How to Make Linux Faster
Without Buying New Hardware
Kyle Wheeler
April 18, 2002
1 Introduction
A few years ago, most computers had a “turbo” button (and before that, they had a physical turbo switch)
to switch between normal (say, 10MHz) and extra-fast (say, 12MHz). The reason was that while most of
the time you wanted your programs to go faster, sometimes—especially playing games—12MHz was just
a little too fast.1 These days, programs are (usually) a little smarter than that, and there’s no such thing
as a computer that’s too fast. You already have a very fast computer, but chances are, you aren’t taking
advantage of how fast it really is.
2 The Disk
Most of the components of the computer are not all the same speed. Your CPU may be slower than
your RAM, the main bus may be slower than either of those, your PCI bus may be slower than that, but
still faster than your hard drive—all of which is faster than the Zip/floppy/CD-ROM drive. This creates
bottlenecks in the computer—pieces of the computer that slow the rest of it down.
One of the main bottlenecks in a computer is disk access. The disk is the obstacle that slows down
copying files (usually), and since most operating systems use the disk for ”swap space” to augment their
memory management, the disk can slow things down even more.
2.1 hdparm
Chances are your hard drive is a lot faster than it seems to be. Over the last twenty years or so, several
advances in hard drive technology has made them faster and faster. However, not all hard drives support
the same features. Thus, when the typical Linux distribution installs itself it plays things safe, and only
talks to the hard drive in a way that is practically guaranteed to work, be stable, and unfortunately, be very
slow. Fortunately, if your hard drive has been purchased in the last five to ten years, you can probably get
a little more out of it than you’re currently getting with a tool called .
There are many options and features about your hard drive that can affect, and not all of them
are particularly useful (from a speed perspective anyway). For example, you can put the drive to sleep, or
you can set it to read-only, or whatever. The speed-related options, however, can be a little dangerous, es-
pecially if you are working with older hardware and a certain feature is not supported. It is recommended
that you switch your machine to single-user mode (
) before fiddling with any of these flags. 2
1 The reason there is that the games used “timing-loops” to pace out the game. Essentially, they would do some simple routine a
few thousand times in order to space out, for example, screen updates. At the slower speeds, a simple routine a few thousand times
could be counted on to take a little while. The faster the computer got, the quicker the loop was processed, and the pacing of the
game sped up along with the processor.
2 That said, I’ve done it over the network with few processes running without ill effects, but some of these options warn that you
can cause massive file-system corruption if you abuse them. You have been warned.
1
2.1.1 Speed Testing
Probably the first thing you want to do is test the speed of your drive in a quantitative way. There are two
test options, one of which ( ) simply tests the speed of the Linux buffer cache. That option is generally
a function of the speed of your CPU, main bus, RAM, and kernel version, but kind of interesting to know.
The other option, is the raw disk I/O speed. You should run it 2-3 times to get meaningful results.
2.1.2 DMA
The 2.4.x Linux kernels turn DMA on by default, but on some types of hardware or on 2.2.x kernels it is an
option. This is probably the biggest speed boost you will see—access speeds can jump up to 300
There are some other
options that are related - the and
options. These are almost never
useful, and usually dangerous (this is a very quick and easy way to corrupt your filesystem). On modern
kernels (2.2.x) these options are nearly always set to the correct and optimal value for the drive.
2.1.3 Multiple Sector I/O
Multiple Sector Mode (also known as IDE Block Mode) is a feature of most modern IDE hard drives, per-
mitting the transfer of multiple sectors per I/O interrupt, rather than the usual one sector per interrupt.
When this feature is enabled, it typically reduces OS overhead for disk access by 30-50%. On many sys-
tems, it also provides increased data throughput of anywhere from 5 to 50%. Some drives, however, run
more slowly with this option turned on (like Western Digital Caviar drives). Do testing to find out which
is the best setting. Generally, you should do
, look for the
record, and
use that value initially, test, and back down from there (in powers of two). Try it like this:
2.1.4 Bandwidth
The most compatible and safe method for talking to hard drives is 16-bits at a time. All (E)IDE drives
still have only a 16-bit connection over the ribbon cable , but the IDE chipsets in modern computers can
usually talk to the computer (over the PCI or VLB bus) 32-bits at a time. I say usually, because there are
a few IDE chipsets that get confused by 32-bits, and require a special “sync” sequence. This is still faster
than 16-bit mode, but requires a little more overhead for the OS than pure 32-bit mode. For , there
are three possible arguments here: 0 (16-bit mode), 1 (32-bit mode), and 3 (32-bit with sync). Sync mode
works with all modern chipsets. Switching from 16-bit to 32-bit mode typically doubles your disk access
speed. Use it like this:
2.1.5 Interrupts
You can allow Linux to unmask other interrupts during processing of a disk interrupt. This greatly im-
proves Linux’s responsiveness and eliminates ”serial port overrun” errors (where accessing the disk (e.g.
one that is asleep and has to wake up) can delay the computer enough that the modem fills the serial port
buffer, for example). However, some chipsets (specifically CMD-640B and RZ1000) are unreliable in this
mode due to a hardware flaw (disabling “IDE prefetch” in the BIOS may help with those chipsets). Set this
option like this:
2.1.6 Finally
Now that you’ve set all the options, test the speed again. See how much faster that is? To quote the O’Reilly
website, “Now launch Netscape. And try not to fall out of your chair.”
2
2.2 Filesystem
Depending on what you’re doing, the “format” or filesystem of your disk partition may be slowing you
down. For example, if you used to dual-boot your machine into Windows and Linux and one of your
partitions is all MS-DOS, that will slow down access to files on that partition. What filesystem to use
depends on what you will use the disk for. EXT2 or EXT3 is generally stable and fast for desktop machines.
Webservers or systems with many users may benefit from using something like SGI’s XFS.
As a tip, there are a lot of “journaling” filesystems out there for Linux (EXT3 is one of them). Journaling
filesystems are generally safer than standard filesystems in that the disk is much less likely to be in an
inconsistent state if your machine loses power or crashes (you were probably doing kernel development,
right?). Journaling filesystems are not inherently any faster or slower during normal use, but using one
avoids the need to run on your partitions after a crash—which can save you all kinds of time.
If you have a kernel that supports it (any kernel past 2.4.9, as long as you compiled in EXT3 support),
you should probably convert your partitions to EXT3 format (unless you’re running a high-availability,
high-load server that may stress the EXT3 code in ways that hasn’t been thought of before). You’ll thank
me the next time the power goes out and you think about buying a UPS again. Converting your partitions
is very easy. First, run this command on them:
Then edit
and change the partition’s filetype. Then unmount or remount the partition (you
can reboot if you like). It’s that simple. And EXT3 is backwards-compatible with EXT2, so if you have an
old rescue disk, it will be able to read the EXT3 partition as if it was an EXT2 partition.
Another thing that will increase your disk performance is telling the OS not to record file access time.
Normally, every time you access a file (like on a webserver when a page is requested) the access time
is incremented in the file’s inode. You can add the
option to a partition’s entry in
to disable this, which will speed up performance (because it requires fewer disk operations for each file
access), particularly in busy servers.
2.3 Swap, Partition Position, and More
If you have multiple drives, move the partitions that are the most active (the
partition, for example)
to the faster drive. Also, if you have multiple drives, put a swap partition on each drive. This speeds up
swapping (because the kernel can interleave swap reads and writes and can then read from both swap
partitions at the same time).
Also, leave empty space on your disk. If your disk is more than about 75
3 Memory
The other main bottleneck in your computer is RAM. RAM is very fast memory, and the CPU talks to it
a lot for whatever it’s currently doing. Ideally, the CPU keeps every program currently executing entirely
in RAM. This makes switching between applications and doing any computation or memory access any
of the running applications want to have done very fast and easy. However, if you don’t have very much
RAM, the OS will keep what it’s working on right now in RAM, and will save what doesn’t fit in memory to
disk temporarily (read: swap space). The more things you do and the less RAM you have, the more things
will need to be saved to disk while the CPU works on what’s currently in RAM.
3.1 More Swap Space
If you enjoy running many many programs at once, but they aren’t all always active (like, for example,
if you want to keep several Mozilla windows open at once, even though you may only be using one at a
time), then you may want to invest in more swap space. This may be a null-gain, because of the time it
takes to take what was in swap and put it back in RAM and take what’s in RAM and put it on disk when
you change programs or windows, BUT, more swap space will allow more processes to run at once and
will allow the OS to clear out more RAM for the current process.
3
Once upon a time, swap space was limited to being a maximum of 128MB. This is no longer the case
(and hasn’t for a few years), so load up with at least as much swap space as you have RAM.
3.2 Buy More Ram
Yes, the title of this paper includes “Without Buying. . . ”, but RAM is reasonably inexpensive. If you can
afford it, get up to 512MB of RAM, and you will see marked speed improvements.
3.3 Choosing What to Run
The problem with RAM is that there’s never enough of it to go around. One thing you can do to speed
up your machine (and also to secure it) is to examine what exactly you are running, and perhaps use
something else that is less RAM-hungry or don’t use a certain program at all.
Start by disabling and removing default servers and similar items. For example, unless you’re using
NFS, you probably don’t need portmap running and taking up RAM. If you don’t need a web-server, don’t
run one.
In general, GUI programs, particularly GNOME and KDE and the like, are very memory hungry. You
can cut down on that and speed up your computer by switching from them to a simpler, window-manager-
only approach. For example, try blackbox, fluxbox, pwm, or WindowMaker as alternatives to GNOME
and KDE—they provide many similar features, without nearly as great a demand on your computer’s re-
sources. Similarly, try out different terminal programs. Instead of gnome-terminal or kterminal (which
terminal programs like
,
, and
both use large amounts of RAM and have many features you may not need), try out smaller and lighter
.
4 Programs
4.1 Kernel
4.1.1 Version
In general, newer kernels are frequently faster than older ones. The 2.4.x series of kernels contains many
internal optimizations that make them faster than the 2.2.x series. Download the most recent version of
the kernel, compile it and install it—quite likely, it will be faster than it used to be.
4.1.2 Modules
Ideally, you want your kernel to be tailored for your machine—support for exactly what hardware you
have, and what features you want, and no others. This reduces kernel size, and in some cases increases
kernel speed.
When compiling a kernel, many options have the ability to be built as modules, to be loaded during
runtime, rather than being built into the kernel itself. This is something to consider when thinking of
speed. If you want your computer to support something—a feature or a hardware device—but that device
will not usually be attached or that feature will not always be used (such as the kernel pcap functions, or
a USB driver), you should compile that as a module, to keep the kernel from thinking about a feature or
device that may not be attached or used. However, if you have a device or feature that is always in use
(like an ethernet card or video card), you should compile it into the kernel if you can (some people may
have kernel size restrictions). This allows the compiler to optimize the kernel a little more, and doesn’t
require the kernel module auto-loader to go looking for the module.
4.1.3 Low-Latency and Preemptive Patches
There are some kernel patches out there for low-latency (
)
or to make the kernel preemptable (
). These can reduce the latency
of your machine and make it respond faster and *feel* faster, even if it isn’t technically processing things
faster. Be warned that these patches are experimental.
4
4.2 Compiler
The compiler is able to improve execution speed of programs by ”optimizing” the code. The compiler can
compress multiple operations into single operations and can reformat flow control in programs (among
other things) that it compiles in order to improve execution speed (this optimization does not affect the
source code). In general, if you tell the compiler to optimize a program, the optimized version will execute
more quickly and require less CPU-time. Most Linux programs are compiled with level two optimization.
At further levels, the compiler may be unreliable. The next significant version of GCC to be released (3.1)
will include significant improvements in its optimization code, and will be able to better optimize existing
programs.
