THE FOLD / GRIND / WARM CACHE / THE TLB REACH
THE TLB REACH
a unit trick, not a capacity gain
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
The TLB caches address translations, and it holds a fixed number of them — not a fixed amount of memory. Multiply entries by page size and you get the only figure that matters: how much memory the machine can currently name without a walk.
LIT verified live. 1,536 entries with 4 KB pages reach 6 MB. The same 1,536 entries with 2 MB pages reach 3,072 MB — 512× further, with no extra silicon. Against a 512 MB working set that is the difference between 98.8% uncovered and 0%. One-gigabyte pages reach 1,572,864 MB.
LIT verified live. 1,536 entries with 4 KB pages reach 6 MB. The same 1,536 entries with 2 MB pages reach 3,072 MB — 512× further, with no extra silicon. Against a 512 MB working set that is the difference between 98.8% uncovered and 0%. One-gigabyte pages reach 1,572,864 MB.
2 HOW IT WAS WEAVED · AI + HUMAN
TLB reach is standard architecture vocabulary; the number is rarely printed because it is embarrassing.
AVAN (AI) computed the coverage against a stated working set rather than quoting entry counts. 1,536 entries sounds generous and 6 MB does not, and they are the same fact. The whole case for huge pages is in that pair — the cache did not get bigger, the unit of account did.
AVAN (AI) computed the coverage against a stated working set rather than quoting entry counts. 1,536 entries sounds generous and 6 MB does not, and they are the same fact. The whole case for huge pages is in that pair — the cache did not get bigger, the unit of account did.
3 ONE DIMENSION
Same entry count, three page sizes.
4 TWO DIMENSIONS · INTERACTIVE
Grow the working set until the TLB stops covering it.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: a cache measured in names, not bytes.
AVAN’s addition (the inverse-companion): the forward reading is that huge pages extend TLB reach enormously. The inverse is that reach was never a property of the TLB. The hardware is identical in all three columns; what changed is how much territory one name is allowed to claim. Read backwards, this is a unit trick rather than a capacity gain — and it works precisely because a page is an accounting fiction, so making the fiction coarser costs nothing until the day you need to say something finer.
LIT 1,536 entries with 4 KB pages reach 6 MB while the same 1,536 entries with 2 MB pages reach 3,072 MB - 512 times further with no extra silicon - which against a 512 MB working set is the difference between 98.8% uncovered and 0%, and one-gigabyte pages reach 1,572,864 MB
FIG TLB reach is standard architecture vocabulary; the number is rarely printed because it is embarrassing. AVAN computed the coverage against a stated working set rather than quoting entry counts. 1,536 entries sounds generous and 6 MB does not, and they are the same fact. The whole case for huge pages is in that pair - the cache did not get bigger, the unit of account did.
FIG TLB reach is standard architecture vocabulary; the number is rarely printed because it is embarrassing. AVAN computed the coverage against a stated working set rather than quoting entry counts. 1,536 entries sounds generous and 6 MB does not, and they are the same fact. The whole case for huge pages is in that pair - the cache did not get bigger, the unit of account did.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of WARM CACHE · David Lee Wise (ROOT0), with AVAN