How it happens


The most common cycle for a program is to 1) accept some input, 2) do some work, 3) give some output. And this sequence is repeated for as long as the program is running. Typically, the work that has to be done takes a very short time compared to the time spent waiting for input, which gives all the other running programs a chance to use the cpu in the meantime.

If a program instead takes no input and gives no output and only does work all the time, then there is much less time in which the cpu is free for other programs to use. This will make the whole system very slow, because all programs have to wait a long time to get their turn.

Demonstration


Its very easy to simulate this scenario. Here is an example. This program is a loop, which checks the condition (which is always true) and then performs the action in the loop (running the command true, which does nothing). No input, no output.

$ while true; do true; done
Hit Ctrl+C when youve had enough.


How to detect it


The easiest way to check for a cpu bound program is to use top. See if there is a program thats using almost 100% of the cpu. To be sure that it doesnt just occasionally spike leave it running for a while (or hit <space> to refresh the display a few times).

$ top
Cpu0 : 98.2%us, 1.8%sy, 0.0%ni, 0.0%id, 0.0%wa, 0.0%hi, 0.0%si
Cpu1 : 1.4%us, 0.5%sy, 0.0%ni, 98.2%id, 0.0%wa, 0.0%hi, 0.0%si

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
26210 alex 20 0 19428 2368 1492 R 99 0.1 1:14.27 bash
4075 alex 20 0 606m 195m 28m S 0 6.5 23:19.58 firefox
6337 root 20 0 505m 116m 6716 S 0 3.9 34:30.25 Xorg
Here we see that cpu0 is idle (free) 0% of the time, which means it is as busy as it possibly can be. 98.2% is due to user programs (ie. the program we just demonstrated). And when we look in the list of programs, we see that bash (the shell in which we ran our one-liner) is using 99% of the cpu.

What you can do about it


There are two cases of cpu bound programs the intentional and the accidental.

If youre running a program that does a lot of work on purpose, for instance a video encoding task, then the more cpu it uses the quicker it will finish, so using a lot of cpu is good. But if its making your whole system slow, then you can lower the priority at which the programs gets access to the cpu, so that it only uses as much cpu as the other programs leave available. To do this, use the renice command:

$ renice 20 26210
The first number you give to renice describes how nice you are being to other programs, on a scale from -20 (very selfish) to 20 (very nice). The other number is the process id (pid) of the program, which is listed by top above. Doing this will probably still make bash use almost 100% of the cpu, but not at the expense of the other programs.

On the other hand, if the program isnt supposed to be using this much cpu, then its either a bug in the program (certain versions of firefox used to spike to 100% cpu) or its just heavier than the cpu can handle. You can still renice the program, but this will make your system more responsive at the expense of the program (so for instance, firefox may become unusable). The last resort is to kill it:

$ kill 26210
On multi-cpu systems this is less of a problem, because most programs can only use one cpu, which leaves the other cpus to serve all the other programs and keep your system responsive.

How it happens


There are two types of memory on your system: physical (RAM) and virtual (swap). Physical memory is relatively small and very quick to access, while virtual memory is just part of your harddisk being used as extra memory (very slow to access). As long as all the running programs can store their work in physical memory, everything is fine. (This is why its good to have a lot of it.)

But once you fill all of the physical memory, the operating system will start moving some of the work into swap (onto the harddisk) to make space for new programs. You probably wont notice that this is happening. But when you switch from one program to another, and the second program has its work in swap, this work now has to be moved back into memory, and some of the stuff currently in memory has to be moved out to swap. This will definitely be noticeable and will make your system slow until its finished.

The effect of this situation is that your system will feel normal for some of the time (when using the same program), and then very unresponsive from time to time (when switching between programs).


Demonstration


This effect is best demonstrated with a desktop program. Start the [ Register or Signin to view external links. ] and create a canvas so large that it exceeds your available physical memory. For instance, try a canvas 10,00010,000 pixels (gimp will tell you how much memory it needs to create it). It will probably take a while to create the canvas, so just let it finish. (In order to make room for this image in memory, other programs are being moved into swap, this is called swapping.) Then do some painting on the canvas. Now switch back to another program (firefox, for instance). You should now sense that your system is slow to respond, but this is temporary for as long as it takes to restore firefox into memory.

How to detect it


Its a good idea to know how much memory your system uses under normal conditions, that way you can keep an eye on things. The command free -m will tell you about the state of your memory:

$ free -m
total used free shared buffers cached
Mem: 3015 1640 1375 0 9 124
-/+ buffers/cache: 1505 1509
Swap: 2878 0 2878
Here we see that we have 3015mb of physical memory, half of which is free. We also have almost as much swap memory, but none of that is in use.

After we create our huge canvas with the gimp we can run free again and see what has changed.

$ free -m
total used free shared buffers cached
Mem: 3015 2993 22 0 2 957
-/+ buffers/cache: 2033 982
Swap: 2878 819 2059
Were now using 819mb of swap memory, so clearly weve exceeded the capacity of physical memory.

While swapping takes place top will also show that there is a lot of I/O activity taking place. (Press Shift+M to sort the program listing by memory use.)

$ top
Cpu(s): 4.0%us, 1.6%sy, 0.0%ni, 51.0%id, 42.1%wa, 0.6%hi, 0.8%si
Mem: 3088224k total, 3043496k used, 44728k free, 3548k buffers
Swap: 2947888k total, 998812k used, 1949076k free, 957956k cached

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
2579 alex 20 0 1593m 1.4g 13m S 0 47.0 0:22.26 gimp
Here we see that user and system activity adds up to only 6%, the cpu is idle 51% of the time, but spending 42% of the time waiting for I/O operations (ie. harddisk activity). And among the programs, the gimp alone is using 1.4gb of memory.

What you can do about it


If you notice that your system becomes unresponsive when switching between programs, you have a pretty good idea that its because of swapping. There is no way to make swapping faster, so what you should do is less swapping. Keep an eye on how much memory your system is using and you should also have an idea about the memory use of various programs (the big ones). When you notice heavy swapping, the easiest way to fix it is to shut down the program thats using the most memory. When you do this you free up physical memory. The data in swap will not automatically be moved into memory (because this is expensive), but you should notice that your system is performing normally again.

How it happens


Input/Output (I/O, also written io) is an umbrella term for *everything* that happens on your system that does not involve the cpu, the memory or the video card (gpu). When talking about performance, io usually means the harddrive, because thats what your system uses most heavily (and therefore what we spend the most time waiting for), but it can also refer to your network card, your cdrom drive, your keyboard etc.

A program running on the cpu, which does a lot of io (such as reading/writing large files), will spend a lot of time waiting for this io to complete. This leaves the cpu busy and other programs have less opportunity to run. The effect is that the whole system may become consistently unresponsive until the heavy io is completed.


Demonstration


We can demonstrate the effect of heavy io by reading and writing a lot of data to the harddrive. Here we find the device that your root partition is on (probably /dev/sda1) and then read 5gb from it, writing it to a file /tmp/dummy (you may want to check that you have enough free space).

$ device=`mount | grep " / " | awk '{ print $1 }'`
$ sudo dd if=$device of=/tmp/dummy bs=5120 count=1048576
This should take around 10 minutes, so you can see how your system behaves while this is happening.


How to detect it


We can detect heavy io with top.

$ top
Cpu0 : 13.7%us, 6.5%sy, 0.0%ni, 0.0%id, 79.0%wa, 0.8%hi, 0.0%si
Cpu1 : 4.7%us, 3.9%sy, 0.0%ni, 73.2%id, 17.3%wa, 0.8%hi, 0.0%si

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
30956 root 20 0 10388 1812 640 D 5 0.1 0:02.76 dd
Here we see that cpu0 is spending 79% of its time waiting for io. In the list of programs we see the program dd that we ran. Its only using 5% cpu, which seems to conflict with the number 79%, but then we see it has status D, which means waiting for io. The reason for this is that while dd is only doing actual work on the cpu 5% of the time, its still using a lot of cpu time because of all the io.

Most io is harddrive io, but we can see if this is the case with the tool atop, which is similar to top.

$ atop
CPU | sys 14% | user 21% | irq 2% | idle 39% | wait 125% |
cpu | sys 10% | user 11% | irq 2% | idle 0% | cpu000 w 78% |
cpu | sys 4% | user 10% | irq 0% | idle 38% | cpu001 w 48% |
DSK | sda | busy 98% | read 1371 | write 1011 | avio 4 ms |

PID SYSCPU USRCPU VGROW RGROW RDDSK WRDSK ST EXC S CPU CMD 1/4
30956 0.60s 0.00s 0K 0K 113.4M 114.0M ** * D 6% dd
Here we again see that the program dd has status D (io wait), and cpu0 is spending 78% waiting for io. In addition, we see that the harddrive sda (which is the one we are reading and writing to, /dev/sda) is busy 98% of the time. So we know that its the harddrive thats responsible for using 78% of cpu0.


What you can do about it


If your system becomes unresponsive because of io, it is because the cpu is not being shared among the programs in a way that allows them all to stay responsive. So the answer is to prioritize certain programs over others. ionice is the io counterpart to nice.

$ ionice -p30956 -n7
Here we are telling ionice first the process id of the program, and then the io priority it should have, on a scale from 0 (highest priority) to 7 (lowest).