perf: align ALIGNMENT to cache line size (16 → 64 bytes) to fix false sharing

[viknash]

Problem

ALIGNMENT in �lignment.h was set to 16 bytes — only ¼ of a typical CPU cache line (64 bytes). Every type annotated with \class_alignment\ / \�lignas(ALIGNMENT)\ and every struct packed via #pragma pack(ALIGNMENT)\ could have multiple independently-written fields share a single 64-byte cache line.

The most acute hotspot is \queue_rank[num_priority][max_num_threads]\ in \ hreadpool.h:

\\cpp
optimization std::atomic queue_rank[task_type::num_priority][max_num_threads];
\\

ank_type = int64_t\ (8 bytes). With \ALIGNMENT=16, four thread slots shared one cache line. Every worker thread writes its own rank slot on every task dispatch and steal — so those four cores constantly bounce the same cache line across the L3, causing O(N²) coherency traffic under thread count. This is textbook false sharing.

Root cause

\cache.h\ already documents the correct value:
\\cpp
uint32_t get_cache_line_size() { return 64; }
\
\ALIGNMENT\ was just never updated to match.

Fix

One-line change in \�lignment.h:

\\diff
-#define ALIGNMENT 16
+// Match the CPU cache line size (see cache.h) to prevent false sharing between
+// concurrently-written fields that happen to land on the same cache line.
+#define ALIGNMENT 64
\\

This propagates through every \class_alignment\ and #pragma pack\ use site in the codebase, ensuring hot shared-state structs never straddle a cache line boundary.

Impact

• Eliminates false sharing on \queue_rank\ (written by every thread on every dispatch/steal)
• Fixes alignment for all \class_alignment-annotated types (\lock_free_node_stack, \multi_producer_multi_consumer, etc.)
• Zero risk of behavioural change — a larger alignment is always ABI-safe for in-process objects
• Consistent with the existing \get_cache_line_size() = 64\ contract

Other identified improvements (follow-up PRs)

This is the highest-impact, lowest-effort item from a broader audit. Other improvements to address:

#	Issue	Effort
2	CAS spin loops missing _mm_pause()\ backoff	Easy
3	\�dd_task_parallel_work\ copies \std::function\ twice	Easy
4	Per-frame \
ew/\delete\ of \data_dispatcher_type\	Moderate
5	\seq_cst\ where \�cq_rel/\
elaxed\ suffice	Easy
6	\�vent::wait\ spurious-wakeup bug + \�larm::sleep\ copies \std::function\	Easy
7	Virtual dispatch on hot-path \push_back/\pop_back\	Easy
8	\lock_free_node_dispenser\ constructor UB + fallback \
ew\	Moderate
9	\operator new\ profiling overhead + wasted alignment padding	Moderate
10	\ ask_name.c_str()\ called twice per task invocation	Easy

[Copilot]

Pull request overview

Updates the project-wide alignment constant in include/alignment.h to target cache-line alignment, aiming to reduce false sharing in concurrent hot paths by increasing ALIGNMENT from 16 to 64 bytes.

Changes:

• Set ALIGNMENT to 64 and added an explanatory comment referencing cache-line sizing.
• Propagates the new ALIGNMENT value to alignas(ALIGNMENT) usage via class_alignment.
• Indirectly changes #pragma pack(ALIGNMENT) to #pragma pack(64).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.