Can anyone explain why a log-structured file system is a good fit for flash? It seems that a lot of the benefits of an LFS are that it reduces disk seeks.
Furthermore if you are going to overwrite something in flash it usually behooves you to overwrite an entire block at a time. For these reasons flash systems tend to try to often write changes in a log manner, wait for a lot of changes to pile up and then incorporate those changes back into the original by rewriting entire blocks. The log can be placed in an initialized section of the memory where overwrites are not necessary.
Well, all block devices are written a block at a time. The way flash differs is that there are two internal block sizes -- called pages (4-8kB), and eraseblocks (256kB-4MB).
There are two states a page can be in: clear, or used. You can write at a page granularity, but only to clear pages. To clear used pages, you need to clear the entire eraseblock at a time. Eraseblock sizes are now limited by fundamental physical limits -- on every NAND node shrink from now, the eraseblocks are going to double in size. And they are already really huge in modern devices.
Log-structured file systems tend to write data sequentially since the writes append to the log (yes - that's a gross over simplification).
File systems like ext4 maintain separate data structures for data and metadata and generate more random writes.
It's a perfect fit for flash, because there read seeks are (almost) free, and the overwrite-less, sequential writes are a perfect fit. Flash only supports read, erase, and write-on-erased operations, the exact set that LFS uses.