How Slow Are Disk and Network? A CPU's Perspective on Real‑World Latency

Let the CPU Tell You How Slow Disk and Network Really Are

People often say “disk is slow” or “network is laggy” based on human perception, but from a CPU’s point of view the time scales are dramatically different. This article uses vivid analogies and data from public benchmarks to translate hardware latencies into human‑scale equivalents, helping readers grasp the magnitude of cache, memory, I/O and network delays.

Introduction

The CPU runs at about 2.6 GHz, executing roughly 2.6*10^9 instructions per second, each taking ~ 0.38ns. We treat this as the base unit of one second in human terms.

Cache Latencies

L1 cache read time is ~ 0.5ns, which corresponds to about 1.3s for a human—roughly one heartbeat. Branch‑prediction miss costs ~ 5ns (~ 13s). L2 cache takes ~ 7ns (~ 18.2s).

Lock and Memory Access

A mutex lock/unlock takes ~ 25ns (~ 65s, a little over a minute). Main memory access is ~ 100ns (~ 260s, about 4 minutes). A context switch costs ~ 1500ns (~ 1.5µs), which translates to roughly 65 minutes of human time.

Network and Storage

Transferring 2 KB over a 1 Gbps link takes ~ 20µs, equivalent to 14.4 hours. SSD random read (~ 150µs) equals 4.5 days. Reading 1 MB sequentially from SSD (~ 1ms) feels like 1 month. Mechanical HDD seek (~ 10ms) maps to 10 months, and reading 1 MB takes ~ 20ms (~ 20 months).

Data‑Center and Global Latency

A round‑trip within a data‑center (~ 0.5ms) feels like 15 days. A cross‑city round‑trip (~ 150ms) feels like 12.5 years. Rebooting a VM (~ 4s) corresponds to 300+ years, while a physical server reboot (~ 5min) maps to 25,000 years.

Takeaway

Disk and network are indeed extremely slow from the CPU’s perspective, and performance optimization must consider these orders‑of‑magnitude differences.