THE FOLD / CO-OP / SHARED MEMORY / THE NUMA HOP
THE NUMA HOP
the flat address space was the lie, and a load-bearing one
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
On a multi-socket machine the memory is not one pool. Some of it is attached to your socket and some to the other one, and reaching across costs a fixed toll on every single access.
LIT verified live, as a stated model: local 80 ns, remote 140 ns — a 75.0% penalty per access. All-local runs at 80 ns and all-remote at 140, a 1.75× slowdown for identical code on identical data. An interleaved policy over 100,000 accesses landed remote 49.94% of the time and averaged 109.96 ns — almost exactly halfway, because interleaving does not avoid the toll, it splits it.
LIT verified live, as a stated model: local 80 ns, remote 140 ns — a 75.0% penalty per access. All-local runs at 80 ns and all-remote at 140, a 1.75× slowdown for identical code on identical data. An interleaved policy over 100,000 accesses landed remote 49.94% of the time and averaged 109.96 ns — almost exactly halfway, because interleaving does not avoid the toll, it splits it.
2 HOW IT WAS WEAVED · AI + HUMAN
The 80/140 figures are a stated model, not a measurement of your machine — the latencies are hardware-specific. What is measured here is the arithmetic they imply and the interleaving.
AVAN (AI) is explicit about that boundary because it is where this kind of sphere usually cheats. The 1.75× follows from the two constants and nothing else; the 49.94% is a real draw from a real generator over 100,000 trials. Naming which numbers are assumed and which are measured is the difference between a model and a claim.
AVAN (AI) is explicit about that boundary because it is where this kind of sphere usually cheats. The 1.75× follows from the two constants and nothing else; the 49.94% is a real draw from a real generator over 100,000 trials. Naming which numbers are assumed and which are measured is the difference between a model and a claim.
3 ONE DIMENSION
Mean latency against the fraction of accesses that cross.
4 TWO DIMENSIONS · INTERACTIVE
Move the placement policy and watch the mean move with it.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: two pools wearing one address space.
AVAN’s addition (the inverse-companion): the forward reading is that NUMA makes some memory slower. The inverse is that the flat address space was the lie, and it is a load-bearing one. Every pointer looks the same and dereferences the same way; nothing in the type system, the language or the instruction encoding admits that two of them differ by 75%. Read backwards, the uniform address space is what made portable software possible, and NUMA is the bill for a fiction that was worth every penny until the machine stopped being one machine.
LIT with local 80 ns and remote 140 ns as stated constants - a 75.0% penalty per access - all-local runs at 80 ns and all-remote at 140, a 1.75x slowdown for identical code on identical data, while an interleaved policy over 100,000 measured accesses landed remote 49.94% of the time and averaged 109.96 ns, almost exactly halfway
FIG The 80/140 figures are a stated model, not a measurement of your machine - the latencies are hardware-specific. What is measured is the arithmetic they imply and the interleaving. AVAN is explicit about that boundary because it is where this kind of sphere usually cheats: the 1.75x follows from the two constants and nothing else, while the 49.94% is a real draw over 100,000 trials. Naming which numbers are assumed and which are measured is the difference between a model and a claim.
FIG The 80/140 figures are a stated model, not a measurement of your machine - the latencies are hardware-specific. What is measured is the arithmetic they imply and the interleaving. AVAN is explicit about that boundary because it is where this kind of sphere usually cheats: the 1.75x follows from the two constants and nothing else, while the 49.94% is a real draw over 100,000 trials. Naming which numbers are assumed and which are measured is the difference between a model and a claim.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of SHARED MEMORY · David Lee Wise (ROOT0), with AVAN