Sequential and random I/O in media workloads

Sequential and random I/O describe the order and size of storage requests, which can affect media reads and library work very differently.

Sequential I/O accesses successive ranges. Random I/O accesses ranges that are scattered relative to the preceding requests. These descriptions apply to an observed request stream: a file can be large while the application accesses it randomly, and many small files can sometimes be serviced efficiently from cache.

Playback often reads forward through a media file. Seeking moves to another range. Scanning and artwork access produce a mixture of directory, metadata and content operations.

Same device, different cost per useful byte

Consider two original uncached workloads. Assume a device transfers data at 150 decimal MB/s, and every scattered operation adds 8 ms of positioning or other service delay. There is one operation at a time.

4 KiB operation:
  transfer = 4,096 ÷ 150,000,000 seconds = 0.0273 ms
  total    = 8 + 0.0273 = 8.0273 ms
  useful rate ≈ 0.510 MB/s

1 MiB operation with the same added delay:
  transfer = 1,048,576 ÷ 150,000,000 seconds = 6.9905 ms
  total    = 8 + 6.9905 = 14.9905 ms
  useful rate ≈ 69.949 MB/s

The device's assumed transfer rate did not change. Larger operations amortized the fixed delay over more bytes. Long adjacent runs can reduce repeated positioning further. These figures describe a deliberately simplified model, not a representative HDD or SSD benchmark.

A large adjacent read spends proportionally less time on repeated fixed operation costs than separate small reads.

Why IOPS needs an operation size

IOPS means input/output operations per second. Multiplying by the average bytes per operation gives a useful-data rate only for the workload being described.

Operation rate Operation size Data rate
1,000 IOPS 4 KiB 4.096 decimal MB/s
1,000 IOPS 1 MiB 1,048.576 decimal MB/s

The equal operation counts do not imply equal work. Read/write mix, queue depth, caching and whether operations are dependent also belong in a meaningful comparison.

Individually sequential, collectively mixed

Four players can each read sequentially while the device alternates among four files. On an HDD, this may require head movement. Read-ahead can gather larger ranges and the operating system can combine or reorder requests, so the physical device may see a different pattern from the application's calls.

This is why a single large-file copy is incomplete evidence for an Emby server's mixed workload. If a film reads smoothly alone and slows during a scan, the relevant change may be the ordering and latency of requests, rather than a lack of average byte-rate capacity.

See Also

References