@@ -122,3 +122,5 @@ xfs.txt
- info and mount options for the XFS filesystem.
xip.txt
- info on execute-in-place for file mappings.
+hot_tracking.txt
+ - info on hot data tracking in VFS layer
new file mode 100644
@@ -0,0 +1,256 @@
+Hot Data Tracking
+
+April, 2013 Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>
+
+CONTENTS
+
+1. Introduction
+2. Motivation
+3. The Design
+4. How to Calc Frequency of Reads/Writes & Temperature
+5. Git Development Tree
+6. Usage Example
+
+
+1. Introduction
+
+ The feature adds the support for tracking data temperature
+information in VFS layer. Essentially, this means maintaining some key
+stats(like number of reads/writes, last read/write time, frequency of
+reads/writes), then distilling those numbers down to a single
+"temperature" value that reflects what data is "hot", and filesystem
+can use this information to move hot data from slow devices to fast
+devices.
+
+ The long-term goal of the feature is to allow some FSs,
+e.g. Btrfs to intelligently utilize SSDs in a heterogenous volume.
+Incidentally, this project has been motivated by
+the Project Ideas page on the Btrfs wiki.
+
+
+2. Motivation
+
+ This is essentially the traditional cache argument: SSD is fast and
+expensive; HDD is cheap but slow. ZFS, for example, can already take
+advantage of SSD caching. Btrfs should also be able to take advantage of
+hybrid storage without many broad, sweeping changes to existing code.
+
+ The overall goal of enabling hot data relocation to SSD has been
+motivated by the Project Ideas page on the Btrfs wiki at
+<https://btrfs.wiki.kernel.org/index.php/Project_ideas>.
+It will divide into two parts. VFS provide hot data tracking function
+while specific FS will provide hot data relocation function.
+So as the first step of this goal, this feature provides the first part
+of the functionality.
+
+
+3. The Design
+
+These include the following parts:
+
+ * Hooks in existing vfs functions to track data access frequency
+
+ * New rb-trees for tracking access frequency of inodes and sub-file
+ranges
+ The relationship between super_block and rb-trees is as below:
+hot_info.hot_inode_tree
+ Each FS instance can find hot tracking info s_hot_root.
+ hot_info has hot_inode_tree and it has inode's hot information,
+and it has hot_range_tree, which has range's hot information.
+
+ * A list of hot inodes and hot ranges by its temperature
+
+ * A debugfs interface for dumping data from the rb-trees
+
+ * A work queue for updating inode heat info
+
+ * Mount options for enabling temperature tracking(-o hot_track,
+default mean disabled)
+ * An ioctl to retrieve the frequency information collected for a certain
+file
+ * Ioctls to enable/disable frequency tracking per inode.
+
+Let us see their relationship as below:
+
+ * hot_info.hot_inode_tree indexes hot_inode_items, one per inode
+
+ * hot_inode_item contains access frequency data for that inode
+
+ * hot_inode_item holds a heat list node to link the access frequency
+data for that inode
+
+ * hot_inode_item.hot_range_tree indexes hot_range_items for that inode
+
+ * hot_range_item contains access frequency data for that range
+
+ * hot_range_item holds a heat list node to index the access
+frequency data for that range
+
+ * hot_info.heat_inode_map indexes per-inode heat list nodes
+
+ * hot_info.heat_range_map indexes per-range heat list nodes
+
+ How about some ascii art? :) Just looking at the hot inode item case
+(the range item case is the same pattern, though), we have:
+
+ super_block
+ |
+ V
+ hot_info
+ |
+ +-------------------------+----------------------------------------+
+ | | |
+ | | |
+ V V V
+heat_inode_map hot_inode_tree heat_range_map
+ | | |
+ | V |
+ | +-------hot_comm_item--------+ |
+ | | frequency data | |
++---+ | list_head | |
+| V V |
+| ...<--hot_comm_item-->... ...<--hot_comm_item-->... |
+ frequency data frequency data |
+ list_head list_head |
+ hot_range_tree hot_range_tree |
+ | |
+ V |
+ +-------hot_comm_item--------+ |
+ | frequency data | |
+ | list_head | +---+
+ V ^ | V |
+ <--hot_comm_item-->... | | ...<--hot_comm_item-->... |
+ frequency data frequency data
+ list_head list_head
+
+
+4. How to Calc Frequency of Reads/Writes & Temperature
+
+1.) hot_rw_freq_calc()
+
+ This function does the actual work of updating the frequency numbers.
+FREQ_POWER determines how many atime deltas we keep track of (as a power of 2).
+So, setting it to anything above 16ish is probably overkill. Also,
+the higher the power, the more bits get right shifted out of the timestamp,
+reducing precision, so take note of that as well.
+
+ FREQ_POWER, defined immediately below, determines how heavily to weight
+the current frequency numbers against the newest access. For example, a value
+of 4 means that the new access information will be weighted 1/16th (ie 2^-4)
+as heavily as the existing frequency info. In essence, this is a kludged-
+together version of a weighted average, since we can't afford to keep all of
+the information that it would take to get a _real_ weighted average.
+
+2.) hot_temp_calc()
+
+ The following comments explain what exactly comprises a unit of heat.
+Each of six values of heat are calculated and combined in order to form an
+overall temperature for the data:
+
+ * NRR - number of reads since mount
+ * NRW - number of writes since mount
+ * LTR - time elapsed since last read (ns)
+ * LTW - time elapsed since last write (ns)
+ * AVR - average delta between recent reads (ns)
+ * AVW - average delta between recent writes (ns)
+
+ These values are divided (right-shifted) according to the *_DIVIDER_POWER
+values defined below to bring the numbers into a reasonable range. You can
+modify these values to fit your needs. However, each heat unit is a u32 and
+thus maxes out at 2^32 - 1. Therefore, you must choose your dividers quite
+carefully or else they could max out or be stuck at zero quite easily.
+(E.g., if you chose AVR_DIVIDER_POWER = 0, nothing less than 4s of atime
+delta would bring the temperature above zero, ever.)
+
+ Finally, each value is added to the overall temperature between 0 and 8
+times, depending on its *_COEFF_POWER value. Note that the coefficients are
+also actually implemented with shifts, so take care to treat these values
+as powers of 2. (I.e., 0 means we'll add it to the temp once; 1 = 2x, etc.)
+
+ * AVR/AVW cold unit = 2^X ns of average delta
+ * AVR/AVW heat unit = HEAT_MAX_VALUE - cold unit
+
+ E.g., data with an average delta between 0 and 2^X ns will have a cold
+value of 0, which means a heat value equal to HEAT_MAX_VALUE.
+
+ This function is responsible for distilling the six heat
+criteria, which are described in detail in hot_tracking.h) down into a single
+temperature value for the data, which is an integer between 0
+and HEAT_MAX_VALUE.
+
+ To accomplish this, the raw values from the hot_freq_data structure
+are shifted in order to make the temperature calculation more
+or less sensitive to each value.
+
+ Once this calibration has happened, we do some additional normalization and
+make sure that everything fits nicely in a u32. From there, we take a very
+rudimentary kind of "average" of each of the values, where the *_COEFF_POWER
+values act as weights for the average.
+
+ Finally, we use the MAP_BITS value, which determines the size of the
+heat list array, to normalize the temperature to the proper granularity.
+
+
+5. Git Development Tree
+
+ This feature is still on development and review, so if you're interested,
+you can pull from the git repository at the following location:
+
+ https://github.com/wuzhy/kernel.git hot_tracking
+ git://github.com/wuzhy/kernel.git hot_tracking
+
+
+6. Usage Example
+
+1.) To use hot tracking, you should mount like this:
+
+$ mount -o hot_track /dev/sdb /mnt
+[ 1505.894078] device label test devid 1 transid 29 /dev/sdb
+[ 1505.952977] btrfs: disk space caching is enabled
+[ 1506.069678] vfs: turning on hot data tracking
+
+2.) Mount debugfs at first:
+
+$ mount -t debugfs none /sys/kernel/debug
+$ ls -l /sys/kernel/debug/hot_track/
+total 0
+drwxr-xr-x 2 root root 0 Aug 8 04:40 sdb
+$ ls -l /sys/kernel/debug/hot_track/sdb
+total 0
+-rw-r--r-- 1 root root 0 Aug 8 04:40 inode_stat
+-rw-r--r-- 1 root root 0 Aug 8 04:40 extent_stat
+
+3.) View information about hot tracking from debugfs:
+
+$ echo "hot tracking test" > /mnt/file
+$ cat /sys/kernel/debug/hot_track/sdb/inode_stat
+inode 279, reads 0, writes 1, temp 109
+$ cat /sys/kernel/debug/hot_track/sdb/extent_stat
+inode 279, extent 0+1048576, reads 0, writes 1, temp 64
+
+$ echo "hot data tracking test" >> /mnt/file
+$ cat /sys/kernel/debug/hot_track/sdb/inode_stat
+inode 279, reads 0, writes 2, temp 109
+$ cat /sys/kernel/debug/hot_track/sdb/extent_stat
+inode 279, extent 0+1048576 reads 0, writes 2, temp 64
+
+4.) Check temp sorting result of some nodes:
+
+$ cat /sys/kernel/debug/hot_track/loop0/inode_spot
+inode 5248773, reads 0, writes 244, temp 111
+inode 878523, reads 0, writes 1, temp 109
+inode 878524, reads 0, writes 1, temp 109
+
+5.) Tune some hot tracking parameters as below:
+
+$ cat /proc/sys/fs/hot-age-interval
+300
+$ echo 360 > /proc/sys/fs/hot-age-interval
+$ cat /proc/sys/fs/hot-age-interval
+360
+$ cat /proc/sys/fs/hot-update-interval
+300
+$ echo 360 > /proc/sys/fs/hot-update-interval
+$ cat /proc/sys/fs/hot-update-interval
+360
+