Alternative To the 200-Line Linux Kernel Patch

climenole writes "Phoronix recently published an article regarding a ~200 line Linux Kernel patch that improves responsiveness under system strain. Well, Lennart Poettering, a Red Hat developer, replied to Linus Torvalds on a mailing list with an alternative to this patch that does the same thing yet all you have to do is run 2 commands and paste 4 lines in your ~/.bashrc file."

Comments

Also from the article

[Weaselmancer]

Two things:

1) There isn't a difference between the kernel patch and the command line hack. They are equivalent. The command line bit was known beforehand because that was the method used to figure out if this kernel hack would be a good idea. The kernel hack just makes the process transparent.

Linus says: Right. And that's basically how this "patch" was actually tested originally - by doing this by hand, without actually having a patch in hand. I told people: this seems to work really well.

2) Linus recommends the kernel patch:

Linus also says:Put another way: if we find a better way to do something, we should _not_ say "well, if users want it, they can do this *technical thing here*". If it really is a better way to do something, we should just do it. Requiring user setup is _not_ a feature.

Source.

Re:Also from the article

[SharpFang]

The difference is the kernel patch is 200 lines of C code, which compiles to several kilobytes of machine code. The shell code needs to spawn a bash process upon startup of every other process, that's several megabytes of RAM and interpreting contents of text scripts that perform the operations.

The final effect may be the same but the overhead of performing the operation is much smaller with the kernel patch.

Re:Also from the article

[Geoffreyerffoeg]

No, incorrect. This is a modification to your .bashrc, which is (already) run every time you start a bash process, within that process (i.e., not a new process). Nothing needs to be spawned on every single process.

Admittedly the bash script does spawn some processes, but a) that's the way .bashrc works, and you have dozens of those in there, and b) it's only one process, a mkdir. The echo and the conditional run within bash itself.

The way that the configuration works, whether done in the kernel or in your .bashrc, is to associate all processes spawned from a single bash shell with a single new scheduling group. This gets you better performance when you're running processes from terminals, by associating logically-similar groups of processes in the kernel instead of letting it see all the processes as a giant pile.

The intended use case, which is pretty clear from the LKML discussion, is to make performance between something intensive (like a compilation) in a terminal and something non-terminal-associated (like watching a movie) better-balanced.

Re:Also from the article

[Anonymous Coward]

One never sets PS1 for non-interactive shells, and it's the primary way the shell tells the user's startup scripts whether they're interactive. There's a good chance the PS1 method spares a system call, too :-) It's also what the documentation says to do.

Your [ -t 0 ] approach also fails in cases where an interactive shell is being run on a non-tty. Although almost any shell since about 1990 tends to complain in such cases, at least the PS1 method will still run the right .profile code, and the -t method will not.

Re:How does this work?

[greg1104]

It makes every process spawned by the user that passes through the bash shell add their process ID to a per-user task control group. See the documentation on control groups for more information about exactly what that means, and what what some of the commands involved aim to do. I'm not sure if this is exactly the same impact as the kernel-level patch, which aimed at per-TTY control groups. That might includes some processes that don't pass through something that executes the .bashrc file along the way.

Re:How does this work?

[Corbet]

Thanks for the nice words about LWN! Here's a special link to the LWN article on per-tty group scheduling for Slashdot folks. Hopefully a few of you will like what you see and decide to subscribe.

Putting the code in the wrong place

[ras]

An early comment on LWN captured the technical argument best, I think, which I guess illustrates both the quality of the articles and posters on LWN. The background to this is we are discussing CPU scheduling. If you don't know what CPU scheduling is, think of it as form of mind reading. I'll illustrate.

Lets say you have asked your computer to do several things, in fact so many that if it follows the usual method of simply dividing its time equally between them it is going to annoy you. The video you watching might start flickering, or the music you are listening will drop out. So obviously the computer must now give more CPU time to playing your movie and less to whatever background task you started, such as that MP3 transcode of your 20,000 song library. Except how is the computer is supposed to know this? This is how we get to mind reading.

The hack we are discussing is essentially the discovery of a way to read the minds of one particular type of computer user - the Linux Kernel developer. The Linux Kernel developer is in the habit of starting huge background jobs called kernel compiles. These kernel compiles take a looong while, so the kernel developers, being very clever people, have invented all sorts of ways of speeding them up. One of those ways is to divide the task into lots of little bits, and then fire off separate tasks to do each. This takes maximum advantage of available CPU cores, soaking up every skerrick of available CPU time. This naturally enough leaves none left over for other important tasks like watching a movie while waiting your kernel compile. In this particular case the default CPU scheduling strategy of giving each task an equal share of CPU is woefully poor, because there might be 20 kernel compile tasks and just one movie watching task, so the movie player ends up with 1/20 of the available CPU time. This isn't enough to play a movie.

The mind reading trick discovered boils down to this: Linux Kernel developers use the linux command line interface to fire off the kernel compile. And it turns out that for years now the kernel has been able group the tasks started from a command line and give that group a single portion of CPU time, as opposed to a equal portion to each task in the group. Thus you only have to split up the CPU time into 2, one portion going to the kernel compiler group and the other going to the movie player. Naturally enough the movie player works real well with a 50% allocation of CPU, and so we have a happy kernel developer.

Now we come to the merits of the two hacks. They both do the job I just described equally well. The difference between them is that one, the kernel patch, is automagic, meaning it happens automatically without anybody having to lift a finger. But it comes at the expense of bloating linux kernel a tiny bit, even for users who won't benefit from it. The other way currently has to be done applied manually using a process the vast majority of Linux users will at best find difficult, tedious and error prone.

Seems like a simple decision eh - lets take the tiny bloat hit and not inflict our long suffering desktop users with yet another Linux user-unfriendly idiosyncrasy. But here is the rub: it doesn't help them. In fact, for some it might have a negative impact (a gstreamer pipeline started from the command line strings to mind). The people who will benefit from this are the ones that use the command line heavily and regularly. People like Linus. Which is why he liked it so much I guess. But these are precisely the people who will have no absolutely no trouble doing it the manual way.