Understanding False Sharing: CPU Cache, MESI Protocol, and Java Mitigation Techniques

In concurrent programming the focus is often on controlling access to shared variables, yet hardware factors such as CPU caches and the JVM can silently degrade performance through false sharing, where independent variables share the same cache line.

CPU Cache – Cache memory sits between the CPU and main memory, providing faster access to frequently used data. Modern CPUs have multiple cache levels (L1, L2, L3) with decreasing size and speed, and cache misses force costly memory accesses.

MESI Protocol and RFO – The MESI (Modified, Exclusive, Shared, Invalid) protocol maintains cache coherence across cores. When a core needs a cache line owned by another core, it issues a Request‑For‑Owner (RFO) message, causing the line to transition states and incurring high latency.

Cache Line – Typically 64 bytes, a cache line holds a contiguous block of memory. If several threads modify different variables that reside on the same line, each write triggers RFO traffic, leading to the false‑sharing problem.

False Sharing Example – The article presents a Java benchmark (see public class FalseShareTest implements Runnable { ... } ) where four threads update separate elements of an array of VolatileLong . When the padding fields are omitted, all elements share a cache line and the execution time is roughly 2.5× higher than when padding is added.

Mitigation Techniques – Adding unused long fields (padding) to expand each object to 64 bytes separates objects onto different cache lines. Alternative approaches include using compiler directives such as __declspec(align(64)) or manually aligning structures and arrays, and inserting dummy calculations to prevent compiler optimization of padding fields.

Practical Considerations – False sharing is hard to detect, varies across architectures, and excessive padding can waste cache space. Therefore, developers should weigh the cost‑benefit of mitigation, especially on platforms where cache hit rates are already high.

References are provided for further reading on false sharing from Java perspectives and data‑fusion methods.