memory-hierarchy-and-caches

Memory Hierarchy and Caches

Purpose

Teach CPU memory hierarchy: cache levels, associativity, line size, coherence protocols, false sharing, and prefetch behavior — architectural depth complementing optimization practice in skills/low-level-programming/cpu-cache-opt.

When to Use

• Explaining why padding fixes scalability
• Choosing struct layout for multicore
• Interpreting cache miss counters
• Understanding DMA vs CPU cache on embedded SoCs

Workflow

1. Hierarchy (typical desktop/ server)

Registers
├── L1d / L1i (per core, ~32 KiB, ~4 cycles)
├── L2 (per core, ~256 KiB – 1 MiB)
├── L3 (shared last-level, MiB – tens of MiB)
├── DRAM (hundreds of cycles)
└── Storage / NUMA remote (much slower)

Embedded MCUs may have only tightly-coupled memory (no L2/L3).

2. Cache line and associativity

• Line size: commonly 64 bytes on x86/ARM64 (verify with getconf LEVEL1_DCACHE_LINESIZE)
• Set-associative: line maps to one set, competes within ways
• Conflict misses: many aliases same set

3. False sharing

/* Bad — two atomics on same cache line */
struct {
    atomic_int counter_a;
    atomic_int counter_b;
} stats;

/* Good — pad to cache line */
struct alignas(64) {
    atomic_int counter_a;
    char pad[64 - sizeof(atomic_int)];
    atomic_int counter_b;
} stats;

4. Coherence (multicore)

MESI states: Modified, Exclusive, Shared, Invalid. Writes invalidate other cores' copies of the line — why atomics and locks ping cache lines.

5. Prefetching

#ifdef __builtin_prefetch
for (int i = 0; i < n; i++) {
    __builtin_prefetch(&data[i + 8], 0, 3);
    process(data[i]);
}
#endif

Hardware stride prefetchers detect sequential access; random access misses.

6. Measurement

perf stat -e cache-references,cache-misses,L1-dcache-load-misses ./app

7. Agent usage

/memory-hierarchy-and-caches Diagnose false sharing in this per-thread stats array

Common Problems

Symptom	Cause	Fix
Scaling collapses	False sharing	Line-align per-thread data
High LLC misses	Working set > cache	Block algorithms; NUMA pin
DMA incoherence	CPU cache vs device	Flush/invalidate on MCU; dma_sync on Linux
Prefetch hurt	irregular access	Remove manual prefetch
Huge struct copies	AoS cold lines	SoA layout — see cpu-cache-opt