How to write fast Java code – thinking about memory by Anders Peterson

In this talk I'll discuss things that affect (CPU bound) performance. A key message is that Java developers DO need to worry about memory, even if garbage collection rarely is a problem.

Much of the talk is focused around a demo of a simple benchmark.

Anders Peterson, Optimatika

Recorded at Jfokus 2023 in Stockholm 7th of february

[RomanTchekashov]

The method loopICJ works much faster than loopJCI due to better cache locality and memory access patterns.
Here's a detailed explanation:
1. Cache Locality:
- Modern CPUs have multiple levels of cache (L1, L2, L3) to speed up memory access.
- When data is accessed, it is loaded into the cache in blocks (cache lines).
- Accessing data sequentially (in a linear fashion) is faster because it takes advantage of spatial locality, meaning consecutive memory accesses are likely to be in the same cache line.

2. Memory Access Patterns:
- In loopICJ, the innermost loop iterates over the right matrix's rows (right[c][j]), which means it accesses elements in a row-major order (sequentially in memory).
- In loopJCI, the innermost loop iterates over the left matrix's columns (left[i][c]), which means it accesses elements in a column-major order (non-sequentially in memory).

3. Impact on Performance:
- loopICJ benefits from better cache utilization because it accesses memory in a more cache-friendly manner.
- loopJCI suffers from poor cache performance because it accesses memory in a way that causes more cache misses (accessing columns instead of rows).

In summary, loopICJ is faster because it accesses memory in a way that is more efficient for the CPU cache, leading to fewer cache misses and better overall performance.