[lkml]   [2008]   [Oct]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
Subject[PATCH -mm 1/6] i/o controller documentation
Documentation of the block device I/O controller: description, usage,
advantages and design.

Signed-off-by: Andrea Righi <>
Documentation/controllers/io-throttle.txt | 409 +++++++++++++++++++++++++++++
1 files changed, 409 insertions(+), 0 deletions(-)
create mode 100644 Documentation/controllers/io-throttle.txt

diff --git a/Documentation/controllers/io-throttle.txt b/Documentation/controllers/io-throttle.txt
new file mode 100644
index 0000000..2a3bbd1
--- /dev/null
+++ b/Documentation/controllers/io-throttle.txt
@@ -0,0 +1,409 @@
+ Block device I/O bandwidth controller
+This controller allows to limit the I/O bandwidth of specific block devices for
+specific process containers (cgroups [1]) imposing additional delays on I/O
+requests for those processes that exceed the limits defined in the control
+group filesystem.
+Bandwidth limiting rules offer better control over QoS with respect to priority
+or weight-based solutions that only give information about applications'
+relative performance requirements. Nevertheless, priority based solutions are
+affected by performance bursts, when only low-priority requests are submitted
+to a general purpose resource dispatcher.
+The goal of the I/O bandwidth controller is to improve performance
+predictability from the applications' point of view and provide performance
+isolation of different control groups sharing the same block devices.
+NOTE #1: If you're looking for a way to improve the overall throughput of the
+system probably you should use a different solution.
+NOTE #2: The current implementation does not guarantee minimum bandwidth
+levels, the QoS is implemented only slowing down I/O "traffic" that exceeds the
+limits specified by the user; minimum I/O rate thresholds are supposed to be
+guaranteed if the user configures a proper I/O bandwidth partitioning of the
+block devices shared among the different cgroups (theoretically if the sum of
+all the single limits defined for a block device doesn't exceed the total I/O
+bandwidth of that device).
+A new I/O limitation rule is described using the files:
+- blockio.bandwidth-max
+- blockio.iops-max
+The I/O bandwidth (blockio.bandwidth-max) can be used to limit the throughput
+of a certain cgroup, while blockio.iops-max can be used to throttle cgroups
+containing applications doing a sparse/seeky I/O workload. Any combination of
+them can be used to define more complex I/O limiting rules, expressed both in
+terms of iops/s and bandwidth.
+The same files can be used to set multiple rules for different block devices
+relative to the same cgroup.
+The following syntax can be used to configure any limiting rule:
+- DEV is the name of the device the limiting rule is applied to.
+- LIMIT is the maximum I/O activity allowed on DEV by CGROUP; LIMIT can
+ represent a bandwidth limitation (expressed in bytes/s) when writing to
+ blockio.bandwidth-max, or a limitation to the maximum I/O operations per
+ second (expressed in iops/s) issued by CGROUP.
+ A generic I/O limiting rule for a block device DEV can be removed setting the
+ LIMIT to 0.
+- STRATEGY is the throttling strategy used to throttle the applications' I/O
+ requests from/to device DEV. At the moment two different strategies can be
+ used [2][3]:
+ 0 = leaky bucket: the controller accepts at most B bytes (B = LIMIT * time)
+ or O operations (O = LIMIT * time); further I/O requests
+ are delayed scheduling a timeout for the tasks that made
+ those requests.
+ Different I/O flow
+ | | |
+ | v |
+ | v
+ v
+ .......
+ \ /
+ \ / leaky-bucket
+ ---
+ |||
+ vvv
+ Smoothed I/O flow
+ 1 = token bucket: LIMIT tokens are added to the bucket every seconds; the
+ bucket can hold at the most BUCKET_SIZE tokens; I/O
+ requests are accepted if there are available tokens in the
+ bucket; when a request of N bytes arrives N tokens are
+ removed from the bucket; if fewer than N tokens are
+ available the request is delayed until a sufficient amount
+ of token is available in the bucket.
+ Tokens (I/O rate)
+ o
+ o
+ o
+ ....... <--.
+ \ / | Bucket size (burst limit)
+ \ooo/ |
+ --- <--'
+ |ooo
+ Incoming --->|---> Conforming
+ I/O |oo I/O
+ requests -->|--> requests
+ |
+ ---->|
+ Leaky bucket is more precise than token bucket to respect the limits, because
+ bursty workloads are always smoothed. Token bucket, instead, allows a small
+ irregularity degree in the I/O flows (burst limit), and, for this, it is
+ better in terms of efficiency (bursty workloads are not smoothed when there
+ are sufficient tokens in the bucket).
+- BUCKET_SIZE is used only with token bucket (STRATEGY == 1) and defines the
+ size of the bucket in bytes (blockio.bandwidth-max) or in I/O operations
+ (blockio.iops-max).
+- CGROUP is the name of the limited process container.
+Also the following syntaxes are allowed:
+- remove an I/O bandwidth limiting rule
+# /bin/echo DEV:0 > CGROUP/blockio.bandwidth-max
+- configure a limiting rule using leaky bucket throttling (ignore bucket size):
+# /bin/echo DEV:LIMIT:0 > CGROUP/blockio.bandwidth-max
+- configure a limiting rule using token bucket throttling
+ (with bucket size == LIMIT):
+# /bin/echo DEV:LIMIT:1 > CGROUP/blockio.bandwidth-max
+2.2. Show I/O limiting rules
+All the defined rules and statistics for a specific cgroup can be shown reading
+the files blockio.bandwidth-max for bandwidth constraints and blockio.iops-max
+for I/O operations per second constraints.
+The following syntax is used:
+$ cat CGROUP/blockio.bandwidth-max
+- MAJOR is the major device number of DEV (defined above)
+- MINOR is the minor device number of DEV (defined above)
+- LIMIT, STRATEGY and BUCKET_SIZE are the same parameters defined above
+- LEAKY_STAT is the amount of bytes (blockio.bandwidth-max) or I/O operations
+ (blockio.iops-max) currently allowed by the I/O controller (only used with
+ leaky bucket strategy - STRATEGY == 0)
+- BUCKET_FILL represents the amount of tokens present in the bucket (only used
+ with token bucket strategy - STRATEGY == 1)
+- TIME_DELTA can be one of the following:
+ - the amount of jiffies elapsed from the last I/O request (token bucket)
+ - the amount of jiffies during which the bytes or the number of I/O
+ operations given by LEAKY_STAT have been accumulated (leaky bucket)
+Multiple per-block device rules are reported in multiple rows
+(DEVi, i = 1 .. n):
+$ cat CGROUP/blockio.bandwidth-max
+The same fields are used to describe I/O operations/sec rules. The only
+difference is that the cost of each I/O operation is scaled up by a factor of
+1000. This allows to apply better fine grained sleeps and provide a more
+precise throttling.
+$ cat CGROUP/blockio.iops-max
+2.3. Additional I/O statistics
+Additional cgroup I/O throttling statistics are reported in
+$ cat CGROUP/blockio.throttlecnt
+ - MAJOR, MINOR are respectively the major and the minor number of the device
+ the following statistics refer to
+ - BW_COUNTER gives the number of times that the cgroup bandwidth limit of
+ this particular device was exceeded
+ - BW_SLEEP is the amount of sleep time measured in clock ticks (divide
+ by sysconf(_SC_CLK_TCK)) imposed to the processes of this cgroup that
+ exceeded the bandwidth limit for this particular device
+ - IOPS_COUNTER gives the number of times that the cgroup I/O operation per
+ second limit of this particular device was exceeded
+ - IOPS_SLEEP is the amount of sleep time measured in clock ticks (divide
+ by sysconf(_SC_CLK_TCK)) imposed to the processes of this cgroup that
+ exceeded the I/O operations per second limit for this particular device
+$ cat CGROUP/blockio.throttlecnt
+8 0 0 0 0 0
+^ ^ ^ ^ ^ ^
+ \ \ \ \ \ \___iops sleep (in clock ticks)
+ \ \ \ \ \____iops throttle counter
+ \ \ \ \_____bandwidth sleep (in clock ticks)
+ \ \ \______bandwidth throttle counter
+ \ \_______minor dev. number
+ \________major dev. number
+Distinct statistics for each process are reported in
+$ cat /proc/PID/io-throttle-stat
+$ cat /proc/$$/io-throttle-stat
+0 0 0 0
+^ ^ ^ ^
+ \ \ \ \_____global iops sleep (in clock ticks)
+ \ \ \______global iops counter
+ \ \_______global bandwidth sleep (clock ticks)
+ \________global bandwidth counter
+2.4. Examples
+* Mount the cgroup filesystem (blockio subsystem):
+ # mkdir /mnt/cgroup
+ # mount -t cgroup -oblockio blockio /mnt/cgroup
+* Instantiate the new cgroup "foo":
+ # mkdir /mnt/cgroup/foo
+ --> the cgroup foo has been created
+* Add the current shell process to the cgroup "foo":
+ # /bin/echo $$ > /mnt/cgroup/foo/tasks
+ --> the current shell has been added to the cgroup "foo"
+* Give maximum 1MiB/s of I/O bandwidth on /dev/sda for the cgroup "foo", using
+ leaky bucket throttling strategy:
+ # /bin/echo /dev/sda:$((1024 * 1024)):0:0 > \
+ > /mnt/cgroup/foo/blockio.bandwidth-max
+ # sh
+ --> the subshell 'sh' is running in cgroup "foo" and it can use a maximum I/O
+ bandwidth of 1MiB/s on /dev/sda
+* Give maximum 8MiB/s of I/O bandwidth on /dev/sdb for the cgroup "foo", using
+ token bucket throttling strategy, bucket size = 8MiB:
+ # /bin/echo /dev/sdb:$((8 * 1024 * 1024)):1:$((8 * 1024 * 1024)) > \
+ > /mnt/cgroup/foo/blockio.bandwidth-max
+ # sh
+ --> the subshell 'sh' is running in cgroup "foo" and it can use a maximum I/O
+ bandwidth of 1MiB/s on /dev/sda (controlled by leaky bucket throttling)
+ and 8MiB/s on /dev/sdb (controlled by token bucket throttling)
+* Run a benchmark doing I/O on /dev/sda and /dev/sdb; I/O limits and usage
+ defined for cgroup "foo" can be shown as following:
+ # cat /mnt/cgroup/foo/blockio.bandwidth-max
+ 8 16 8388608 1 0 8388608 -522560 48
+ 8 0 1048576 0 737280 0 0 216
+* Extend the maximum I/O bandwidth for the cgroup "foo" to 16MiB/s on /dev/sda:
+ # /bin/echo /dev/sda:$((16 * 1024 * 1024)):0:0 > \
+ > /mnt/cgroup/foo/blockio.bandwidth-max
+ # cat /mnt/cgroup/foo/blockio.bandwidth-max
+ 8 16 8388608 1 0 8388608 -84432 206436
+ 8 0 16777216 0 0 0 0 15212
+* Remove limiting rule on /dev/sdb for cgroup "foo":
+ # /bin/echo /dev/sdb:0:0:0 > /mnt/cgroup/foo/blockio.bandwidth-max
+ # cat /mnt/cgroup/foo/blockio.bandwidth-max
+ 8 0 16777216 0 0 0 0 110388
+* Set a maximum of 100 I/O operations/sec (leaky bucket strategy) to /dev/sdc
+ for cgroup "foo":
+ # /bin/echo /dev/sdc:100:0 > /mnt/cgroup/foo/blockio.iops-max
+ # cat /mnt/cgroup/foo/blockio.iops-max
+ 8 32 100000 0 846000 0 2113
+ ^ ^
+ /________/
+ /
+ Remember: these values are scaled up by a factor of 1000 to apply a fine
+ grained throttling (i.e. LIMIT == 100000 means a maximum of 100 I/O operation
+ per second)
+* Remove limiting rule for I/O operations from /dev/sdc for cgroup "foo":
+ # /bin/echo /dev/sdc:0 > /mnt/cgroup/foo/blockio.iops-max
+* Allow I/O traffic shaping for block device shared among different cgroups
+* Improve I/O performance predictability on block devices shared between
+ different cgroups
+* Limiting rules do not depend of the particular I/O scheduler (anticipatory,
+ deadline, CFQ, noop) and/or the type of the underlying block devices
+* The bandwidth limitations are guaranteed both for synchronous and
+ asynchronous operations, even the I/O passing through the page cache or
+ buffers and not only direct I/O (see below for details)
+* It is possible to implement a simple user-space application to dynamically
+ adjust the I/O workload of different process containers at run-time,
+ according to the particular users' requirements and applications' performance
+ constraints
+The I/O throttling is performed imposing an explicit timeout, via
+schedule_timeout_killable() on the processes that exceed the I/O limits
+dedicated to the cgroup they belong to. I/O accounting happens per cgroup.
+It just works as expected for read operations: the real I/O activity is reduced
+synchronously according to the defined limitations.
+Multiple re-reads of pages already present in the page cache are not considered
+to account the I/O activity, since they actually don't generate any real I/O
+This means that a process that re-reads multiple times the same blocks of a
+file is affected by the I/O limitations only for the actual I/O performed from
+the underlying block devices.
+For write operations the scenario is a bit more complex, because the writes in
+the page cache are processed asynchronously by kernel threads (pdflush), using
+a write-back policy. So the real writes to the underlying block devices occur
+in a different I/O context respect to the task that originally generated the
+dirty pages.
+The I/O bandwidth controller uses the following solution to resolve this
+The cost of each I/O operation is always accounted when the operation is
+submitted to the I/O subsystem (submit_bio()).
+If the operation is a read then we automatically know that the context of the
+request is the current task and so we can charge the cgroup the current task
+belongs to. And throttle the current task as well, if it exceeded the cgroup
+If the operation is a write, we can charge the right cgroup looking at the
+owner of the first page involved in the I/O operation, that gives the context
+that generated the I/O activity at the source. This information can be
+retrieved using the page_cgroup functionality provided by the cgroup memory
+controller [4]. In this way we can correctly account the I/O cost to the right
+cgroup, but we cannot throttle the current task in this stage, because, in
+general, it is a different task (e.g. a kernel thread that is processing
+asynchronously the dirty page). For this reason, throttling of write operations
+is always performed asynchronously in balance_dirty_pages_ratelimited_nr(), a
+function always called by processes which are dirtying memory.
+Multiple rules for different block devices are stored in a linked list, using
+the dev_t number of each block device as key to uniquely identify each element
+of the list. RCU synchronization is used to protect the whole list structure,
+since the elements in the list are not supposed to change frequently (they
+change only when a new rule is defined or an old rule is removed or updated),
+while the reads in the list occur at each operation that generates I/O. This
+allows to provide zero overhead for cgroups that do not use any limitation.
+WARNING: per-block device limiting rules always refer to the dev_t device
+number. If a block device is unplugged (i.e. a USB device) the limiting rules
+defined for that device persist and they are still valid if a new device is
+plugged in the system and it uses the same major and minor numbers.
+NOTE: explicit sleeps are *not* imposed on tasks doing asynchronous I/O (AIO)
+operations; AIO throttling is performed returning -EAGAIN from sys_io_submit().
+Userspace applications must be able to handle this error code opportunely.
+5. TODO
+* Implement a rbtree per request queue; all the requests queued to the I/O
+ subsystem first will go in this rbtree. Then based on cgroup grouping and
+ control policy dispatch the requests and pass them to the elevator associated
+ with the queue. This would allow to provide both bandwidth limiting and
+ proportional bandwidth functionalities using a generic approach (suggested by
+ Vivek Goyal)
+* Improve fair throttling: distribute the time to sleep among all the tasks of
+ a cgroup that exceeded the I/O limits, depending of the amount of IO activity
+ previously generated in the past by each task (see task_io_accounting)
+* Try to reduce the cost of calling cgroup_io_throttle() on every submit_bio();
+ this is not too much expensive, but the call of task_subsys_state() has
+ surely a cost. A possible solution could be to temporarily account I/O in the
+ current task_struct and call cgroup_io_throttle() only on each X MB of I/O.
+ Or on each Y number of I/O requests as well. Better if both X and/or Y can be
+ tuned at runtime by a userspace tool
+* Think an alternative design for general purpose usage; special purpose usage
+ right now is restricted to improve I/O performance predictability and
+ evaluate more precise response timings for applications doing I/O. To a large
+ degree the block I/O bandwidth controller should implement a more complex
+ logic to better evaluate real I/O operations cost, depending also on the
+ particular block device profile (i.e. USB stick, optical drive, hard disk,
+ etc.). This would also allow to appropriately account I/O cost for seeky
+ workloads, respect to large stream workloads. Instead of looking at the
+ request stream and try to predict how expensive the I/O cost will be, a
+ totally different approach could be to collect request timings (start time /
+ elapsed time) and based on collected informations, try to estimate the I/O
+ cost and usage
+[1] Documentation/cgroups/cgroups.txt
+[4] Documentation/controllers/memory.txt

 \ /
  Last update: 2008-10-07 12:07    [W:0.038 / U:36.468 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site