Laying out the disk

A subject the papers are about. The loosest grouping, and the one to reach for last.

File system design chasing a moving target: each layout is optimal for the hardware behaviour of its decade, and then the hardware changes.

Four records, and the argument is legible only in sequence. Ritchie and Thompson's UNIX paper describes a file system that is simple and does not think about where blocks land, because on the hardware of 1974 it did not have to matter as much as simplicity did. McKusick's fast file system in 1984 is the correction: cylinder groups, larger blocks, and an explicit model of the disk's geometry, because seek time had become the cost. Rosenblum and Ousterhout's log-structured file system in 1992 inverts it again -- if memory is large enough to absorb the reads, then writes are what remain, so lay everything out as a log and never seek to update in place.

Each is right about its decade and wrong about the next. That is the set's whole claim, and it is why the newest member is a survey of *distributed* file system design choices rather than another layout: the target moved off the platter entirely, and the question became which node holds the block rather than which track. Flash removed the seek that the fast file system was built around, and the network reintroduced a latency an order of magnitude larger.

A reader wanting the shortest version should read McKusick and Rosenblum back to back. They are eight years apart, both correct, and they disagree about the most basic thing a file system does, which is where to put the next block.