1 Block IO Controller 2 =================== 3Overview 4======== 5cgroup subsys "blkio" implements the block io controller. There seems to be 6a need of various kinds of IO control policies (like proportional BW, max BW) 7both at leaf nodes as well as at intermediate nodes in a storage hierarchy. 8Plan is to use the same cgroup based management interface for blkio controller 9and based on user options switch IO policies in the background. 10 11In the first phase, this patchset implements proportional weight time based 12division of disk policy. It is implemented in CFQ. Hence this policy takes 13effect only on leaf nodes when CFQ is being used. 14 15HOWTO 16===== 17You can do a very simple testing of running two dd threads in two different 18cgroups. Here is what you can do. 19 20- Enable Block IO controller 21 CONFIG_BLK_CGROUP=y 22 23- Enable group scheduling in CFQ 24 CONFIG_CFQ_GROUP_IOSCHED=y 25 26- Compile and boot into kernel and mount IO controller (blkio). 27 28 mount -t cgroup -o blkio none /cgroup 29 30- Create two cgroups 31 mkdir -p /cgroup/test1/ /cgroup/test2 32 33- Set weights of group test1 and test2 34 echo 1000 > /cgroup/test1/blkio.weight 35 echo 500 > /cgroup/test2/blkio.weight 36 37- Create two same size files (say 512MB each) on same disk (file1, file2) and 38 launch two dd threads in different cgroup to read those files. 39 40 sync 41 echo 3 > /proc/sys/vm/drop_caches 42 43 dd if=/mnt/sdb/zerofile1 of=/dev/null & 44 echo $! > /cgroup/test1/tasks 45 cat /cgroup/test1/tasks 46 47 dd if=/mnt/sdb/zerofile2 of=/dev/null & 48 echo $! > /cgroup/test2/tasks 49 cat /cgroup/test2/tasks 50 51- At macro level, first dd should finish first. To get more precise data, keep 52 on looking at (with the help of script), at blkio.disk_time and 53 blkio.disk_sectors files of both test1 and test2 groups. This will tell how 54 much disk time (in milli seconds), each group got and how many secotors each 55 group dispatched to the disk. We provide fairness in terms of disk time, so 56 ideally io.disk_time of cgroups should be in proportion to the weight. 57 58Various user visible config options 59=================================== 60CONFIG_BLK_CGROUP 61 - Block IO controller. 62 63CONFIG_DEBUG_BLK_CGROUP 64 - Debug help. Right now some additional stats file show up in cgroup 65 if this option is enabled. 66 67CONFIG_CFQ_GROUP_IOSCHED 68 - Enables group scheduling in CFQ. Currently only 1 level of group 69 creation is allowed. 70 71Details of cgroup files 72======================= 73- blkio.weight 74 - Specifies per cgroup weight. This is default weight of the group 75 on all the devices until and unless overridden by per device rule. 76 (See blkio.weight_device). 77 Currently allowed range of weights is from 100 to 1000. 78 79- blkio.weight_device 80 - One can specify per cgroup per device rules using this interface. 81 These rules override the default value of group weight as specified 82 by blkio.weight. 83 84 Following is the format. 85 86 #echo dev_maj:dev_minor weight > /path/to/cgroup/blkio.weight_device 87 Configure weight=300 on /dev/sdb (8:16) in this cgroup 88 # echo 8:16 300 > blkio.weight_device 89 # cat blkio.weight_device 90 dev weight 91 8:16 300 92 93 Configure weight=500 on /dev/sda (8:0) in this cgroup 94 # echo 8:0 500 > blkio.weight_device 95 # cat blkio.weight_device 96 dev weight 97 8:0 500 98 8:16 300 99 100 Remove specific weight for /dev/sda in this cgroup 101 # echo 8:0 0 > blkio.weight_device 102 # cat blkio.weight_device 103 dev weight 104 8:16 300 105 106- blkio.time 107 - disk time allocated to cgroup per device in milliseconds. First 108 two fields specify the major and minor number of the device and 109 third field specifies the disk time allocated to group in 110 milliseconds. 111 112- blkio.sectors 113 - number of sectors transferred to/from disk by the group. First 114 two fields specify the major and minor number of the device and 115 third field specifies the number of sectors transferred by the 116 group to/from the device. 117 118- blkio.io_service_bytes 119 - Number of bytes transferred to/from the disk by the group. These 120 are further divided by the type of operation - read or write, sync 121 or async. First two fields specify the major and minor number of the 122 device, third field specifies the operation type and the fourth field 123 specifies the number of bytes. 124 125- blkio.io_serviced 126 - Number of IOs completed to/from the disk by the group. These 127 are further divided by the type of operation - read or write, sync 128 or async. First two fields specify the major and minor number of the 129 device, third field specifies the operation type and the fourth field 130 specifies the number of IOs. 131 132- blkio.io_service_time 133 - Total amount of time between request dispatch and request completion 134 for the IOs done by this cgroup. This is in nanoseconds to make it 135 meaningful for flash devices too. For devices with queue depth of 1, 136 this time represents the actual service time. When queue_depth > 1, 137 that is no longer true as requests may be served out of order. This 138 may cause the service time for a given IO to include the service time 139 of multiple IOs when served out of order which may result in total 140 io_service_time > actual time elapsed. This time is further divided by 141 the type of operation - read or write, sync or async. First two fields 142 specify the major and minor number of the device, third field 143 specifies the operation type and the fourth field specifies the 144 io_service_time in ns. 145 146- blkio.io_wait_time 147 - Total amount of time the IOs for this cgroup spent waiting in the 148 scheduler queues for service. This can be greater than the total time 149 elapsed since it is cumulative io_wait_time for all IOs. It is not a 150 measure of total time the cgroup spent waiting but rather a measure of 151 the wait_time for its individual IOs. For devices with queue_depth > 1 152 this metric does not include the time spent waiting for service once 153 the IO is dispatched to the device but till it actually gets serviced 154 (there might be a time lag here due to re-ordering of requests by the 155 device). This is in nanoseconds to make it meaningful for flash 156 devices too. This time is further divided by the type of operation - 157 read or write, sync or async. First two fields specify the major and 158 minor number of the device, third field specifies the operation type 159 and the fourth field specifies the io_wait_time in ns. 160 161- blkio.io_merged 162 - Total number of bios/requests merged into requests belonging to this 163 cgroup. This is further divided by the type of operation - read or 164 write, sync or async. 165 166- blkio.io_queued 167 - Total number of requests queued up at any given instant for this 168 cgroup. This is further divided by the type of operation - read or 169 write, sync or async. 170 171- blkio.avg_queue_size 172 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. 173 The average queue size for this cgroup over the entire time of this 174 cgroup's existence. Queue size samples are taken each time one of the 175 queues of this cgroup gets a timeslice. 176 177- blkio.group_wait_time 178 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. 179 This is the amount of time the cgroup had to wait since it became busy 180 (i.e., went from 0 to 1 request queued) to get a timeslice for one of 181 its queues. This is different from the io_wait_time which is the 182 cumulative total of the amount of time spent by each IO in that cgroup 183 waiting in the scheduler queue. This is in nanoseconds. If this is 184 read when the cgroup is in a waiting (for timeslice) state, the stat 185 will only report the group_wait_time accumulated till the last time it 186 got a timeslice and will not include the current delta. 187 188- blkio.empty_time 189 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. 190 This is the amount of time a cgroup spends without any pending 191 requests when not being served, i.e., it does not include any time 192 spent idling for one of the queues of the cgroup. This is in 193 nanoseconds. If this is read when the cgroup is in an empty state, 194 the stat will only report the empty_time accumulated till the last 195 time it had a pending request and will not include the current delta. 196 197- blkio.idle_time 198 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. 199 This is the amount of time spent by the IO scheduler idling for a 200 given cgroup in anticipation of a better request than the exising ones 201 from other queues/cgroups. This is in nanoseconds. If this is read 202 when the cgroup is in an idling state, the stat will only report the 203 idle_time accumulated till the last idle period and will not include 204 the current delta. 205 206- blkio.dequeue 207 - Debugging aid only enabled if CONFIG_DEBUG_BLK_CGROUP=y. This 208 gives the statistics about how many a times a group was dequeued 209 from service tree of the device. First two fields specify the major 210 and minor number of the device and third field specifies the number 211 of times a group was dequeued from a particular device. 212 213- blkio.reset_stats 214 - Writing an int to this file will result in resetting all the stats 215 for that cgroup. 216 217CFQ sysfs tunable 218================= 219/sys/block/<disk>/queue/iosched/group_isolation 220----------------------------------------------- 221 222If group_isolation=1, it provides stronger isolation between groups at the 223expense of throughput. By default group_isolation is 0. In general that 224means that if group_isolation=0, expect fairness for sequential workload 225only. Set group_isolation=1 to see fairness for random IO workload also. 226 227Generally CFQ will put random seeky workload in sync-noidle category. CFQ 228will disable idling on these queues and it does a collective idling on group 229of such queues. Generally these are slow moving queues and if there is a 230sync-noidle service tree in each group, that group gets exclusive access to 231disk for certain period. That means it will bring the throughput down if 232group does not have enough IO to drive deeper queue depths and utilize disk 233capacity to the fullest in the slice allocated to it. But the flip side is 234that even a random reader should get better latencies and overall throughput 235if there are lots of sequential readers/sync-idle workload running in the 236system. 237 238If group_isolation=0, then CFQ automatically moves all the random seeky queues 239in the root group. That means there will be no service differentiation for 240that kind of workload. This leads to better throughput as we do collective 241idling on root sync-noidle tree. 242 243By default one should run with group_isolation=0. If that is not sufficient 244and one wants stronger isolation between groups, then set group_isolation=1 245but this will come at cost of reduced throughput. 246 247/sys/block/<disk>/queue/iosched/slice_idle 248------------------------------------------ 249On a faster hardware CFQ can be slow, especially with sequential workload. 250This happens because CFQ idles on a single queue and single queue might not 251drive deeper request queue depths to keep the storage busy. In such scenarios 252one can try setting slice_idle=0 and that would switch CFQ to IOPS 253(IO operations per second) mode on NCQ supporting hardware. 254 255That means CFQ will not idle between cfq queues of a cfq group and hence be 256able to driver higher queue depth and achieve better throughput. That also 257means that cfq provides fairness among groups in terms of IOPS and not in 258terms of disk time. 259 260/sys/block/<disk>/queue/iosched/group_idle 261------------------------------------------ 262If one disables idling on individual cfq queues and cfq service trees by 263setting slice_idle=0, group_idle kicks in. That means CFQ will still idle 264on the group in an attempt to provide fairness among groups. 265 266By default group_idle is same as slice_idle and does not do anything if 267slice_idle is enabled. 268 269One can experience an overall throughput drop if you have created multiple 270groups and put applications in that group which are not driving enough 271IO to keep disk busy. In that case set group_idle=0, and CFQ will not idle 272on individual groups and throughput should improve. 273 274What works 275========== 276- Currently only sync IO queues are support. All the buffered writes are 277 still system wide and not per group. Hence we will not see service 278 differentiation between buffered writes between groups. 279