.\"
.\" swapcache - Cache clean filesystem data & meta-data on SSD-based swap
.\"
.\" Redistribution and use in source and binary forms, with or without
.\" modification, are permitted provided that the following conditions
.\" are met:
.\" 1. Redistributions of source code must retain the above copyright
.\"    notice, this list of conditions and the following disclaimer.
.\" 2. Redistributions in binary form must reproduce the above copyright
.\"    notice, this list of conditions and the following disclaimer in the
.\"    documentation and/or other materials provided with the distribution.
.Dd February 7, 2010
.Dt SWAPCACHE 8
.Os
.Sh NAME
.Nm swapcache
.Nd a mechanism to use fast swap to cache filesystem data and meta-data
.Sh SYNOPSIS
.Cd sysctl vm.swapcache.accrate=100000
.Cd sysctl vm.swapcache.maxfilesize=0
.Cd sysctl vm.swapcache.maxburst=2000000000
.Cd sysctl vm.swapcache.curburst=4000000000
.Cd sysctl vm.swapcache.minburst=10000000
.Cd sysctl vm.swapcache.read_enable=0
.Cd sysctl vm.swapcache.meta_enable=0
.Cd sysctl vm.swapcache.data_enable=0
.Cd sysctl vm.swapcache.use_chflags=1
.Cd sysctl vm.swapcache.maxlaunder=256
.Cd sysctl vm.swapcache.hysteresis=(vm.stats.vm.v_inactive_target/2)
.Sh DESCRIPTION
.Nm
is a system capability which allows a solid state disk (SSD) in a swap
space configuration to be used to cache clean filesystem data and meta-data
in addition to its normal function of backing anonymous memory.
.Pp
Sysctls are used to manage operational parameters and can be adjusted at
any time.
Typically a large initial burst is desired after system boot,
controlled by the initial
.Va vm.swapcache.curburst
parameter.
This parameter is reduced as data is written to swap by the swapcache
and increased at a rate specified by
.Va vm.swapcache.accrate .
Once this parameter reaches zero write activity ceases until it has
recovered sufficiently for write activity to resume.
.Pp
.Va vm.swapcache.meta_enable
enables the writing of filesystem meta-data to the swapcache.
Filesystem
metadata is any data which the filesystem accesses via the disk device
using buffercache.
Meta-data is cached globally regardless of file or directory flags.
.Pp
.Va vm.swapcache.data_enable
enables the writing of clean filesystem file-data to the swapcache.
Filesystem filedata is any data which the filesystem accesses via a
regular file.
In technical terms, when the buffer cache is used to access
a regular file through its vnode.
Please do not blindly turn on this option, see the
.Sx PERFORMANCE TUNING
section for more information.
.Pp
.Va vm.swapcache.use_chflags
enables the use of the
.Va cache
and
.Va noscache
.Xr chflags 1
flags to control which files will be data-cached.
If this sysctl is disabled and
.Va data_enable
is enabled, the system will ignore file flags and attempt to
swapcache all regular files.
.Pp
.Va vm.swapcache.read_enable
enables reading from the swapcache and should be set to 1 for normal
operation.
.Pp
.Va vm.swapcache.maxfilesize
controls which files are to be cached based on their size.
If set to non-zero only files smaller than the specified size
will be cached.
Larger files will not be cached.
.Pp
.Va vm.swapcache.maxlaunder
controls the maximum number of clean VM pages which will be added to
the swap cache and written out to swap on each poll.
Swapcache polls ten times a second.
.Pp
.Va vm.swapcache.hysteresis
controls how many pages swapcache waits to be added to the inactive page
queue before continuing its scan.
Once it decides to scan it continues subject to the above limitations
until it reaches the end of the inactive page queue.
This parameter is designed to make swapcache generate more bulky bursts
to swap which helps SSDs reduce write amplification effects.
.Sh PERFORMANCE TUNING
Best operation is achieved when the active data set fits within the
swapcache.
.Pp
.Bl -tag -width 4n -compact
.It Va vm.swapcache.accrate
This specifies the burst accumulation rate in bytes per second and
ultimately controls the write bandwidth to swap averaged over a long
period of time.
This parameter must be carefully chosen to manage the write endurance of
the SSD in order to avoid wearing it out too quickly.
Even though SSDs have limited write endurance, there is massive
cost/performance benefit to using one in a swapcache configuration.
.Pp
Let's use the Intel X25V 40GB MLC SATA SSD as an example.
This device has approximately a
40TB (40 terabyte) write endurance, but see later
notes on this, it is more a minimum value.
Limiting the long term average bandwidth to 100KB/sec leads to no more
than ~9GB/day writing which calculates approximately to a 12 year endurance.
Endurance scales linearly with size.
The 80GB version of this SSD
will have a write endurance of approximately 80TB.
.Pp
MLC SSDs have a 1000-10000x write endurance, while the lower density
higher-cost SLC SSDs have a 10000-100000x write endurance, approximately.
MLC SSDs can be used for the swapcache (and swap) as long as the system
manager is cognizant of its limitations.
.Pp
.It Va vm.swapcache.meta_enable
Turning on just
.Va meta_enable
causes only filesystem meta-data to be cached and will result
in very fast directory operations even over millions of inodes
and even in the face of other invasive operations being run
by other processes.
.Pp
For
.Nm HAMMER
filesystems meta-data includes the B-Tree, directory entries,
and data related to tiny files.
Approximately 6 GB of swapcache is needed
for every 14 million or so inodes cached, effectively giving one the
ability to cache all the meta-data in a multi-terabyte filesystem using
a fairly small SSD.
.Pp
.It Va vm.swapcache.data_enable
Turning on
.Va data_enable
(with or without other features) allows bulk file data to be cached.
This feature is very useful for web server operation when the
operational data set fits in swap.
The usefulness is somewhat mitigated by the maximum number
of vnodes supported by the system via
.Va kern.maxfiles ,
because the bulk data in the cache is lost when the related
vnode is recycled.
In this case it might be desirable to
take the plunge into running a 64-bit kernel which can support
far more vnodes.
32-bit kernels have limited kernel virtual
memory (KVM) and cannot reliably support more than around
100,000 active vnodes.
64-bit kernels can support 300,000+ active vnodes.
.Pp
Data caching is definitely more wasteful of the SSD's write durability
than meta-data caching.
The swapcache may exhaust its burst and smack against the long term
average bandwidth limit, causing the SSD to wear out at the maximum rate
you programmed.
Data caching is far less wasteful and more efficient
if (on a 64-bit system only) you provide a sufficiently large SSD and
increase
.Va kern.maxvnodes
to cover the entire directory topology being served.
Each vnode requires about 1KB of physical RAM.
.Pp
Due to the higher SSD write rate you may want to use a
medium-sized SSD with good write performance to reduce interference
between reading and writing.
Write durability also scales with larger SSDs.
For example, an Intel X25-V only has 40MB/s in write performance
and burst writing by swapcache will seriously interfere with
concurrent read operation on the SSD.
The 80GB X25-M on the otherhand has double the write performance.
.Pp
When data caching is turned on you generally want to use
.Xr chflags 1
with the
.Va cache
flag to enable data caching on a directory.
This flag is tracked by the namecache and does not need to be
recursively set in the directory tree.
Simply setting the flag in a top level directory or mount point
is usually sufficient.
However, the flag does not track across mount points.
A typical setup is something like this:
.Pp
.Dl chflags cache /etc /sbin /bin /usr /home
.Dl chflags noscache /usr/obj
.Pp
If that doesn't work you can turn off
.Va vm.swapcache.use_chflags
entirely and not bother with any
.Nm chflag Ns 'ing .
.Pp
Filesystems such as NFS which do not support flags generally
have a
.Va cache
mount option which enables swapcache operation on the mount.
.Pp
.It Va vm.swapcache.maxfilesize
This may be used to reduce cache thrashing when a focus on a small
potentially fragmented filespace is desired, leaving the
larger files alone.
.Pp
.It Va vm.swapcache.minburst
This controls hysteresis and prevents nickel-and-dime write bursting.
Once
.Va curburst
drops to zero, writing to the swapcache ceases until it has recovered past
.Va minburst .
The idea here is to avoid creating a heavily fragmented swapcache where
reading data from a file must alternate between the cache and the primary
filesystem.
Doing so does not save disk seeks on the primary filesystem
so we want to avoid doing small bursts.
This parameter allows us to do larger bursts.
The larger bursts also tend to improve SSD performance as the SSD itself
can do a better job write-combining and erasing blocks.
.Pp
.It Va vm_swapcache.maxswappct
This controls the maximum amount of swapspace
.Nm
may use, in percentage terms.
.El
.Pp
It is important to note that you should always use
.Xr disklabel64 8
to label your SSD.
Disklabel64 will properly align the base of the
partition space relative to the physical drive regardless of how badly
aligned the fdisk slice is.
This will significantly reduce write amplification and write combining
inefficiencies on the SSD.
.Pp
Finally, interleaved swap (multiple SSDs) may be used to increase
performance even further.
A single SATA SSD is typically capable of reading 120-220MB/sec.
Configuring two SSDs for your swap will
improve aggregate swapcache read performance by 1.5x to 1.8x.
In tests with two Intel 40GB SSDs 300MB/sec was easily achieved.
.Pp
At this point you will be configuring more swap space than a 32 bit
.Dx
kernel can handle (due to KVM limitations).
By default, 32 bit
.Dx
systems only support 32GB of configured swap and while this limit
can be increased somewhat in
.Pa /boot/loader.conf
you should really be using a 64-bit
.Dx
kernel instead.
64-bit systems support up to 512GB of swap by default
and can be boosted to up to 8TB if you are really crazy and have enough RAM.
Each 1GB of swap requires around 1MB of physical memory to manage it so
the practical limit is more around 1TB of swap.
.Pp
Of course, a 1TB SSD is something on the order of $3000+ as of this writing.
Even though a 1TB configuration might not be cost effective, storage levels
more in the 100-200GB range certainly are.
If the machine has only a 1GigE
ethernet (100MB/s) there's no point configuring it for more SSD bandwidth.
A single SSD of the desired size would be sufficient.
.Sh INITIAL BURSTING & REPEATED BURSTING
Even though the average write bandwidth is limited it is desirable
to have a large initial burst after boot to load the cache.
.Va curburst
is initialized to 4GB by default and you can force rebursting
by adjusting it with a sysctl.
Remember that
.Va curburst
dynamically tracks burst and will go up and down depending.
.Pp
In addition there will be periods of time where the system is in
steady state and not writing to the swapcache.
During these periods
.Va curburst
will inch back up but will not exceed
.Va maxburst .
Thus the
.Va maxburst
value controls how large a repeated burst can be.
.Pp
A second bursting parameter called
.Va vm.swapcache.minburst
controls bursting when the maximum write bandwidth has been reached.
When
.Va minburst
reaches zero write activity ceases and
.Va curburst
is allowed to recover up to
.Va minburst
before write activity resumes.
The recommended range for the
.Va minburst
parameter is 1MB to 50MB.
This parameter has a relationship to
how fragmented the swapcache gets when not in a steady state.
Large bursts reduce fragmentation and reduce incidences of
excessive seeking on the hard drive.
If set too low the
swapcache will become fragmented within a single regular file
and the constant back-and-forth between the swapcache and the
hard drive will result in excessive seeking on the hard drive.
.Sh SWAPCACHE SIZE & MANAGEMENT
The swapcache feature will use up to 75% of configured swap space
by default.
The remaining 25% is reserved for normal paging operation.
The system operator should configure at least 4 times the SWAP space
versus main memory and no less than 8GB of swap space.
If a 40GB SSD is used the recommendation is to configure 16GB to 32GB of
swap (note: 32-bit is limited to 32GB of swap by default, for 64-bit
it is 512GB of swap), and to leave the remainder unwritten and unused.
.Pp
The
.Va vm_swapcache.maxswappct
sysctl may be used to change the default.
You may have to change this default if you also use
.Xr tmpfs 5 ,
.Xr vn 4 ,
or if you have not allocated enough swap for reasonable normal paging
activity to occur (in which case you probably shouldn't be using
.Nm
anyway).
.Pp
If swapcache reaches the 75% limit it will begin tearing down swap
in linear bursts by iterating through available VM objects, until
swap space use drops to 70%.
The tear-down is limited by the rate at
which new data is written and this rate in turn is often limited by
.Va vm.swapcache.accrate ,
resulting in an orderly replacement of cached data and meta-data.
The limit is typically only reached when doing full data+meta-data
caching with no file size limitations and serving primarily large
files, or (on a 64-bit system) bumping
.Va kern.maxvnodes
up to very high values.
.Sh NORMAL SWAP PAGING ACTIVITY WITH SSD SWAP
This is not a function of
.Nm
per se but instead a normal function of the system.
Most systems have
sufficient memory that they do not need to page memory to swap.
These types of systems are the ones best suited for MLC SSD
configured swap running with a
.Nm
configuration.
Systems which modestly page to swap, in the range of a few hundred
megabytes a day worth of writing, are also well suited for MLC SSD
configured swap.
Desktops usually fall into this category even if they
page out a bit more because swap activity is governed by the actions of
a single person.
.Pp
Systems which page anonymous memory heavily when
.Nm
would otherwise be turned off are not usually well suited for MLC SSD
configured swap.
Heavy paging activity is not governed by
.Nm
bandwidth control parameters and can lead to excessive uncontrolled
writing to the MLC SSD, causing premature wearout.
You would have to use the lower density, more expensive SLC SSD
technology (which has 10x the durability).
This isn't to say that
.Nm
would be ineffective, just that the aggregate write bandwidth required
to support the system would be too large for MLC flash technologies.
.Pp
With this caveat in mind, SSD based paging on systems with insufficient
RAM can be extremely effective in extending the useful life of the system.
For example, a system with a measly 192MB of RAM and SSD swap can run
a -j 8 parallel build world in a little less than twice the time it
would take if the system had 2GB of RAM, whereas it would take 5x to 10x
as long with normal HD based swap.
.Sh USING SWAPCACHE WITH NORMAL HARD DRIVES
Although
.Nm
is designed to work with SSD-based storage it can also be used with
HD-based storage as an aid for offloading the primary storage system.
Here we need to make a distinction between using RAID for fanning out
storage verses using RAID for redundancy.  There are numerous situations
where RAID-based redundancy does not make sense.
.Pp
A good example would be in an environment where the servers themselves
are redundant and can suffer a total failure without effecting
ongoing operations.  When the primary storage requirements easily fit onto
a single large-capacity drive it doesn't make a whole lot of sense to
use RAID if your only desire is to improve performance.  If you had a farm
of, say, 20 servers supporting the same facility adding RAID to each one
would not accomplish anything other than to bloat your deployment and
maintenance costs.
.Pp
In these sorts of situations it may be desirable and convenient to have
the primary filesystem for each machine on a single large drive and then
use the
.Nm
facility to offload the drive and make the machine more effective without
actually distributing the filesystem itself across multiple drives.
For the purposes of offloading while a SSD would be the most effective
from a performance standpoint, a second medium sized HD with its much lower
cost and higher capacity might actually be more cost effective.
.Pp
In cases where you might desire to use
.Nm
with a normal hard drive you should probably consider running a 64-bit
.Dx
instead of a 32-bit system.
The 64-bit build is capable of supporting much larger swap configurations
(upwards of 512G) and would be a more suitable match against a medium-sized
HD.
.Sh EXPLANATION OF STATIC VS DYNAMIC WEARING LEVELING, AND WRITE-COMBINING
Modern SSDs keep track of space that has never been written to.
This would also include space freed up via TRIM, but simply not
touching a bit of storage in a factory fresh SSD works just as well.
Once you touch (write to) the storage all bets are off, even if
you reformat/repartition later.  It takes sending the SSD a
whole-device TRIM command or special format command to take it back
to its factory-fresh condition (sans wear already present).
.Pp
SSDs have wear leveling algorithms which are responsible for trying
to even out the erase/write cycles across all flash cells in the
storage.  The better a job the SSD can do the longer the SSD will
remain usable.
.Pp
The more unused storage there is from the SSDs point of view the
easier a time the SSD has running its wear leveling algorithms.
Basically the wear leveling algorithm in a modern SSD (say Intel or OCZ)
uses a combination of static and dynamic leveling.  Static is the
best, allowing the SSD to reuse flash cells that have not been
erased very much by moving static (unchanging) data out of them and
into other cells that have more wear.  Dynamic wear leveling involves
writing data to available flash cells and then marking the cells containing
the previous copy of the data as being free/reusable.  Dynamic wear leveling
is the worst kind but the easiest to implement.  Modern SSDs use a combination
of both algorithms plus also do write-combining.
.Pp
USB sticks often use only dynamic wear leveling and have short life spans
because of that.
.Pp
In anycase, any unused space in the SSD effectively makes the dynamic
wear leveling the SSD does more efficient by giving the SSD more 'unused'
space above and beyond the physical space it reserves beyond its stated
storage capacity to cycle data throgh, so the SSD lasts longer in theory.
.Pp
Write-combining is a feature whereby the SSD is able to reduced write
amplification effects by combining OS writes of smaller, discrete,
non-contiguous logical sectors into a single contiguous 128KB physical
flash block.
.Pp
On the flip side write-combining also results in more complex lookup tables
which can become fragmented over time and reduce the SSDs read performance.
Fragmentation can also occur when write-combined blocks are rewritten
piecemeal.
Modern SSDs can regain the lost performance by de-combining previously
write-combined areas as part of their static wear leveling algorithm, but
at the cost of extra write/erase cycles which slightly increase write
amplification effects.
Operating systems can also help maintain the SSDs performance by utilizing
larger blocks.
Write-combining results in a net-reduction
of write-amplification effects but due to having to de-combine later and
other fragmentary effects it isn't 100%.
From testing with Intel devices write-amplification can be well controlled
in the 2x-4x range with the OS doing 16K writes, verses a worst-case
8x write-amplification with 16K blocks, 32x with 4K blocks, and a truly
horrid worst-case with 512 byte blocks.
.Pp
The
.Dx
.Nm
feature utilizes 64K-128K writes and is specifically designed to minimize
write amplification and write-combining stresses.
In terms of placing an actual filesystem on the SSD, the
.Dx
.Xr hammer 8
filesystem utilizes 16K blocks and is well behaved as long as you limit
reblocking operations.
For UFS you should create the filesystem with at least a 4K fragment
size, verses the default 2K.
Modern Windows filesystems use 4K clusters but it is unclear how SSD-friendly
NTFS is.
.Sh WARNINGS
I am going to repeat and expand a bit on SSD wear.
Wear on SSDs is a function of the write durability of the cells,
whether the SSD implements static or dynamic wear leveling (or both),
write amplification effects when the OS does not issue write-aligned 128KB
ops or when the SSD is unable to write-combine adjacent logical sectors,
or if the SSD has a poor write-combining algorithm for non-adjacent sectors.
In addition some additional erase/rewrite activity occurs from cleanup
operations the SSD performs as part of its static wear leveling algorithms
and its write-decombining algorithms (necessary to maintain performance over
time).  MLC flash uses 128KB physical write/erase blocks while SLC flash
typically uses 64KB physical write/erase blocks.
.Pp
The algorithms the SSD implements in its firmware are probably the most
important part of the device and a major differentiator between e.g. SATA
and USB-based SSDs.  SATA form factor drives will universally be far superior
to USB storage sticks.
SSDs can also have wildly different wearout rates and wildly different
performance curves over time.
For example the performance of a SSD which does not implement
write-decombining can seriously degrade over time as its lookup
tables become severely fragmented.
For the purposes of this manual page we are primarily using Intel and OCZ
drives when describing performance and wear issues.
.Pp
.Nm
parameters should be carefully chosen to avoid early wearout.
For example, the Intel X25V 40GB SSD has a minimum write durability
of 40TB and an actual durability that can be quite a bit higher.
Generally speaking, you want to select parameters that will give you
at least 10 years of service life.
The most important parameter to control this is
.Va vm.swapcache.accrate .
.Nm
uses a very conservative 100KB/sec default but even a small X25V
can probably handle 300KB/sec of continuous writing and still last 10 years.
.Pp
Depending on the wear leveling algorithm the drive uses, durability
and performance can sometimes be improved by configuring less
space (in a manufacturer-fresh drive) than the drive's probed capacity.
For example, by only using 32GB of a 40GB SSD.
SSDs typically implement 10% more storage than advertised and
use this storage to improve wear leveling.
As cells begin to fail
this overallotment slowly becomes part of the primary storage
until it has been exhausted.
After that the SSD has basically failed.
Keep in mind that if you use a larger portion of the SSD's advertised
storage the SSD will not know if/when you decide to use less unless
appropriate TRIM commands are sent (if supported), or a low level
factory erase is issued.
.Pp
.Nm smartctl
(from pkgsrc's sysutils/smartmontools) may be used to retrieve
the wear indicator from the drive.
One usually runs something like
.Ql smartctl -d sat -a /dev/daXX
(for AHCI/SILI/SCSI), or
.Ql smartctl -a /dev/adXX
for NATA.
Some SSDs
(particularly the Intels) will brick the SATA port when smart operations
are done while the drive is busy with normal activity, so the tool should
only be run when the SSD is idle.
.Pp
ID 232 (0xe8) in the SMART data dump indicates available reserved
space and ID 233 (0xe9) is the wear-out meter.
Reserved space
typically starts at 100 and decrements to 10, after which the SSD
is considered to operate in a degraded mode.
The wear-out meter typically starts at 99 and decrements to 0,
after which the SSD has failed.
.Pp
.Nm
tends to use large 64KB writes and tends to cluster multiple writes
linearly.
The SSD is able to take significant advantage of this
and write amplification effects are greatly reduced.
If we take a 40GB Intel X25V as an example the vendor specifies a write
durability of approximately 40TB, but
.Nm
should be able to squeeze out upwards of 200TB due the fairly optimal
write clustering it does.
The theoretical limit for the Intel X25V is 400TB (10,000 erase cycles
per MLC cell, 40GB drive), but the firmware doesn't do perfect static
wear leveling so the actual durability is less.
In tests over several hundred days we have validated a write endurance
greater than 200TB on the 40G Intel X25V using
.Nm .
.Pp
In contrast, filesystems directly stored on a SSD could have
fairly severe write amplification effects and will have durabilities
ranging closer to the vendor-specified limit.
.Pp
Power-on hours, power cycles, and read operations do not really affect wear.
There is something called read-disturb but it is unclear what sort of
ratio would be needed.  Since the data is cached in ram and thus not
re-read at a high rate there is no expectation of a practical effect.
For all intents and purposes only write operations effect wear.
.Pp
SSD's with MLC-based flash technology are high-density, low-cost solutions
with limited write durability.
SLC-based flash technology is a low-density,
higher-cost solution with 10x the write durability as MLC.
The durability also scales with the amount of flash storage.
SLC based flash is typically
twice as expensive per gigabyte.
From a cost perspective, SLC based flash
is at least 5x more cost effective in situations where high write
bandwidths are required (because it lasts 10x longer).
MLC is at least 2x more cost effective in situations where high
write bandwidth is not required.
When wear calculations are in years, these differences become huge, but
often the quantity of storage needed trumps the wear life so we expect most
people will be using MLC.
.Nm
is usable with both technologies.
.Sh SEE ALSO
.Xr chflags 1 ,
.Xr fstab 5 ,
.Xr disklabel64 8 ,
.Xr hammer 8 ,
.Xr swapon 8
.Sh HISTORY
.Nm
first appeared in
.Dx 2.5 .
.Sh AUTHORS
.An Matthew Dillon