Understand NUMA: Non-Uniform Memory Access in Multi-Socket Systems

Non-Uniform Memory Access

Multi-socket servers attach local DRAM to each CPU. Accessing another socket's memory crosses an interconnect (Intel UPI, AMD Infinity Fabric) with higher latency and limited bandwidth. The OS and allocator should prefer local nodes.

bashCopy
Loading…

Observing Remote Penalty

• Local DRAM: ~80 ns
• Remote DRAM: ~120 to 150 ns
• First-touch policy: page fault on thread's current CPU binds memory there

For example, allocating on node 0 then letting threads on node 1 hammer the buffer can halve bandwidth versus node-local allocation.

For this exercise, you will run the same bandwidth microbenchmark under numactl policies and compare nodes. This task asks you to quantify why PostgreSQL, Redis, and HPC jobs pin threads near their memory.

Working Through the Exercise

Keep the relevant datasheet, ISA manual, or architecture textbook chapter open while you implement. When your output disagrees with the reference trace on the same program, the bug is usually a mis-decoded opcode, a stale register read, or a flag bit left unchanged after arithmetic.

• Trace first: walk one failing input by hand before changing random lines
• Read the manual: skim the official doc for the mechanism named in this task
• Compare baselines: save measurements from the prior step so speedups are honest

For this exercise, you will use those habits while implementing the requirement in the starter code. Microarchitectural product names change across CPU generations, but the control ideas (fetch, bypass, cache lines, vector lanes) stay stable enough to debug from first principles.