blob: cb751b467e6ed1d2c2268b25fe6d8591f942e40a (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
|
#!/bin/bash
#
# Wrapper script for easily running a survey of zpios based tests
#
basedir="$(dirname $0)"
SCRIPT_COMMON=common.sh
if [ -f "${basedir}/${SCRIPT_COMMON}" ]; then
. "${basedir}/${SCRIPT_COMMON}"
else
echo "Missing helper script ${SCRIPT_COMMON}" && exit 1
fi
PROG=zpios-survey.sh
usage() {
cat << EOF
USAGE:
$0 [hvp] <-c config> <-t test>
DESCRIPTION:
Helper script for easy zpios survey benchmarking.
OPTIONS:
-h Show this message
-v Verbose
-p Enable profiling
-c Zpool configuration
-t Zpios test
-l Zpios survey log
EOF
}
print_header() {
tee -a ${ZPIOS_SURVEY_LOG} << EOF
================================================================
Test: $1
EOF
}
# Baseline performance for an out of the box config with no manual tuning.
# Ideally, we want everything to be automatically tuned for your system and
# for this to perform reasonably well.
zpios_survey_base() {
TEST_NAME="${ZPOOL_CONFIG}+${ZPIOS_TEST}+baseline"
print_header ${TEST_NAME}
${ZFS_SH} ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZPIOS_SH} ${VERBOSE_FLAG} -c ${ZPOOL_CONFIG} -t ${ZPIOS_TEST} | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZFS_SH} -u ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
}
# Disable ZFS's prefetching. For some reason still not clear to me
# current prefetching policy is quite bad for a random workload.
# Allowing the algorithm to detect a random workload and not do
# anything may be the way to address this issue.
zpios_survey_prefetch() {
TEST_NAME="${ZPOOL_CONFIG}+${ZPIOS_TEST}+prefetch"
print_header ${TEST_NAME}
${ZFS_SH} ${VERBOSE_FLAG} \
tee -a ${ZPIOS_SURVEY_LOG}
${ZPIOS_SH} ${VERBOSE_FLAG} -c ${ZPOOL_CONFIG} -t ${ZPIOS_TEST} \
-o "--noprefetch" | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZFS_SH} -u ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
}
# Simulating a zerocopy IO path should improve performance by freeing up
# lots of CPU which is wasted move data between buffers.
zpios_survey_zerocopy() {
TEST_NAME="${ZPOOL_CONFIG}+${ZPIOS_TEST}+zerocopy"
print_header ${TEST_NAME}
${ZFS_SH} ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZPIOS_SH} ${VERBOSE_FLAG} -c ${ZPOOL_CONFIG} -t ${ZPIOS_TEST} \
-o "--zerocopy" | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZFS_SH} -u ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
}
# Disabling checksumming should show some (if small) improvement
# simply due to freeing up a modest amount of CPU.
zpios_survey_checksum() {
TEST_NAME="${ZPOOL_CONFIG}+${ZPIOS_TEST}+checksum"
print_header ${TEST_NAME}
${ZFS_SH} ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZPIOS_SH} ${VERBOSE_FLAG} -c ${ZPOOL_CONFIG} -t ${ZPIOS_TEST} \
-s "set checksum=off" | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZFS_SH} -u ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
}
# Increasing the pending IO depth also seems to improve things likely
# at the expense of latency. This should be explored more because I'm
# seeing a much bigger impact there that I would have expected. There
# may be some low hanging fruit to be found here.
zpios_survey_pending() {
TEST_NAME="${ZPOOL_CONFIG}+${ZPIOS_TEST}+pending"
print_header ${TEST_NAME}
${ZFS_SH} ${VERBOSE_FLAG} \
zfs="zfs_vdev_max_pending=1024" | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZPIOS_SH} ${VERBOSE_FLAG} -c ${ZPOOL_CONFIG} -t ${ZPIOS_TEST} | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZFS_SH} -u ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
}
# To avoid memory fragmentation issues our slab implementation can be
# based on a virtual address space. Interestingly, we take a pretty
# substantial performance penalty for this somewhere in the low level
# IO drivers. If we back the slab with kmem pages we see far better
# read performance numbers at the cost of memory fragmention and general
# system instability due to large allocations. This may be because of
# an optimization in the low level drivers due to the contigeous kmem
# based memory. This needs to be explained. The good news here is that
# with zerocopy interfaces added at the DMU layer we could gaurentee
# kmem based memory for a pool of pages.
#
# 0x100 = KMC_KMEM - Force kmem_* based slab
# 0x200 = KMC_VMEM - Force vmem_* based slab
zpios_survey_kmem() {
TEST_NAME="${ZPOOL_CONFIG}+${ZPIOS_TEST}+kmem"
print_header ${TEST_NAME}
${ZFS_SH} ${VERBOSE_FLAG} \
zfs="zio_bulk_flags=0x100" | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZPIOS_SH} ${VERBOSE_FLAG} -c ${ZPOOL_CONFIG} -t ${ZPIOS_TEST} | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZFS_SH} -u ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
}
# Apply all possible turning concurrently to get a best case number
zpios_survey_all() {
TEST_NAME="${ZPOOL_CONFIG}+${ZPIOS_TEST}+all"
print_header ${TEST_NAME}
${ZFS_SH} ${VERBOSE_FLAG} \
zfs="zfs_vdev_max_pending=1024" \
zfs="zio_bulk_flags=0x100" | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZPIOS_SH} ${VERBOSE_FLAG} -c ${ZPOOL_CONFIG} -t ${ZPIOS_TEST} \
-o "--noprefetch --zerocopy" \
-s "set checksum=off" | \
tee -a ${ZPIOS_SURVEY_LOG}
${ZFS_SH} -u ${VERBOSE_FLAG} | \
tee -a ${ZPIOS_SURVEY_LOG}
}
PROFILE=
ZPOOL_NAME=zpios-survey
ZPOOL_CONFIG=zpool-config.sh
ZPIOS_TEST=zpios-test.sh
ZPIOS_SURVEY_LOG=/dev/null
while getopts 'hvpc:t:l:' OPTION; do
case $OPTION in
h)
usage
exit 1
;;
v)
VERBOSE=1
VERBOSE_FLAG="-v"
;;
p)
PROFILE=1
PROFILE_FLAG="-p"
;;
c)
ZPOOL_CONFIG=${OPTARG}
;;
t)
ZPIOS_TEST=${OPTARG}
;;
l)
ZPIOS_SURVEY_LOG=${OPTARG}
;;
?)
usage
exit
;;
esac
done
if [ $(id -u) != 0 ]; then
die "Must run as root"
fi
zpios_survey_base
zpios_survey_prefetch
zpios_survey_zerocopy
zpios_survey_checksum
zpios_survey_pending
zpios_survey_kmem
zpios_survey_all
exit 0
|
