Allocation decisions and the cost of fragmented space

Internal fragmentation is unused space within allocated units. If a simplified filesystem dedicates whole blocks to each small file, the final block may contain slack. Free-space fragmentation instead describes how unused blocks are split into runs. File fragmentation describes one file's contents spread across multiple mappings. A volume with many scattered free blocks can still store a new file, yet require more extents. Do not add these quantities as though they were the same percentage; each uses a different denominator and describes a different mechanism.

Application access order matters. A large sequential scan can benefit from long contiguous mappings because they reduce mapping work and support larger requests. An application that reads isolated random blocks may receive less benefit from making every file sequential. On a rotating disk, movement between distant requests adds a mechanical cost. An SSD has no seek arm, but request count, controller parallelism, translation, and transfer size still matter. Therefore a fragmented-file count is a clue, not a direct measurement of elapsed-time loss across all devices.

Ext4 delayed allocation can postpone choosing data-block locations until dirty data is written back. Seeing a larger span of writes together can improve placement compared with allocating independently for every tiny arrival. The tradeoff is that accepting buffered writes is not equivalent to having permanently stored them, and allocation or I/O errors still require handling. A reservation, an allocated mapping, and durable contents are different milestones. Record which milestone an API promises before using a free-space observation as evidence that a checkpoint can be published.

Preallocation can establish space before a stream of later writes and reduce some fragmentation risks, but its precise semantics depend on the operation and filesystem. It can also reserve capacity that another task needed. Workloads that grow many files concurrently, retain snapshots, or rewrite small regions can defeat a placement strategy that looked effective on an empty volume. Keep room for metadata and future updates rather than targeting a universal safe free-space percentage. Capacity plans should include measured behavior near their intended operating range, not merely a successful initial allocation.

Compare a representative operation before and after a placement change while controlling file contents, cache state, concurrency, and available capacity. Observe both extent layout and block-request behavior. A faster second run might simply be cached. A defragmentation operation can itself generate heavy I/O, alter shared-extent storage consumption on some copy-on-write filesystems, or compete with user traffic. Treat it as a workload with a purpose and a rollback or recovery plan, rather than a cosmetic way to reduce a counter.

The quadratic lab is a deliberate scaling analogy for metadata-heavy all-pairs inspection. Interpret positive x as normalized extent count and positive a as work per pair for a naive comparison procedure. Double x and predict fourfold y before moving the control. This illustrates why a poor diagnostic algorithm becomes expensive as mappings multiply; it does not predict fragmentation or defragment a filesystem. A practical investigation should distinguish the cost of the layout from the cost of the tool inspecting it, and favor indexed or streaming analysis when appropriate.