The Hairy State of Linux Filesystems

RazvanM writes "Do the OSes really shrink? Perhaps the user space (MySQL, CUPS) is getting slimmer, but how about the internals? Using as a metric the number of external calls between the filesystem modules and the rest of the Linux kernel I argue that this is not the case. The evidence is a graph that shows the evolution of 15 filesystems from 2.6.11 to 2.6.28 along with the current state (2.6.28) for 24 filesystems. Some filesystems that stand out are: nfs for leading in both number of calls and speed of growth; ext4 and fuse for their above-average speed of growth and 9p for its roller coaster path."

Comments

Yes/no

[EmbeddedJanitor]

The number of calls in the interface do matter because they increase complexity. This makes fs maintainability and development a bit harder from version to version as it gets less clear what each call should do. Many of the calls are optional, or can be performed by defaults, which does help to simplify things.

There is little calling overhead from using multiple calls. Of course these interface changes are all done for a good reason: performance, stability, security.

Re:Yes/no

[Suzuran]

Function calls are not free. Especially in kernel space. Everything costs time. You need to do the most you can with the least instructions. 100 lines of inline code will probably run faster than a function call.

Re:Yes/no

[hackerjoe]

What's your point? Processors can pipeline across branches just fine, and the main effect of cache is to give a performance boost to smaller code -- code that separates and reuses functions rather than inlining them willy-nilly.

Inlining can still be a win sometimes, but compilers will do that for you automatically anyway...

Re:Yes/no

[ultranova]

Funny thing? A hundred increments runs within 1% of the speed of 100 calls to a function to do the increment. And yes I unrolled those calls to isolate the cost of what I was measuring. So... rather surprisingly, the cost of these function calls is as close as doesn't matter, to exactly zero.

Not at all surprisingly, since 100 function calls and 100 integer additions will take so little time on a modern processor - and, I suspect, would even on an 8088 - that they amount to a rounding error. The machine's clock doesn't have sufficient resolution to measure them. You'd need a hundred million for a meaningful comparison.

not following

[Eil]

What's your argument here? That filesystem code in the kernel shouldn't be growing more sophisticated over time?

This rings of the armchair-pundit argument that the kernel is getting more and more "bloated" and a breath later crying out that there still aren't Linux hardware drivers for every computing device ever made.

Re:not following

[doshell]

I have a good idea to get the drivers while still eliminating the bloat.

Have an option to compile the kernel during installation, based on detected devices.

And then whenever you buy a new webcam, replace your graphics card, or whatever, the kernel must be recompiled. People will love that.

Also: the Linux kernel is modular. This means you don't actually hold in memory, at every time, the drivers to all the hardware devices known to man. Only those your machine actually needs to run. The remaining are just files sitting on your hard disk, ready to be loaded should the need arise. This is an adequate way to keep down the bloat while not inconveniencing the user every time a new piece of hardware pops up.

Re:Where's NTFS ?

[quickOnTheUptake]

Did you know that Linux has limited NTFS support? I usually have to create a FAT32 partition to copy files between Windows XP and Linux. NTFS is usually read only or not available.

Have you heard of NTFS-3G?

The NTFS-3G driver is a freely and commercially available and supported read/write NTFS driver for Linux, FreeBSD, Mac OS X, NetBSD, Solaris, Haiku, and other operating systems. It provides safe and fast handling of the Windows XP, Windows Server 2003, Windows 2000, Windows Vista and Windows Server 2008 file systems.

Goofy metric, too.

[Ungrounded Lightning]

Unless I've misread it, TFA's definition of "size" for a filesystem is "how many distinct external/kernel subroutines does it call?"

That seems to be a very strange metric. Seems to me that a well-designed filesystem will have little code and make extensive (re)use of well-defined services elsewhere (rather than reinventing those wheels). This "slims" the filesystem's own unique code down to its own core functionality.

Now maybe, if the functions are hooks in the OS for the filesystem, a larger number of them implies a poorer abstraction boundary. But it's not clear to me, absent more explanation, that this is what TFA is claiming.