0 records in
5120+0 records out
21474836480 bytes (21 GB) copied, 44.612 s, 481 MB/s
Considering my numbers were ~ 90MB/s writes and 300MB/s reads when I started, I'm very pleased with the results.
Many thanks to the OP for providing a starting point for me.
Joo
