Locality — where it appears
Named by 36 essays across 7 fields — each of them below, with the objects they name alongside it.
One access, eight kilobytes
An operation count charges one for an array access. A machine charges for a block. The same 65,536 accesses cost 1,024 transfers in one order and 65,536 in another, with nothing about the algorithm's work changed — a factor of 64, which is exactly the number of elements in a block, and which no counter here could see until now.
The count is not the time
An operation count is exact, machine-independent, and not a running time. The gap between them is mostly memory, and it is large enough to reorder the rankings. A second count is kept alongside it — modelled cache misses from the same runs — and the two are required to disagree, because if they agreed the second one would carry no information.
The cliff where the data stops fitting
Below the cache's capacity, almost every access hits. A factor of eight above it, almost every access misses. The transition is not gradual and it is not a property of any algorithm — it is a property of how much data there is, and an algorithm's complexity class says nothing about which side of it a program is working on.
Where an algorithm looks
Plotted as index against time, every array access an algorithm makes becomes a picture that no count contains. Merge sort's is a set of sweeps. Heapsort's is a spray. Quicksort's is a narrowing triangle. These shapes decide how fast the algorithms run and they are entirely absent from the analysis that says all three are Θ(n log n).
A list and a block of memory
The same traversal, over the same graph, examining the same edges in the same order, laid out two ways. Twelve thousand two hundred and eighty-eight edge slots either way; 11,812 modelled cache misses against 3,258. This is the site's largest gap between two counts of one run, and it exists because one of the layouts is a pointer chase and the other is a sweep.
Two searches, one comparison count
Three arrangements of the same binary search tree over the same million keys, walking the same path, making the same twenty comparisons. One costs 15 block transfers, one costs 13, and one costs 3. Nothing about the algorithm differs between them — only where the nodes were put — and no counter this site had before this phase could tell them apart.
A column computed in machine words
Adjacent cells of a distance table differ by at most one, so a whole column is two bits per cell — and thirty-two of them fit in one register. Fifteen word operations per character replace three cell evaluations per cell, and below a pattern of fifteen characters the trade is a loss.
Permuting is the harder problem here
Rearranging 65,536 elements into a stated order costs 63,601 transfers one at a time and 4,096 by sorting them into place. In the model every other field on this site uses, the first is the cheap method and beats the second by a factor of eight. The two models disagree about which problem is easy, and they disagree by about the same factor in opposite directions.
The index that is not worth reading
An index turns a query over 65,536 rows from 1,024 transfers into four. At a thousand matching rows it costs 654 and still wins; at sixteen thousand it costs 1,027 and has lost. Where it turns is decided by the block size — a number the query does not contain, the schema does not mention, and nobody writing either has seen.
A triangle stored in a square
An interval table has a cell for every range of keys and nothing below its diagonal, and it can be stored as a square array, as packed rows, or as packed diagonals — the last matching the order it is filled in. On sixty-four keys, with every read replayed through a small cache, the square misses 39.7% of its reads, packed rows 38.8%, and packed diagonals 78.4%. Storing a table in the order it is written is storing it in the order it is not read.
The permutation that moves almost nothing
Two ways to scramble sixteen thousand elements. Shuffling them inside windows of five hundred and twelve puts two million pairs out of order and costs 3,095 block transfers to carry out. Swapping a thousand pairs across the whole array puts seven million out of order and costs 1,189. Inversions are the textbook measure of disorder, and on a disk they rank these two backwards.
The split scan cut into blocks
Every way of filling an interval table one cell at a time stops at about one cache miss per split point considered once the table outgrows the cache — 1.01 at 128 keys, whether the cells go by length, by rows, or in a recursive tiling. Cut each cell's scan into blocks instead, and apply a block of split points to a block of cells whose inputs are all in hand, recursively at every scale, and the same 357,760 split points cost 0.094 misses each. The fill is told nothing about the cache, blocks of one and of four do equally well, and it needs no extra memory, where storing the table twice gets to 0.151 by doubling it.
The bound a block can and cannot have
Knuth's condition turns an interval table's cubic fill into a quadratic one by bounding each cell's best split between its two neighbours'. A blocked fill cannot use it a cell at a time, and the two cells that bound a block lie outside the block — one to its left, one below it. The schedule has finished both for ten per cent of the blocks, the bound then removes eleven per cent of the splits, and it removes half a per cent of the cache misses, because the splits it skips are the ones already in the cache.
Positions confined to one line
A Bloom filter lookup reads 5.55 cache lines because its six positions are scattered across the whole filter. Confining them to a 512-bit block makes it exactly one, and costs 7% more false positives at eight bits a key. At sixteen bits a key the same block costs 91%, and the two-value trick that is free across a whole filter costs another 135% inside one — because a block is a small filter, and small filters are where the penalty lives.
The record that forgets on purpose
A table of every pair selection sort might ask is 131,072 bits at 256 elements and catches every repeated question. A direct-mapped cache of 1,024 pairs is 16,384 bits and catches 91% of them, at a break-even of 3.53 word operations a comparison against the table's 3.12. What decides the cache is not how long the sort waits before asking again — 98% of its repeats come within one pass — but how many other questions it asks meanwhile, and a model with nothing in it but collisions predicts every measured share to within 2.6%.
A lookup that stops caring how wide an entry is
Buckets of eight entries aligned to a cache line read 1.20 lines a lookup when an entry is eight bytes and 19.25 when it is 128, because the bound was arithmetic about alignment and the arithmetic stops holding. Keeping one byte of each key's hash in a separate array and the entries in a parallel one reads 2.21 lines at every width from four bytes to sixty-four — and for a key the table does not hold, 2.05 against 31.98.
Two blocks and the chances they add
Send each key to the emptier of two 512-bit blocks and the busiest block of a filter at sixteen bits a key holds 35 keys instead of 51. The false-positive rate gets worse: 0.0997% against 0.0854%. At eight bits a key it nearly doubles. A lookup cannot tell which block a key went to, so it has to ask both, and asking twice is two chances to be wrong. At every density a filter is commonly built at, the repair that tames a hash table's worst bucket buys a filter nothing that a block twice as wide does not.
A block the lookup can work out
Sending each key to the emptier of two blocks flattens the load and doubles the chance of a false positive, because a lookup cannot know which block was chosen. Give each block a threshold on a hash of the key instead and the lookup can work it out: one block read, one chance, and 46% of the load balance kept. At sixteen bits a key that is 0.0618% against one block's 0.0854% and two blocks' 0.0997%. The threshold costs 1.5% of the budget and two bits of it are nearly as good as sixteen.
The record measured where it would run
A table of every pair selection sort has compared saves half its comparisons at 1,024 keys, and priced in word operations it was the better record at every size. Replayed through a 32 KB cache it misses 477,824 times to save 258,290 comparisons — nearly two misses for every comparison it removes — while a cache of 1,024 pairs misses 192. The table's size is not the problem. Its layout is: it looks up and writes each question at [key][minimum], a new row every time. Indexed by the ordered pair instead, the same bits miss 3,079 times, and that repair works only for a sort whose smaller operand stays put.
The digit a pass chooses for itself
A radix sort's digit width trades passes against the streams each pass scatters into, and the best width depends on the keys and the cache together: 8 bits for uniform keys through 32 KB, 11 through 256 KB, 16 for keys with few distinct values. A digit widened because the previous pass's counts were concentrated cannot find it, since in a least-significant-first sort one digit's counts say nothing about the next. One read that counts every byte at once, and a sample to size each field, can. Charged for both, the resulting plan is never worse than 1.23 times the best fixed width on any keys or cache measured. The best single fixed width reaches 1.50, and 11 bits reaches 3.61.
A penalty set by the block
A blocked Bloom filter pays a penalty for its uneven block loads, and a threshold that lets a lookup work out its block repays about a third of it — measured on 2,048 keys, with a prediction that larger filters would widen the gain. From 2,048 keys to 131,072 nothing moves: one block holds 0.086% to 0.090% against a computed 0.086%, and the threshold design 0.060% to 0.065%. The busiest block grows and the rate does not follow it. What sets both the penalty and the repair is the keys a block holds. A threshold turns out to be worth one doubling of the block's width, and its advantage shrinks from 40% at sixteen keys a block to between 5% and 13% at 128.
A record the size of a pass's minima
Selection sort's table of every pair it has compared is 256 KB at 1,024 keys, and indexed by the ordered pair it lives almost entirely in the rows of the current minima. Keep only those rows — a ring of r rows, the least recently used given up — and 12 rows, 3 KB, catch 99% of the comparisons the whole table skips, with no misses beyond the first touch of each row. The ring beats the cache of 1,024 pairs on both axes at once. The rows it needs grow with the logarithm of the input, a few more than the minima one pass passes through, and which row it gives up matters as much as how many it has.
A record of stretches, not pairs
Selection sort asks many of its comparisons again, and a record of past answers can skip them. A table of every pair skips them all in 256 KB at 1,024 keys, and a ring of the recent minima's rows in 3 KB. The questions a pass repeats have more structure than rows record: each minimum of the previous pass was compared with one contiguous stretch of positions. Kept as a list of stretches — each minimum's position, where its stretch began, and where the key that beat it stood — the previous pass's chain answers 99.3% of what the table would, in 312 bytes. The failure predicted for it, a key the swap moves into a stretch, costs 0.04%. What the chain cannot do is stay small on every input: on reversed keys it needs a link for every key.
Two swaps a pass, and the stretches hold
Bidirectional selection sort finds the smallest and the largest key in one scan and places both, so a record of its past answers has to keep two chains, and each pass makes two swaps that land where the other chain reaches. The prediction was that the second swap would break stretches at a rate of whole per cent. It does not. Each chain answers its own side's repeats as the one-sided chain does — 99.1% and 99.2% at 1,024 keys — and a swap spoils about a tenth of a question whichever sort makes it, so the broken share falls as 1/n for both. What the second chain changes is the input that makes a record long: sorted keys, which the one-sided sort never repeats a question on, cost the two-ended record a link a key.
The order with the best depth
An edit-distance table can be filled row by row, column by column, or one anti-diagonal at a time, and the anti-diagonal order is the one that needs the fewest rounds — 513 against 65,793 on two strings of 256 characters, because every cell on an anti-diagonal is independent of the others. Stored the usual way, row by row, it also misses the cache on 31.1% of its reads, where row order misses 6.3%. The order that is best for parallel work is worst for the memory it runs on.
Two probes are two misses
Cuckoo hashing's lookup reads at most two slots, and at a load of 0.45 it reads 1.27 on average where linear probing reads 1.39. Replayed through a cache, it misses 1.18 times a lookup where linear probing misses 0.98. The table that wins the count the analysis uses loses the count the machine charges, because two slots in unrelated places are two cache lines, and a run of adjacent slots is usually one.
The bucket that fits a line
Make each of a cuckoo table's two candidates a bucket of eight slots laid out on one cache line, and no lookup ever touches more than two lines, the table builds past a load of 0.95, and at that load it misses 1.21 times a lookup where linear probing misses 1.79. The prediction that it would lose to linear probing at low loads was wrong — it misses less at every load measured, 0.94 against 0.96 at 0.3 — because a key it holds almost never lives in its second bucket. The guarantee belongs to the alignment, not the bucket; eight slots on lines of four put a lookup on four lines.
The table stored the way it is filled
Store an edit-distance table by anti-diagonals instead of by rows, and the anti-diagonal fill keeps its 513 rounds while its cache misses fall from 31.1% of reads to 8.7%. It does not fall to row order's 6.3%, and the gap is not noise — on caches of four and eight lines the two rates are 9.4% and 6.3%, exactly three to two, because a cell reads from two earlier diagonals and only one earlier row. The same layout turns row order into the order that strides, at 28.3%. How a table is stored and the order it is filled in are one decision, and its price is the number of earlier fronts the recurrence reads.
Eight cells at once
The anti-diagonal fill order exists because its cells do not depend on one another, and every table filled here has been walked one cell at a time anyway. Computed eight at a time, a step touches 5.71 cache lines on the layout that stores the table by diagonals and 10.87 on the one that stores it by rows — and per cell the first keeps falling to 0.42 while the second stops at 1.27. The prediction that a diagonal step would touch three or four lines was wrong, and line-aligning each diagonal only takes it to 4.94.
A product paid for in blocks
The published floor under sorting puts a bound on time times space, and a read-only family that scans its input over and over meets it, while every in-place sort sits far above it. Counted in block transfers rather than reads, the floor itself moves to n²/B: the family meets it at every block size, and so does every sort once memory holds the whole array. With a memory of 512 cells, quicksort sits a fixed eight times above the family at every block size, because partitioning is itself a scan. Only heapsort's distance grows with the block, from ten times at a line of one to 285 times at 64. The prediction that the family's lead would widen towards B holds against heapsort and nothing else.
A tag that answers more than yes
A byte of hash in front of each entry has one bit of output per slot: this tag matches or it does not. Cut it from an ordered field instead and a bucket's tags can be kept sorted, so a lookup stops at the first tag above its own — 8.97 comparisons for an absent key instead of 14.43, for 3.55 tag-byte shifts an insertion. The obvious ordered field is the trap. Take the tag from the bucket index's own bits and every key in a bucket shares them: 5.74 entry reads on an absent key instead of 0.054, a hundred and seven times more.
Entries found by the rank of their tag
A bucket whose tags are kept sorted is an index: the position a tag is found at is its rank, and a rank can address an entry. Drop the fixed slots and pack each bucket's entries in tag order, and a cuckoo table of 32-byte records holds 36.1 bytes a key at half load where slots hold 66.1, and 34.3 against 36.7 at 0.9. Every insertion pays for it by moving the entries after its rank: 99 bytes against 36 at half load, 586 against 105 at 0.9. The prediction that packing would lose above 70% load was about the wrong quantity.
The slack an insertion can reach
A cuckoo table whose entries are packed by tag rank holds half the bytes of a slotted one at half load, and pays by shifting entries on every insertion. It was proposed to keep a few spare entries at the end of each group's region. Charged the copies a full region costs, that slack helps only with the copies: at groups of 16 buckets an insertion still moves over a thousand bytes, against 36 for slots. The same slack divided among the buckets, with an insertion borrowing the nearest free entry, moves 347 — and the group size, which multiplied every earlier cost, stops mattering.
The memory a product does not charge
Time times space charges a sort for the cells it writes and not for the fast memory it needs to write them cheaply. Counted three ways, on random keys in lines of eight, a read-only family stays within a factor of eight of the floor with 32 to 256 cells of fast memory at every length, while heapsort needs all of n and quicksort an eighth of it. Quicksort's need falls far faster than predicted as the factor loosens. And the product the floor leaves out, transfers times fast memory, separates the two kinds of sort completely: no read-only sort can push it far below n²/B, and quicksort reaches a twentieth of it.
The lines a sort writes back
A replay that charges a transfer for every line brought into fast memory and none for a line sent back flatters every sort that writes in place. Charged for its write-backs, quicksort's transfers rise by 1.9 at every memory short of n, as predicted, and heapsort's by 1.8, not the predicted three, because no line can go back more often than it came in. The cost that changes the account is the last one: at n itself every in-place sort pays two reads of its array, so none of them can come within a factor of two of the floor, while the read-only family reaches 1.07.
A buffer that does not shift its tail
The read-only family of sorts scans its input once for each batch of the smallest values not yet output, keeping them in a buffer; charged for the dirty lines it sends back to slow memory, its least transfers × fast memory rose from 1.29 of n²/B to 1.87, because a sorted buffer takes each value by shifting every larger entry one place. Kept as a max-heap instead, the buffer was predicted to write back about a third of the lines it reads below its own size and to bring the product back towards 1.29. At a buffer of 64 cells it writes back 33% at its best memory, against 45%, and its least product falls to 1.71: right about the share, wrong about the product, since a third written back is still a third more transfers. At 256 cells the heap does more than predicted — 1.47 million accesses against 6.94 million, a third of the misses — and its least product, 1.62, is under any buffer's on the earlier charge. At 16 cells it does nothing.
Named alongside it
The objects these essays reach for when they reach for this one.
CacheWorking setCost modelMemory layoutDesign parameterHonest limitMeasured countSpace time tradeBlock transferAccess patternComparison countMiss rate