The I/O Cost Model¶

I/O operations are orders of magnitude slower than computation, making I/O the dominant cost in most systems and the primary target for optimization.

Tension Addressed¶

Latency vs. Throughput

Recognizable As¶

Any system design that batches operations, uses caching, or carefully manages when data is read/written; performance analyses that focus on I/O counts rather than CPU cycles; the phrase "I/O bound"

Canonical Sources¶

Patterson & Hennessy, "Computer Architecture: A Quantitative Approach" (multiple editions); Jim Gray, "The Five-Minute Rule" (1987, updated 1997, 2007)

Transfers To¶

Database design, distributed systems (network I/O), web performance, mobile development, any system with external data dependencies

The latency hierarchy spans roughly six orders of magnitude: CPU registers (sub-nanosecond), L1 cache (~1ns), main memory (~100ns), SSD (~100 microseconds), HDD (~10 milliseconds), network (milliseconds to hundreds of milliseconds). Understanding where you sit on this hierarchy determines your optimization strategy.