freebsd-dev/share/man/man7/tuning.7
Robert Watson fe3e42064f o When describing kern.ipc.nsfbufs, use the slightly less colloquial
"modify" rather than the slightly more colloquial "mess with".
2001-12-06 19:48:48 +00:00

781 lines
26 KiB
Groff

.hlm 0
.\" Copyright (c) 2001, Matthew Dillon. Terms and conditions are those of
.\" the BSD Copyright as specified in the file "/usr/src/COPYRIGHT" in
.\" the source tree.
.\"
.\" $FreeBSD$
.\"
.Dd May 25, 2001
.Dt TUNING 7
.Os
.Sh NAME
.Nm tuning
.Nd performance tuning under FreeBSD
.Sh SYSTEM SETUP - DISKLABEL, NEWFS, TUNEFS, SWAP
When using
.Xr disklabel 8
to lay out your filesystems on a hard disk it is important to remember
that hard drives can transfer data much more quickly from outer tracks
than they can from inner tracks.
To take advantage of this you should
try to pack your smaller filesystems and swap closer to the outer tracks,
follow with the larger filesystems, and end with the largest filesystems.
It is also important to size system standard filesystems such that you
will not be forced to resize them later as you scale the machine up.
I usually create, in order, a 128M root, 1G swap, 128M
.Pa /var ,
128M
.Pa /var/tmp ,
3G
.Pa /usr ,
and use any remaining space for
.Pa /home .
.Pp
You should typically size your swap space to approximately 2x main memory.
If you do not have a lot of RAM, though, you will generally want a lot
more swap.
It is not recommended that you configure any less than
256M of swap on a system and you should keep in mind future memory
expansion when sizing the swap partition.
The kernel's VM paging algorithms are tuned to perform best when there is
at least 2x swap versus main memory.
Configuring too little swap can lead
to inefficiencies in the VM page scanning code as well as create issues
later on if you add more memory to your machine.
Finally, on larger systems
with multiple SCSI disks (or multiple IDE disks operating on different
controllers), we strongly recommend that you configure swap on each drive
(up to four drives).
The swap partitions on the drives should be approximately the same size.
The kernel can handle arbitrary sizes but
internal data structures scale to 4 times the largest swap partition.
Keeping
the swap partitions near the same size will allow the kernel to optimally
stripe swap space across the N disks.
Don't worry about overdoing it a
little, swap space is the saving grace of
.Ux
and even if you don't normally use much swap, it can give you more time to
recover from a runaway program before being forced to reboot.
.Pp
How you size your
.Pa /var
partition depends heavily on what you intend to use the machine for.
This
partition is primarily used to hold mailboxes, the print spool, and log
files.
Some people even make
.Pa /var/log
its own partition (but except for extreme cases it isn't worth the waste
of a partition ID).
If your machine is intended to act as a mail
or print server,
or you are running a heavily visited web server, you should consider
creating a much larger partition \(en perhaps a gig or more.
It is very easy
to underestimate log file storage requirements.
.Pp
Sizing
.Pa /var/tmp
depends on the kind of temporary file usage you think you will need.
128M is
the minimum we recommend.
Also note that sysinstall will create a
.Pa /tmp
directory, but it is usually a good idea to make
.Pa /tmp
a softlink to
.Pa /var/tmp
after the fact.
Dedicating a partition for temporary file storage is important for
two reasons: first, it reduces the possibility of filesystem corruption
in a crash, and second it reduces the chance of a runaway process that
fills up
.Oo Pa /var Oc Ns Pa /tmp
from blowing up more critical subsystems (mail,
logging, etc).
Filling up
.Oo Pa /var Oc Ns Pa /tmp
is a very common problem to have.
.Pp
In the old days there were differences between
.Pa /tmp
and
.Pa /var/tmp ,
but the introduction of
.Pa /var
(and
.Pa /var/tmp )
led to massive confusion
by program writers so today programs haphazardly use one or the
other and thus no real distinction can be made between the two.
So it makes sense to have just one temporary directory.
However you handle
.Pa /tmp ,
the one thing you do not want to do is leave it sitting
on the root partition where it might cause root to fill up or possibly
corrupt root in a crash/reboot situation.
.Pp
The
.Pa /usr
partition holds the bulk of the files required to support the system and
a subdirectory within it called
.Pa /usr/local
holds the bulk of the files installed from the
.Xr ports 7
hierarchy.
If you do not use ports all that much and do not intend to keep
system source
.Pq Pa /usr/src
on the machine, you can get away with
a 1 gigabyte
.Pa /usr
partition.
However, if you install a lot of ports
(especially window managers and linux-emulated binaries), we recommend
at least a 2 gigabyte
.Pa /usr
and if you also intend to keep system source
on the machine, we recommend a 3 gigabyte
.Pa /usr .
Do not underestimate the
amount of space you will need in this partition, it can creep up and
surprise you!
.Pp
The
.Pa /home
partition is typically used to hold user-specific data.
I usually size it to the remainder of the disk.
.Pp
Why partition at all?
Why not create one big
.Pa /
partition and be done with it?
Then I don't have to worry about undersizing things!
Well, there are several reasons this isn't a good idea.
First,
each partition has different operational characteristics and separating them
allows the filesystem to tune itself to those characteristics.
For example,
the root and
.Pa /usr
partitions are read-mostly, with very little writing, while
a lot of reading and writing could occur in
.Pa /var
and
.Pa /var/tmp .
By properly
partitioning your system fragmentation introduced in the smaller more
heavily write-loaded partitions will not bleed over into the mostly-read
partitions.
Additionally, keeping the write-loaded partitions closer to
the edge of the disk (i.e. before the really big partitions instead of after
in the partition table) will increase I/O performance in the partitions
where you need it the most.
Now it is true that you might also need I/O
performance in the larger partitions, but they are so large that shifting
them more towards the edge of the disk will not lead to a significant
performance improvement whereas moving
.Pa /var
to the edge can have a huge impact.
Finally, there are safety concerns.
Having a small neat root partition that
is essentially read-only gives it a greater chance of surviving a bad crash
intact.
.Pp
Properly partitioning your system also allows you to tune
.Xr newfs 8 ,
and
.Xr tunefs 8
parameters.
Tuning
.Xr newfs 8
requires more experience but can lead to significant improvements in
performance.
There are three parameters that are relatively safe to tune:
.Em blocksize , bytes/inode ,
and
.Em cylinders/group .
.Pp
.Fx
performs best when using 8K or 16K filesystem block sizes.
The default filesystem block size is 8K.
For larger partitions it is usually a good
idea to use a 16K block size.
This also requires you to specify a larger
fragment size.
We recommend always using a fragment size that is 1/8
the block size (less testing has been done on other fragment size factors).
The
.Xr newfs 8
options for this would be
.Dq Li "newfs -f 2048 -b 16384 ..." .
Using a larger block size can cause fragmentation of the buffer cache and
lead to lower performance.
.Pp
If a large partition is intended to be used to hold fewer, larger files, such
as a database files, you can increase the
.Em bytes/inode
ratio which reduces the number of inodes (maximum number of files and
directories that can be created) for that partition.
Decreasing the number
of inodes in a filesystem can greatly reduce
.Xr fsck 8
recovery times after a crash.
Do not use this option
unless you are actually storing large files on the partition, because if you
overcompensate you can wind up with a filesystem that has lots of free
space remaining but cannot accommodate any more files.
Using 32768, 65536, or 262144 bytes/inode is recommended.
You can go higher but
it will have only incremental effects on
.Xr fsck 8
recovery times.
For example,
.Dq Li "newfs -i 32768 ..." .
.Pp
Finally, increasing the
.Em cylinders/group
ratio has the effect of packing the inodes closer together.
This can increase directory performance and also decrease
.Xr fsck 8
times.
If you use this option at all, we recommend maxing it out.
Use
.Dq Li "newfs -c 999"
and
.Xr newfs 8
will error out and tell you what the maximum is, then use that.
.Pp
.Xr tunefs 8
may be used to further tune a filesystem.
This command can be run in
single-user mode without having to reformat the filesystem.
However, this is possibly the most abused program in the system.
Many people attempt to
increase available filesystem space by setting the min-free percentage to 0.
This can lead to severe filesystem fragmentation and we do not recommend
that you do this.
Really the only
.Xr tunefs 8
option worthwhile here is turning on
.Em softupdates
with
.Dq Li "tunefs -n enable /filesystem" .
(Note: in
.Fx
5.x
softupdates can be turned on using the
.Fl U
option to
.Xr newfs 8 ) .
Softupdates drastically improves meta-data performance, mainly file
creation and deletion.
We recommend enabling softupdates on all of your
filesystems.
There are two downsides to softupdates that you should be
aware of.
First, softupdates guarantees filesystem consistency in the
case of a crash but could very easily be several seconds (even a minute!)
behind updating the physical disk.
If you crash you may lose more work
than otherwise.
Secondly, softupdates delays the freeing of filesystem
blocks.
If you have a filesystem (such as the root filesystem) which is
close to full, doing a major update of it, e.g.\&
.Dq Li "make installworld" ,
can run it out of space and cause the update to fail.
.Pp
A number of run-time
.Xr mount 8
options exist that can help you tune the system.
The most obvious and most dangerous one is
.Cm async .
Don't ever use it, it is far too dangerous.
A less dangerous and more
useful
.Xr mount 8
option is called
.Cm noatime .
.Ux
filesystems normally update the last-accessed time of a file or
directory whenever it is accessed.
This operation is handled in
.Fx
with a delayed write and normally does not create a burden on the system.
However, if your system is accessing a huge number of files on a continuing
basis the buffer cache can wind up getting polluted with atime updates,
creating a burden on the system.
For example, if you are running a heavily
loaded web site, or a news server with lots of readers, you might want to
consider turning off atime updates on your larger partitions with this
.Xr mount 8
option.
However, you should not gratuitously turn off atime
updates everywhere.
For example, the
.Pa /var
filesystem customarily
holds mailboxes, and atime (in combination with mtime) is used to
determine whether a mailbox has new mail.
You might as well leave
atime turned on for mostly read-only partitions such as
.Pa /
and
.Pa /usr
as well.
This is especially useful for
.Pa /
since some system utilities
use the atime field for reporting.
.Sh STRIPING DISKS
In larger systems you can stripe partitions from several drives together
to create a much larger overall partition.
Striping can also improve
the performance of a filesystem by splitting I/O operations across two
or more disks.
The
.Xr vinum 8
and
.Xr ccdconfig 8
utilities may be used to create simple striped filesystems.
Generally
speaking, striping smaller partitions such as the root and
.Pa /var/tmp ,
or essentially read-only partitions such as
.Pa /usr
is a complete waste of time.
You should only stripe partitions that require serious I/O performance,
typically
.Pa /var , /home ,
or custom partitions used to hold databases and web pages.
Choosing the proper stripe size is also
important.
Filesystems tend to store meta-data on power-of-2 boundaries
and you usually want to reduce seeking rather than increase seeking.
This
means you want to use a large off-center stripe size such as 1152 sectors
so sequential I/O does not seek both disks and so meta-data is distributed
across both disks rather than concentrated on a single disk.
If
you really need to get sophisticated, we recommend using a real hardware
RAID controller from the list of
.Fx
supported controllers.
.Sh SYSCTL TUNING
There are several hundred
.Xr sysctl 8
variables in the system, including many that appear to be candidates for
tuning but actually aren't.
In this document we will only cover the ones
that have the greatest effect on the system.
.Pp
The
.Va kern.ipc.shm_use_phys
sysctl defaults to 0 (off) and may be set to 0 (off) or 1 (on).
Setting
this parameter to 1 will cause all System V shared memory segments to be
mapped to unpageable physical RAM.
This feature only has an effect if you
are either (A) mapping small amounts of shared memory across many (hundreds)
of processes, or (B) mapping large amounts of shared memory across any
number of processes.
This feature allows the kernel to remove a great deal
of internal memory management page-tracking overhead at the cost of wiring
the shared memory into core, making it unswappable.
.Pp
The
.Va vfs.vmiodirenable
sysctl defaults to 1 (on).
This parameter controls how directories are cached
by the system.
Most directories are small and use but a single fragment
(typically 1K) in the filesystem and even less (typically 512 bytes) in
the buffer cache.
However, when operating in the default mode the buffer
cache will only cache a fixed number of directories even if you have a huge
amount of memory.
Turning on this sysctl allows the buffer cache to use
the VM Page Cache to cache the directories.
The advantage is that all of
memory is now available for caching directories.
The disadvantage is that
the minimum in-core memory used to cache a directory is the physical page
size (typically 4K) rather than 512 bytes.
We recommend turning this option off in memory-constrained environments;
however, when on, it will substantially improve the performance of services
which manipulate large numbers of files.
Such services can include web caches, large mail systems, and news systems.
Turning on this option will generally not reduce performance even with the
wasted memory but you should experiment to find out.
.Pp
There are various buffer-cache and VM page cache related sysctls.
We do not recommend modifying those values.
As of
.Fx 4.3 ,
the VM system does an extremely good job tuning itself.
.Pp
The
.Va net.inet.tcp.sendspace
and
.Va net.inet.tcp.recvspace
sysctls are of particular interest if you are running network intensive
applications.
This controls the amount of send and receive buffer space
allowed for any given TCP connection.
The default sending buffer is 32k; the default receiving buffer
is 64k.
You can often
improve bandwidth utilization by increasing the default at the cost of
eating up more kernel memory for each connection.
We do not recommend
increasing the defaults if you are serving hundreds or thousands of
simultaneous connections because it is possible to quickly run the system
out of memory due to stalled connections building up.
But if you need
high bandwidth over a fewer number of connections, especially if you have
gigabit ethernet, increasing these defaults can make a huge difference.
You can adjust the buffer size for incoming and outgoing data separately.
For example, if your machine is primarily doing web serving you may want
to decrease the recvspace in order to be able to increase the
sendspace without eating too much kernel memory.
Note that the routing table (see
.Xr route 8 )
can be used to introduce route-specific send and receive buffer size
defaults.
As an additional management tool you can use pipes in your
firewall rules (see
.Xr ipfw 8 )
to limit the bandwidth going to or from particular IP blocks or ports.
For example, if you have a T1 you might want to limit your web traffic
to 70% of the T1's bandwidth in order to leave the remainder available
for mail and interactive use.
Normally a heavily loaded web server
will not introduce significant latencies into other services even if
the network link is maxed out, but enforcing a limit can smooth things
out and lead to longer term stability.
Many people also enforce artificial
bandwidth limitations in order to ensure that they are not charged for
using too much bandwidth.
.Pp
Setting the send or receive TCP buffer to values larger then 65535 will result
in a marginal performance improvement unless both hosts support the window
scaling extension of the TCP protocol, which is controlled by the
.Va net.inet.tcp.rfc1323
sysctl.
These extensions should be enabled and the TCP buffer size should be set
to a value larger than 65536 in order to obtain good performance out of
certain types of network links; specifically, gigabit WAN links and
high-latency satellite links.
.Pp
We recommend that you turn on (set to 1) and leave on the
.Va net.inet.tcp.always_keepalive
control.
The default is usually off.
This introduces a small amount of
additional network bandwidth but guarantees that dead TCP connections
will eventually be recognized and cleared.
Dead TCP connections are a
particular problem on systems accessed by users operating over dialups,
because users often disconnect their modems without properly closing active
connections.
.Pp
The
.Va kern.ipc.somaxconn
sysctl limits the size of the listen queue for accepting new TCP connections.
The default value of 128 is typically too low for robust handling of new
connections in a heavily loaded web server environment.
For such environments,
we recommend increasing this value to 1024 or higher.
The service daemon
may itself limit the listen queue size (e.g.\&
.Xr sendmail 8 ,
apache) but will
often have a directive in its configuration file to adjust the queue size up.
Larger listen queues also do a better job of fending off denial of service
attacks.
.Pp
The
.Va kern.maxfiles
sysctl determines how many open files the system supports.
The default is
typically a few thousand but you may need to bump this up to ten or twenty
thousand if you are running databases or large descriptor-heavy daemons.
.Pp
The
.Va vm.swap_idle_enabled
sysctl is useful in large multi-user systems where you have lots of users
entering and leaving the system and lots of idle processes.
Such systems
tend to generate a great deal of continuous pressure on free memory reserves.
Turning this feature on and adjusting the swapout hysteresis (in idle
seconds) via
.Va vm.swap_idle_threshold1
and
.Va vm.swap_idle_threshold2
allows you to depress the priority of pages associated with idle processes
more quickly then the normal pageout algorithm.
This gives a helping hand
to the pageout daemon.
Do not turn this option on unless you need it,
because the tradeoff you are making is to essentially pre-page memory sooner
rather then later, eating more swap and disk bandwidth.
In a small system
this option will have a detrimental effect but in a large system that is
already doing moderate paging this option allows the VM system to stage
whole processes into and out of memory more easily.
.Sh LOADER TUNABLES
Some aspects of the system behavior may not be tunable at runtime because
memory allocates they perform must occur early in the boot process.
To change loader tunables, you must set their value in
.Xr loader.conf 5
and reboot the system.
.Pp
The
.Va kern.maxusers
tunable defaults to an incredibly low value.
For most modern machines,
you probably want to increase this value to 64, 128, or 256.
We do not
recommend going above 256 unless you need a huge number of file descriptors.
Network buffers are also affected but can be controlled with a separate
kernel option.
Do not increase maxusers just to get more network mbufs.
Systems older than
.Fx 4.4
do not have this loader tunable and require that
the kernel
.Xr config 8
option
.Cd maxusers
be set instead.
.Pp
.Va kern.ipc.nmbclusters
may be adjusted to increase the number of network mbufs the system is
willing to allocate.
Each cluster represents approximately 2K of memory,
so a value of 1024 represents 2M of kernel memory reserved for network
buffers.
You can do a simple calculation to figure out how many you need.
If you have a web server which maxes out at 1000 simultaneous connections,
and each connection eats a 16K receive and 16K send buffer, you need
approximate 32MB worth of network buffers to deal with it.
A good rule of
thumb is to multiply by 2, so 32MBx2 = 64MB/2K = 32768.
So for this case
you would want to set
.Va kern.ipc.nmbclusters
to 32768.
We recommend values between
1024 and 4096 for machines with moderates amount of memory, and between 4096
and 32768 for machines with greater amounts of memory.
Under no circumstances
should you specify an arbitrarily high value for this parameter, it could
lead to a boot-time crash.
The
.Fl m
option to
.Xr netstat 1
may be used to observe network cluster use.
Older versions of
.Fx
do not have this tunable and require that the
kernel
.Xr config 8
option
.Dv NMBCLUSTERS
be set instead.
.Pp
More and more programs are using the
.Xr sendfile 2
system call to transmit files over the network.
The
.Va kern.ipc.nsfbufs
sysctl controls the number of filesystem buffers
.Xr sendfile 2
is allowed to use to perform its work.
This parameter nominally scales
with
.Va kern.maxusers
so you should not need to modify this parameter except under extreme
circumstances.
.Sh KERNEL CONFIG TUNING
There are a number of kernel options that you may have to fiddle with in
a large scale system.
In order to change these options you need to be
able to compile a new kernel from source.
The
.Xr config 8
manual page and the handbook are good starting points for learning how to
do this.
Generally the first thing you do when creating your own custom
kernel is to strip out all the drivers and services you don't use.
Removing things like
.Dv INET6
and drivers you don't have will reduce the size of your kernel, sometimes
by a megabyte or more, leaving more memory available for applications.
.Pp
.Dv SCSI_DELAY
and
.Dv IDE_DELAY
may be used to reduce system boot times.
The defaults are fairly high and
can be responsible for 15+ seconds of delay in the boot process.
Reducing
.Dv SCSI_DELAY
to 5 seconds usually works (especially with modern drives).
Reducing
.Dv IDE_DELAY
also works but you have to be a little more careful.
.Pp
There are a number of
.Dv *_CPU
options that can be commented out.
If you only want the kernel to run
on a Pentium class CPU, you can easily remove
.Dv I386_CPU
and
.Dv I486_CPU ,
but only remove
.Dv I586_CPU
if you are sure your CPU is being recognized as a Pentium II or better.
Some clones may be recognized as a Pentium or even a 486 and not be able
to boot without those options.
If it works, great!
The operating system
will be able to better-use higher-end CPU features for MMU, task switching,
timebase, and even device operations.
Additionally, higher-end CPUs support
4MB MMU pages which the kernel uses to map the kernel itself into memory,
which increases its efficiency under heavy syscall loads.
.Sh IDE WRITE CACHING
.Fx 4.3
flirted with turning off IDE write caching.
This reduced write bandwidth
to IDE disks but was considered necessary due to serious data consistency
issues introduced by hard drive vendors.
Basically the problem is that
IDE drives lie about when a write completes.
With IDE write caching turned
on, IDE hard drives will not only write data to disk out of order, they
will sometimes delay some of the blocks indefinitely when under heavy disk
loads.
A crash or power failure can result in serious filesystem
corruption.
So our default was changed to be safe.
Unfortunately, the
result was such a huge loss in performance that we caved in and changed the
default back to on after the release.
You should check the default on
your system by observing the
.Va hw.ata.wc
sysctl variable.
If IDE write caching is turned off, you can turn it back
on by setting the
.Va hw.ata.wc
kernel variable back to 1.
This must be done from the boot
.Xr loader 8
at boot time.
Attempting to do it after the kernel boots will have no effect.
Please see
.Xr ata 4
and
.Xr loader 8 .
.Pp
There is a new experimental feature for IDE hard drives called
.Va hw.ata.tags
(you also set this in the boot loader) which allows write caching to be safely
turned on.
This brings SCSI tagging features to IDE drives.
As of this
writing only IBM DPTA and DTLA drives support the feature.
Warning!
These
drives apparently have quality control problems and I do not recommend
purchasing them at this time.
If you need performance, go with SCSI.
.Sh CPU, MEMORY, DISK, NETWORK
The type of tuning you do depends heavily on where your system begins to
bottleneck as load increases.
If your system runs out of CPU (idle times
are perpetually 0%) then you need to consider upgrading the CPU or moving to
an SMP motherboard (multiple CPU's), or perhaps you need to revisit the
programs that are causing the load and try to optimize them.
If your system
is paging to swap a lot you need to consider adding more memory.
If your
system is saturating the disk you typically see high CPU idle times and
total disk saturation.
.Xr systat 1
can be used to monitor this.
There are many solutions to saturated disks:
increasing memory for caching, mirroring disks, distributing operations across
several machines, and so forth.
If disk performance is an issue and you
are using IDE drives, switching to SCSI can help a great deal.
While modern
IDE drives compare with SCSI in raw sequential bandwidth, the moment you
start seeking around the disk SCSI drives usually win.
.Pp
Finally, you might run out of network suds.
The first line of defense for
improving network performance is to make sure you are using switches instead
of hubs, especially these days where switches are almost as cheap.
Hubs
have severe problems under heavy loads due to collision backoff and one bad
host can severely degrade the entire LAN.
Second, optimize the network path
as much as possible.
For example, in
.Xr firewall 7
we describe a firewall protecting internal hosts with a topology where
the externally visible hosts are not routed through it.
Use 100BaseT rather
than 10BaseT, or use 1000BaseT rather then 100BaseT, depending on your needs.
Most bottlenecks occur at the WAN link (e.g.\&
modem, T1, DSL, whatever).
If expanding the link is not an option it may be possible to use
.Xr dummynet 4
feature to implement peak shaving or other forms of traffic shaping to
prevent the overloaded service (such as web services) from affecting other
services (such as email), or vice versa.
In home installations this could
be used to give interactive traffic (your browser,
.Xr ssh 1
logins) priority
over services you export from your box (web services, email).
.Sh SEE ALSO
.Xr netstat 1 ,
.Xr systat 1 ,
.Xr ata 4 ,
.Xr dummynet 4 ,
.Xr login.conf 5 ,
.Xr firewall 7 ,
.Xr hier 7 ,
.Xr ports 7 ,
.Xr boot 8 ,
.Xr ccdconfig 8 ,
.Xr config 8 ,
.Xr disklabel 8 ,
.Xr fsck 8 ,
.Xr ifconfig 8 ,
.Xr ipfw 8 ,
.Xr loader 8 ,
.Xr mount 8 ,
.Xr newfs 8 ,
.Xr route 8 ,
.Xr sysctl 8 ,
.Xr tunefs 8 ,
.Xr vinum 8
.Sh HISTORY
The
.Nm
manual page was originally written by
.An Matthew Dillon
and first appeared
in
.Fx 4.3 ,
May 2001.