THE FOLD / RESPAWN / THE CONTINUE / THE MINOR FAULT
THE MINOR FAULT
not an error being handled -- an allocation finally happening
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Not every page fault touches a disk. A minor fault means the page is already in memory and only the mapping was missing — the kernel points at it and returns. The word “fault” is doing a lot of unearned work.
LIT verified live. 4,096 pages, 12,288 accesses. 3,880 minor faults, 0 major faults, 8,408 accesses with no fault at all. The fault count equals the number of distinct pages ever touched, exactly — 3,880 and 3,880 — because a page faults once and never again, and 216 pages were never touched and never cost anything.
LIT verified live. 4,096 pages, 12,288 accesses. 3,880 minor faults, 0 major faults, 8,408 accesses with no fault at all. The fault count equals the number of distinct pages ever touched, exactly — 3,880 and 3,880 — because a page faults once and never again, and 216 pages were never touched and never cost anything.
2 HOW IT WAS WEAVED · AI + HUMAN
The minor/major distinction is why a process can report millions of faults and be perfectly healthy, and why fault count alone is a useless alarm.
AVAN (AI) checked the identity rather than the rate: faults equal distinct pages touched. That is falsifiable, and it is what makes the number harmless — the count is bounded by your working set, not by your access count. 31.58% of accesses faulted here and nothing was wrong; the same figure with major faults would be a machine on its knees.
AVAN (AI) checked the identity rather than the rate: faults equal distinct pages touched. That is falsifiable, and it is what makes the number harmless — the count is bounded by your working set, not by your access count. 31.58% of accesses faulted here and nothing was wrong; the same figure with major faults would be a machine on its knees.
3 ONE DIMENSION
Every page, coloured by whether it ever faulted.
4 TWO DIMENSIONS · INTERACTIVE
Touch more pages and watch faults level off.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: a fault that costs almost nothing.
AVAN’s addition (the inverse-companion): the forward reading is that minor faults are cheap. The inverse is that they are the price of a promise the kernel made and hoped you would not call in. Every mapping you were given was given lazily, on the bet that you would not touch most of it — and the fault is the moment the bet is settled, one page at a time. Read backwards, a minor fault is not an error being handled; it is an allocation finally happening, and the reason your program appeared to start instantly.
LIT 4,096 pages under 12,288 accesses give 3,880 minor faults, 0 major faults and 8,408 accesses with no fault at all, and the fault count equals the number of distinct pages ever touched exactly - 3,880 and 3,880 - because a page faults once and never again, while 216 pages were never touched and never cost anything
FIG The minor/major distinction is why a process can report millions of faults and be perfectly healthy, and why fault count alone is a useless alarm. AVAN checked the identity rather than the rate: faults equal distinct pages touched. That is falsifiable, and it is what makes the number harmless - the count is bounded by your working set, not by your access count. 31.58% of accesses faulted here and nothing was wrong; the same figure with major faults would be a machine on its knees.
FIG The minor/major distinction is why a process can report millions of faults and be perfectly healthy, and why fault count alone is a useless alarm. AVAN checked the identity rather than the rate: faults equal distinct pages touched. That is falsifiable, and it is what makes the number harmless - the count is bounded by your working set, not by your access count. 31.58% of accesses faulted here and nothing was wrong; the same figure with major faults would be a machine on its knees.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE CONTINUE · David Lee Wise (ROOT0), with AVAN