A FAT32 fragmenter

https://gist.github.com/ryancdotorg/9f557d3513710ce91aed

by jaimebuelta • 11 years ago

110 50 11 years ago

gist.github.com

iso8859-1• 11 years ago

Could actually be useful for testing, maybe?

I can see it used to test defragmentation programs. Fragment the disk into random pieces and then run your defrag program on it and see how long it takes to organize it.

illektr1k• 11 years ago

Sometimes SysAdmins need relax time too!

https://xkcd.com/303/

DiabloD32• 11 years ago

It is against HN rules to merely comment about how awesome something is and not contribute to the conversation in any way.

Screw it, I can't hear your rules over how awesome this is.

dang2• 11 years ago

It is against HN rules to merely comment about how awesome something is

Why do you say that? Of course it isn't.

DiabloD32• 11 years ago

What, really? My bad.

ludston• 11 years ago

It's not against HN rules. :)

mikeash• 11 years ago

This could make for an excellent subtle LART, for when the more traditional ones can't be used.

f_• 11 years ago

This is a pretty neat concept; I've come across this some years back in bisqwit's video https://www.youtube.com/watch?v=lxZyxxHOw3Y - he call's it "enfragmentation".

ryan-c5• 11 years ago

Neat, I wasn't aware of bisqwit's hack. Looks like he supported FAT12 and FAT16 but not FAT32, whereas I only supported FAT32.

His has better status information.

userbinator4• 11 years ago

Would be interesting to see a version for ext* too.

marvy2• 11 years ago

Request for benchmarks: I'd like to see what effect this has on, say, WinXP boot time, or maybe some database benchmark. In fact, there's no need to get so fancy. Let's see how much this slows down computing a file's checksum, or maybe copying a file.

ryan-c5• 11 years ago

(I wrote this horrible thing)

It's pretty horrendously bad on spinning rust - you get very, very close to 100% fragmentation, so the disk has to seek for each cluster. A lower bound average on seek time is probably something like 4ms, so your max transfer rate is going to be limited to 125 * cluster size. Depending on the disk size[0] the cluster size may be up to 32KB, so we're looking at maybe 4MB/sec best case. I'd guesstimate it'd increase boot time by somewhere between 10x and 40x.

0. https://support.microsoft.com/en-us/kb/140365

ryan-c5• 11 years ago

In theory, you could make the disk performance even worse by alternating between the beginning and end of the disk, though you may need to do a little tweaking to ensure this doesn't inadvertently allow read-ahead to help.

userbinator4• 11 years ago

I wonder how much effect it'd have on an SSD. Those tend to have much higher random read and effective seek times far below 1ms.

marvy2• 11 years ago

That sounds plausible. I'm curious how you tested this. Do you have a FAT-formatted disk to play with, or did you do this in a VM, or...?

Aldo_MX• 11 years ago

Don't try this on pendrives or sdcards.

rsi_oww• 11 years ago

I did something like this when I was a kid, in QuickBasic. Not as sophisticated - it would just randomly create, append to, and delete files.

I tested it on floppy disks - if you let it run for a while, the remaining free space was so fragmented it was almost unusable. You could hear the poor floppy drive seeking like crazy just to open a tiny text file.

The fun part was defragmenting it afterwards - I miss the days of the graphical defrag that Norton and MS had.

upperclasstwit• 11 years ago

That was such a pure pleasure. I can't exactly picture it but I remember it was sort of like watching your dwarf fortress.

nitrogen2• 11 years ago

My favorite ones would show you a grid representing all the clusters on the disk. You would see an open block for free space and a solid block for allocated space.

Then it would highlight a span of blocks to indicate reading.

Finally it would write those clusters elsewhere, indicating so with a different color.

Edit: Uhh.. and HN deleted my diagrams. See http://pastebin.com/raw.php?i=nKHddiKe

digi_owl4• 11 years ago

You can get graphical defragers today...

Vexs2• 11 years ago

Defraggler is the only one I can think of off the top of my head. I love that kind of interactivity in applications, so much better than a loading bar.

ambrop73• 11 years ago

Related, I'm doing a FAT32 driver for embedded systems (fully asynchronous!). It's part of my APrinter firmware project: https://github.com/ambrop72/aprinter/tree/master/aprinter/fs

Currently it has good read support and limited write support (can re-write existing files but not append or create new files).

Before writing that code I made some prototype read-only code in Python, to make sure I understand the FS structure properly: https://github.com/ambrop72/aprinter/blob/master/prototyping...

voltagex_• 11 years ago

You got further than me. I got stuck trying to find the files themselves in a disk image. I think the code is in my junkcode repo on GitHub. My end goal is ext4, and then on to a FAT32 over USB gadget hack to bring back mass storage support to Android.

What did you use as your main source of documentation while developing?

ambrop73• 11 years ago

I've mostly used Wikipedia, https://en.wikipedia.org/wiki/Design_of_the_FAT_file_system . Also a little bit the dosfstools source code.

userbinator4• 11 years ago

Embedded as in "32-bit ARM running Linux", or "8-bit microcontroller"? Either way, as someone who has written a FAT driver for embedded systems in the later category, that looks like it's far more code than it needs to be. Full read/write functionality can be done in approximately 800 (Z80) machine instructions. FAT is a linked list, and if you look at it that way, it doesn't take much code to manipulate one.

I've written a bit more about that before:

https://news.ycombinator.com/item?id=7492318

Append = find an empty cluster and link it into the chain for the file, then write into it. Create is similar except you start with a empty chain and also add an entry to the directory (which is like writing/appending to a file, because directories are files.)

When allocating next clusters you can reduce fragmentation significantly over the dumb "first fit" strategy if you scan the FAT to find contiguous free clusters and apply a next/best/worst fit. Even better if you add an API to allow tuning the amount of "gap" after a file when creating it, based on knowledge of how much it may expand in the future.

I believe that with good tuning and allocation heuristics, FAT32 can outperform the far more complex filesystems (ext*, NTFS, etc.) widely believed to be superior. One idea I've never gotten around to testing out is to modify the Linux FAT driver and do some benchmarking.

JohnBooty• 11 years ago

   I believe that with good tuning and allocation 
   heuristics, FAT32 can outperform the far more complex
   filesystems (ext*, NTFS, etc.) widely believed to be
   superior. One idea I've never gotten around to testing
   out is to modify the Linux FAT driver and do some
   benchmarking.

I don't understand. Has anybody ever claimed that the more complex alternatives were actually faster?

NTFS and other filesystems that followed FAT32 are "superior" because they support things like journaling and more robust permissions... things that unavoidably incur (a least) a small performance hit.

ambrop73• 11 years ago

It is targeted to ARM uC's and also works on AVRs with enough RAM and program space (yes I've tested it). Write support requires a significant amount of RAM; I think about six 512-byte blocks. When compiled in read-only mode, it should need no more than two.

One reason for the RAM needs is that I use a block cache and I take care do to "atomically" certain related operations (on the block cache level). These are things like marking a cluster as allocated, decrementing the number of free clusters in FsInfo, and linking the allocated cluster to the end of a chain. These things require multiple blocks to all be present in the block cache at the same time.

I suppose the asynchronous design also contributes to the code and RAM size (though I did not actually measure the code size or compare it to other drivers). It may look like more that it really compiles to. Anyway would be very surprised if you show me another asynchronous driver - I couldn't find any!

Also consider that my code supports VFATs (decoding only, file creation is not supported anyway).

The design is intentional, I have sacrificed some RAM/program size for asynchronous operation and better reliability in case of random write failures (because of the block cache, all writes can reliably be retried later, avoiding corruption as long as they eventually complete).

Update: I estimate the code compiles to about 10kB with gcc for Coretex-M3 using -Os. This is certainly within limits of most ARM uCs and many AVRs.

raverbashing• 11 years ago

So what happens exactly if you run this on Ext3/4, HFS, NTFS or other modern fss?

delan• 11 years ago

Probably fail spectacularly, as it seems to make modifications to the FAT32 data structures directly. A more generic fragmenter might interact with a file system via pathological file access patterns, but this doesn’t take that approach.

Sami_Lehtinen• 11 years ago

I've usually done intentional fragmentation by filling disk with small files and deleting those while growing a new file to claim those blocks. Works basically with every file system. Some examples where I've used such method.

People often claim that fragmentation doesn't affect SSD drives, but that's not true: http://www.sami-lehtinen.net/blog/ssd-file-fragmentation-myt...

This is slightly related. How contiguously growing files are allocated on different file systems: http://www.sami-lehtinen.net/blog/test-btrfs-ext4-ntfs-simpl...

userbinator4• 11 years ago

On the other hand, your worst fragmented case still gets ~200MB/s with a random read, which is around the best case with a sequential read for most HDDs today. A 4k random read on a HDD will be 1-2MB/s at most.

Relatively speaking the SSD slows down to ~50% of its max speed with fragmentation, but the HDD will be down to around 1%. So fragmentation affects SSDs somewhat, but HDDs are affected much more severely.

(Which SSD was it, and how was the test setup? That's important for comparison purposes, as the layout of the filesystem blocks and how they correspond to the NAND eraseblocks/pages has a huge effect on what fragmentation will do.)

0x4a42• 11 years ago

A fragmenter driven by the game of life with graphics feedback would be prety cool. :D

Shivetya• 11 years ago

Having been out of the FAT32 disk environment for a while, is fragmentation determined by non contiguous sectors on the drive? Do the methods reading data from the drives consider the number of heads/platters involved for performance gains? A file would not need to be contiguous necessarily per platter but some logic employed would map it to where following sectors would pass under a drive head at the same time.

Granted that would not work out for SSD, but I know next to nothing on their addressing

AndrewStephens• 11 years ago

Do the methods reading data from the drives consider the number of heads/platters involved for performance gains?

In general no, at least not since the early 80s. Although you still hear about cylinders, heads, and sectors, spinning hard drives have effectively been a black box to OSes for 30 years and will internally remap sectors wherever they feel like it (and lie about it if you ask them).

The result is that while you can be reasonably confident that sector 33566 is followed by sector 33567 and reading both will not involve seeking, things like reading the whole cylinders at a time are not worth the effort since you don't know where the sectors are.

A FAT32 fragmenter

Related Stories