NFS versus dNFS
Finally got to take Kevin Closson’s SLOB for spin. (BTW can one test dNFS with Orion at all?) Nice write up on SLOB by Yury Velikanov at Pythian: http://www.pythian.com/news/33299/my-slob-io-testing-index/
Also see Karl Arao’s SLOB cheat sheet.
NEW: see slob on steroids at http://www.pythian.com/news/34605/slob-on-steroids-part-1/ I ran into the same issues Pythian came up with
turn off default AWR
dbms_workload_repository.MODIFY_SNAPSHOT_SETTINGS(51120000, 51120000, 100, null)
reduce db_cache_size (cpu_count didn’t seem to work. The only way I got db_cache_size down was by allocating a lot to the shared pool)
*.sga_max_size=554M *.sga_target=554M *.shared_pool_size=450M *.db_cache_size=40M *.cpu_count=1 *.large_pool_size=50M
avoid db file parallel reads
*._db_block_prefetch_limit=0 *._db_block_prefetch_quota=0 *._db_file_noncontig_mblock_read_count=0
The goal was to test NFS verses DNFS.
I didn’t expect DNFS to have much of an impact in the basic case. I think of DNFS as something that adds flexibility like multi pathing and fail over and takes care of all the mount issues. In some case like on LINUX where the maximum outstanding RPC is 128, then Oracle DNFS as I understand it can go beyond these limits. I’m also guessing that in the case of rac DNFS will avoid the overhead of getattr calls that would be required with noac that is required with out DNFS. (on single instance noac can be taken off the mount options)
Setup
DB host Linux version 2.6.18-164.el5
with
/etc/sysctl.conf
sunrpc.tcp_slot_table_entries = 128
sysctl -p
Oracle 11.2.0.2.0Ran SLOB straight through with default reader.sql which does 5000 loops. This took about 12+ hours to complete 18 tests – 9 on NFS and 9 on DNFS of varying user loads.
The stats showed dNFS impact to be variable depending on the load.
I then change reader.sql to do 200 loops in order to run more tests faster and get a handle on the standard deviation between tests.
With that change, then ran the test alternatively between NFS and dNFS, ie I ran 1,2,3,4,8,16,32,64 users loads on NFS then switch to dNFS and ran the same tests. Then did this 4 times.
The stats were a bit different than the first run. I then ran the test 8 times (ie full user ramp up on DNFS then NFS , repeat 8 times)
Here are the stats.
Looks like there is a pretty consistent performance degradation on dNFS at 2 users and a good performance improvement at 8 users. At 16 users the test is just under line speed (1GbE, so around 105MB/s) By 32 users the test is hitting line saturation so I grayed out those lines.
The NFS server caches a lot thus the fast I/O times.
% impact = 100 * ( NFS time – dNFS time )/ NFS time , ie % better or worse of dNFS over NFS5000 loops200 loops x4200 loops x8usersavg ms% impactavg ms% impactStdevavg ms% impactStdev10.440.620.43-3.6810.760.45-7.909.2420.43-16.620.49-19.9515.210.49-20.973.6730.576.610.57-23.4013.490.56-8.625.6340.645.980.58-0.804.970.60-0.865.1280.7617.170.6914.0411.330.6811.802.25161.2911.861.208.5615.701.208.342.23322.23-2.852.311.8316.412.332.042.31644.23-8.384.621.7512.034.672.430.761288.180.731.86
At this point there is a clear degradation at 2 and improvement at 8 and the line maxes out just above 16, so decided to run the tests with 1-16 users, 8 times each, and here is the data:
Again, there is apparent degradation around 2 and 3 users and performance improvement peaks around 10 users, with a maximum saved time of 130us.
Here is the script to run the tests NOTE:you have to be in the directory with “runit.sh” from SLOB and you have to have backed up libodm11.so as libodm11.so.orig
slob.sh:
#!/bin/bash
# see: http://www.pythian.com/news/33299/my-slob-io-testing-index/
# if EVAL=0, then script only outputs commands without running them
EVAL=1
USERS="1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16"
REPEAT="1 2 3 4 5 6 7 8"
TYPES="DNFS NFS "
function DNFS
{
sqlplus / as sysdba << EOF
shutdown abort
EOF
cp $ORACLE_HOME/lib/libnfsodm11.so $ORACLE_HOME/lib/libodm11.so
sqlplus / as sysdba << EOF
startup
EOF
}
function NFS
{
sqlplus / as sysdba << EOF
shutdown abort
EOF
cp $ORACLE_HOME/lib/libodm11.so.orig $ORACLE_HOME/lib/libodm11.so
sqlplus / as sysdba << EOF
startup
EOF
}
function param {
for i in 1; do
cat << EOF select name,value from v\$parameter where ISDEFAULT='FALSE'; select count(*) from v\$dnfs_servers; select count(*) from v\$dnfs_files; select count(*) from v\$dnfs_channels; show sga exit EOF done > param.sql
}
function runcmd {
echo $cmd
if [ $EVAL -eq 1 ] ; then
eval $cmd
fi
}
param
for type in $TYPES; do
[[ ! -d $type ]] && mkdir $type
done
for runnumber in $REPEAT;do
echo "runumber $runnumber"
for type in $TYPES; do
if [ $EVAL -eq 1 ] ; then
eval $type
fi
echo "type $type, run number $runnumber"
for users in $USERS; do
echo "Users $users"
for output in 1; do
cmd="sqlplus / as sysdba @param"
runcmd $cmd
cmd="./runit.sh 0 $users"
runcmd $cmd
cmd="mv awr.txt ${type}/awr_${users}_${runnumber}.txt"
runcmd $cmd
done > ${type}/${type}_${users}_${runnumber}.out 2>&1
done
done
doneTo extract the data after the test you can grep out the sequential read performance in the AWR files that end in “txt”:
$ grep ‘db file sequential read’ *NFS/*txt | grep User
DNFS/awr_1_1.txt:db file sequential read 32,032 14 0 88.1 User I/O NFS/awr_1_1.txt:db file sequential read 34,633 16 0 88.8 User I/O DNFS/awr_2_1.txt:db file sequential read 88,826 52 1 95.4 User I/O NFS/awr_2_1.txt:db file sequential read 92,367 41 0 95.1 User I/O DNFS/awr_3_1.txt:db file sequential read 141,068 92 1 96.0 User I/O NFS/awr_3_1.txt:db file sequential read 145,057 81 1 96.2 User I/O DNFS/awr_4_1.txt:db file sequential read 192,897 111 1 95.2 User I/O NFS/awr_4_1.txt:db file sequential read 197,535 118 1 96.0 User I/O DNFS/awr_5_1.txt:db file sequential read 246,322 153 1 95.3 User I/O NFS/awr_5_1.txt:db file sequential read 249,038 141 1 96.2 User I/O DNFS/awr_6_1.txt:db file sequential read 297,759 181 1 95.4 User I/O NFS/awr_6_1.txt:db file sequential read 301,305 199 1 96.5 User I/O DNFS/awr_7_1.txt:db file sequential read 349,475 216 1 95.5 User I/O NFS/awr_7_1.txt:db file sequential read 352,788 244 1 96.5 User I/O DNFS/awr_8_1.txt:db file sequential read 402,262 266 1 95.8 User I/O NFS/awr_8_1.txt:db file sequential read 405,099 282 1 96.5 User I/O DNFS/awr_9_1.txt:db file sequential read 453,341 306 1 96.2 User I/O NFS/awr_9_1.txt:db file sequential read 456,009 345 1 96.7 User I/O
Alternatively calculate the average single block read time in ms:
grep ‘db file sequential read’ *NFS/*txt | grep User | sort -n -k 5 | sed -e ‘s/,//g’ | awk ‘{ printf(“%20s %10.2f %10d %10d\n”, $1, $6/$5*1000, $5, $6 ) }’ DNFS/awr_1_1.txt:db 0.44 32032 14 NFS/awr_1_1.txt:db 0.46 34633 16 DNFS/awr_2_1.txt:db 0.59 88826 52 NFS/awr_2_1.txt:db 0.44 92367 41 DNFS/awr_3_1.txt:db 0.65 141068 92 NFS/awr_3_1.txt:db 0.56 145057 81 DNFS/awr_4_1.txt:db 0.58 192897 111 NFS/awr_4_1.txt:db 0.60 197535 118 DNFS/awr_5_1.txt:db 0.62 246322 153 NFS/awr_5_1.txt:db 0.57 249038 141 DNFS/awr_6_1.txt:db 0.61 297759 181 NFS/awr_6_1.txt:db 0.66 301305 199 DNFS/awr_7_1.txt:db 0.62 349475 216 NFS/awr_7_1.txt:db 0.69 352788 244 DNFS/awr_8_1.txt:db 0.66 402262 266 NFS/awr_8_1.txt:db 0.70 405099 282 DNFS/awr_9_1.txt:db 0.67 453341 306 NFS/awr_9_1.txt:db 0.76 456009 345 DNFS/awr_10_1.txt:db 0.74 507722 376 NFS/awr_10_1.txt:db 0.85 508800 433
Or to get the averages, max, min and standard deviations,
grep 'db file sequential read' *NFS/*txt | \
grep User | \
# lines look like
# NFS/awr_9_6.txt:db file sequential read 454,048 359 1 97.0 User I/O
# first number is count
# second number is time in seconds
# names are like "NFS/awr_9_6.txt:db file sequential read "
# break 9_6 out as a number 9.6 and sort by those numbers
sed -e 's/_/ /' | \
sed -e 's/_/./' | \
sed -e 's/.txt/ /' | \
sort -n -k 2 | \
sed -e 's/,//g' | \
# lines now look like
# NFS/awr 9.6 db file sequential read 454,048 359 1 97.0 User I/O
# $2 is userload.run_number
# $7 is count
# $8 is time in seconds
# print out test # and avg ms
awk '
{ printf("%10s %10.2f ", $2, $8/$7*1000 ) }
/^NFS/ { printf("\n"); }
' | \
# print out user load
# print out %impact of dns
awk '{printf("%10s %10.2f %10.2f\n",int($1),($4+$2)/2,($4-$2)/$2*100) }
END{printf("%d\n",17) }
' | \
# for each test, each user load
# print avg ms, avg impact, std of impact, min and max impact
awk 'BEGIN{row=1;nrow=0}
{ if ( row != $1 ) {
if ( nrow > 0 ) {
printf("%10s ", row )
row = $1
for(x=1;xmax) {max=val};
if(val< min) {min=val};
}'
1 0.46 1.93 14.69 -14.00 35.00
2 0.47 -13.24 8.14 -25.42 0.00
3 0.61 -13.88 4.14 -21.13 -6.25
4 0.61 0.10 5.40 -11.29 8.62
5 0.60 -1.60 7.94 -9.68 11.67
6 0.64 12.10 6.33 -1.54 19.30
7 0.67 10.73 5.53 3.03 21.31
8 0.70 12.46 4.42 6.06 19.70
9 0.73 14.77 2.94 11.43 19.70
10 0.78 17.08 2.91 13.89 23.61
11 0.85 13.48 2.90 8.64 17.95
12 0.92 10.34 2.76 6.90 16.28
13 0.99 11.83 3.92 6.45 18.28
14 1.04 7.23 4.42 0.00 13.13
15 1.12 8.03 4.18 -0.93 13.21
16 1.21 9.37 2.78 5.08 14.04
Commentaires