Theme

The thread: Space is the other axis

Comparisons and cache misses are both proxies for duration. How much an algorithm keeps while it runs is a separate resource, it is measurable by the same discipline, and it is the one that decides whether a program runs at all.
suffix array + text311,29619.00 b/chcounter array per symbol5,407,710330.06 b/chFM-index, plain100,9476.16 b/chFM-index, compressed36,8042.25 b/chthe packed textone bar shaded darker needs the text · English-likesigma 21, sample 64 The index that replaces the text

An index larger than what it indexes

A suffix array over 16,384 characters is 229,376 bits, and it cannot answer a single question without the 81,920 bits of text beside it. Nearly four times the text, to search the text. Every index on this site had been weighed at zero until somebody put one on a scale.

peak slots held at once, logarithmic18645124096Insertion sort11Selection sort11Bubble sort11Heapsort11Shellsort11Quicksort, median of three22log nQuicksort, random pivot28log nQuicksort, first-element29log nMerge sort with a cutoff4,106nMerge sort4,110nn = 4,096, random inputone slot = one array element or one stack frame The other axis

Measuring what an algorithm keeps

Four counters measure what an algorithm does and none of them measures what it holds. An in-place sort and an out-of-place one with identical comparison counts are different algorithms, and until this phase the site had no way to say so. Two primitives close the gap, and the second of them counts something no array counter can ever see.

1,00010,0000.1bits of state heldrelative errorHyperLogLog (-0.49)LogLog (-0.44)bottom-k (-0.38)50,000 distinct keys · 14 runs per point · truth counted exactlybest: 2.00% at 10,240 bits One pass, and no room

The answer that is allowed to be wrong

Every algorithm on this site so far was checked for correctness before it was measured. A summary of a stream cannot be — the data goes past once and does not fit — so the error becomes a resource, bought with bits, at an exchange rate that is a measurement.

one summary, k = 32769one summary, k = 25608 summaries of 32, merged536worst error over the top keys, in arrivalsconcentration 1.00 — the mean share of a heavy key held by one shard8 shards · hashed · balancedmerge 536 against matched 0 One pass, and no room

The state a merge is standing in for

A merge of eight summaries of thirty-two counters is wrong by 536 where one summary of thirty-two is wrong by 769, which reads as merging helping. One summary of two hundred and fifty-six counters — exactly what the eight were holding between them — is wrong by nothing at all.

10010³1010010³nstack framesa stack of 512 framespivot: first-elementpivot: median of threepivot: random pivotMerge sortalready sorted input, n from 64 to 4,096one frame charged as one slot The other axis

The stack nobody counts

Merge sort makes 8,192 calls to sort 4,096 elements and holds fourteen of them at once. Depth-first search on a grid holds twelve vertices, or sixty-six, or a hundred and forty-four, depending on which of three equally standard implementations is running. The stack is a resource, it is the one that fails hard rather than slowly, and nothing that watches the data can see it.

10⁵10⁶10⁷11010010³10⁴comparisonspeak auxiliary slotsInsertion sortSelection sortBubble sortMerge sortHeapsortQuicksortQuicksortQuicksortShellsortMerge sort with a cutoffn = 8,192, random input5 on the frontier, 5 dominated The other axis

The frontier between time and space

The question of which sorting algorithm to use has an honest answer, and it is a shape rather than a name. Comparisons on one axis, peak auxiliary space on the other, and five of the ten algorithms here are on the Pareto frontier while five are dominated — beaten on both counts at once, so that no weighting of the two costs makes them the right choice. Heapsort is one of the five that lose.

comparisonsswapsInsertion sort63,071 / 0Selection sort130,816 / 504Bubble sort129,688 / 62,563Merge sort3,964 / 0Heapsort7,653 / 4,170Quicksort5,049 / 2,380n = 512, random inputcounted in the same run Counting

The count somebody chose

Six quantities can now be measured for every sort. Ranking the ten algorithms by each of them and comparing the orders, comparisons and peak space disagree about 91% of all pairs, and memory traffic and modelled misses disagree about 7%. There is no ranking of sorting algorithms; there are six, and choosing between them is a statement about the data rather than about the algorithms.

pale: total slots ever allocated · dark: peak held at onceone buffer, allocated once16,38416,400peak 100% of na buffer per merge229,37616,386peak 100% of nn = 16,384, random input208,687 comparisons each — identical in time The other axis

The space the model does not see

A slot is not a byte, a frame is not a slot, sixteen thousand allocations are not one allocation of the same size, and none of these numbers includes the input. The space counters are the newest instrument here and the honest account of what they miss is longer than the account of what they measure — including one bound this phase set out to demonstrate and could not.

executionintention87777776569888888765one unit = one subproblem given a value100 cells computed, 20 held at once The other axis

The table nobody has to keep

A million-cell table, computed cell for cell in the same order, holding two thousand cells at its peak instead of a million. The saving is exactly (n+1)/2, it costs nothing on any operation counter, and what it buys is paid for with the one thing the table was for.

bits per elementε = 0.13.3 → 4.8 (+1.5)ε = 0.035.1 → 7.3 (+2.2)ε = 0.016.6 → 9.6 (+2.9)ε = 0.0038.4 → 12.1 (+3.7)ε = 0.00110.0 → 14.4 (+4.4)ε = 1e-413.3 → 19.2 (+5.9)floor log₂(1/ε) filled; Bloom's log₂(1/ε)/ln 2 outlined44.3% above the floor at every rate The floors

A floor on the bits

Answering membership for n keys with a false-positive rate of 1% and no false negatives requires at least 6.64 bits per key, whatever the structure. A Bloom filter uses 9.59. The gap is 44.27% at that rate and at every other rate, and it is the first bound on this site that a real structure comes close to.

executionintention0123456789112345667822234567773333455678434345667854444567776555555678766666656787777776569888888765one unit = one subproblem given a value110 cells for one divide step, 30 held The other axis

The alignment that fits in one line

Compute the table twice and hold three rows of it. The factor of two is a geometric series and is predicted exactly; measured, it comes down from 2.269 to 2.052 as the strings grow, and the peak is 3(m+1) cells on the nose.

1101001,00010010³position in the textLF stepsspan + sample = 64row sampling onlywith the secondsampling8,192 characters · sample one in 32 · span 32second sampling 3,598 bits The index that replaces the text

The sampling that goes the other way

An FM-index hands the text back, and the way it does it is to walk from the last character to the first. So thirty-two characters from the end cost thirty-three steps and thirty-two characters from the beginning cost eight thousand one hundred and ninety-two. The repair is a second array the same size as the first, indexed the other way round.

W = 256even256–256Poisson225–285bursty256–256drifting79–1,736items held inside the time windowD = 2.56 s · 1 ms clockW would be 256, always What is taught wrongly

The window that is not full

A structure sized for a window of 256 items meets a stream that hands it 1,736 at the worst instant and 79 at the best. Occupancy was a constant in the model the sizing came from, and every per-item bound in that model quietly assumed it.

1,00010³window length W, in arrivalsbits of state heldW bits, exactε = 0.5ε = 0.2ε = 0.1ε = 0.05ε = 0.02sliding-window model · 40,000 arrivals · state from the shape of the structure5,670 bits at W = 8,000 The other axis

What a window costs in bits

The approximate structure grows like the square of a logarithm and the exact one grows like the window, so the approximation wins eventually. Eventually is a window of 6,000 at a 2% tolerance — and below it the summary is larger than the thing it is summarising.

decayedwindowed060120secondscount attributed to key 0half-life 4 s · window 5.77 s · burstythe same key, two questions One pass, and no room

The counter with no window in it

A counter that fades by half every H settles, on a steady stream, at exactly the count of a window of 1.44H. That correspondence holds in the mean, on a steady stream, and nowhere else — and it is the reason a decayed counter is not an estimate of a windowed count for any window.

3,600 — the window opens4,000 — now32161616884the window edge215 in buckets−16 for the oldest= 200true 2000.3% outsliding-window model · ε = 0.2, k = 5 · 1,764 merges406 bits against 400 One pass, and no room

The summary that has to forget

Every structure in this field so far accumulates. Ask instead for the count over only the last thousand arrivals and no counter will do, because a counter has no record of which of its increments are old — and the repair is a row of buckets whose whole error is the oldest one.

100,000110bits held by the whole indexLF steps per located occurrence1 in 11 in 21 in 41 in 81 in 161 in 321 in 641 in 128one point per sampling rate · 66 occurrences located each timeEnglish-like The other axis

The text that does not have to be kept

The index reproduces its text character for character, in 1,024 mapping steps and zero reads of anything. That is the whole justification for weighing it against the text rather than beside it — and the price is a dial that moves the structure by 3.3 times and the cost of locating one match by 72.

windowed HyperLogLog4,088 bitsone stamp per live key8,775 bitsblocks of Misra-Gries18,549 bitsthe last W keys, kept131,072 bitskeysstampspayloadindexthe parts, in the order they stackwindow 4,096 · the popular keys drift131,072 bits at most One pass, and no room

The bits that say when

A windowed cardinality estimator holds 4,592 bits and 2,392 of them are clocks. Every summary in this collection has reported its size from the shape of its own structure, and not one of those numbers has ever been asked what the bits were for — so the resource that half of these structures spend most of their state on has been invisible while being counted.

fades a lazy table computed382,976fades an eager table would compute3,830,256fade multiplications (log scale)11,968 evictions, each comparing all 32 counters at one instant32 counters · half-life 4 s · bursty10× fewer What the libraries do

The fading nobody computes

Twelve thousand arrivals into thirty-two decayed counters cost 382,976 fade multiplications. An implementation that aged every counter on every tick would have cost 3,830,256, and the ratio is exactly the mean gap between arrivals — not a coincidence, and the reason the family is deployable.

11010010⁴10⁵queries answeredcharacter comparisons, cumulative5 queriesScan each timeIndex, then queryone unit = one character comparison · 4 doubling roundsbreak-even at 5 queries Structures

The index that is the text

A suffix array sorts all 4,097 suffixes of a text — 8.4 million characters of string, in total — and examines exactly zero characters doing it. It then answers a search in 91 characters where a scan costs 1,472, and the whole thing pays for itself at six queries. Both halves of that are worth the same amount of attention, and the first is the one that is usually skipped.

10³10⁴10⁵10⁶passes over the datapeak bits of stateone pass, exact: 1,048,576 bitsradix, uniformsample, uniformradix, paretosample, paretoevery point exact · 32,768 values320 bits at best The other axis

What a second pass buys

Exact selection of a median from thirty-two thousand values needs the whole stream in one pass — a million bits — and eleven thousand in two. By nine passes it is three hundred and twenty. The state falls as n to the power one over p, which is a law with an exponent worth fitting, and on skewed data the deterministic rule misses it by three orders of magnitude.

0%25%50%75%100%1.5×shareheadroom over the mean occupancyoverflowingstanding idledrifting · 4.0 s window · 100 Hz47% overflow at the mean What the libraries do

Sized for a rate that does not hold still

A four-second window on a stream at a hundred arrivals a second holds four hundred items on average and between 105 and 2,169 when the rate moves. An allocation set at that average overflows at 47 per cent of instants while 35 per cent of it stands empty, which is the same decision failing in both directions at once.

sampled positions3,855grows with nsample marks2,056grows with nrun starts2,016grows with rrun lengths per symbol1,885grows with rrun heads1,563grows with rC table525grows with r11,900 bits in total16,384 characters · r = 224 · sample one in 6450% is the sampling The other axis

What is still proportional to n

An index whose size is a function of the run count still has an n in it, and at thirty-two copies of a text the n is half of it. Everything that stores the transform grew by 45 per cent; the two arrays that answer "where" grew by eighteen times, and neither of them has anything to do with repetition.

1,00010,00010⁴bits12481632characters in the collection · copies aboveFM-index, compressedr-indexphrase indexEnglish-like · divergence 0z 156 · r 233 The index that replaces the text

An index with z in its size

Over thirty-two copies of one text, an index built on the parse is 8,892 bits, the r-index is 17,047 and the entropy-bounded index is 34,615. Over eight thousand characters of four-symbol text the same three are 7,844, 25,177 and 20,413, and the smallest of the three has changed places twice.

1,000248163264high-biased, α = 0.56low-biased, α = 0.56none-biased, α = 0.74tuples keptshards mergedlog-normal, σ = 1.2 — a latency distribution · 20,000 valuesα 0.56 / 0.56 / 0.74 · worst residual 6.5% The other axis

The cheap tail and the expensive merge

A summary whose tolerance tightens towards the tail keeps seven times the tuples of a plain one on a single pass, and after merging sixty-four shards it keeps three and a half times as many. The error function that buys a useful tail promise is also the one that pays most for never having the values in one place.

ε = 0.0515 → 29 (1.93×)ε = 0.0238 → 73 (1.92×)ε = 0.0177 → 136 (1.77×)ε = 0.005152 → 270 (1.78×)ε = 0.002397 → 674 (1.70×)reportedoccupied at the peak20,000 arrivals · lognormalpeak = resident + period, to 14% The other axis

The tuples a summary does not report

A Greenwald–Khanna summary at ε = 0.01 answers `tuples` with seventy-seven. Watched through the run it holds a hundred and thirty-six. The gap is the compression period, it is 1.70 to 1.93 times across every tolerance measured, and it is the number a deployment has to allocate.

0.020.050.10.20.5121235810filter bits a keywasted reads per absent lookupno filters: all 3 levelssame rate every levelrates sized to levelsn = 65,536, size ratio 4, first run 1,024 keysthe same memory at every point When it does not fit

The filter each run carries

A log-structured store turns every lookup for a missing key into a read of every level, and a Bloom filter on each run buys those reads back with memory. Spread five bits a key evenly across three levels and a missing key still wastes 0.279 reads. Give the small levels more bits and the large one fewer — the same memory — and it wastes 0.201. At four levels the gap is 0.382 against 0.209, because sized filters stop the waste growing with the number of levels.

0.01%0.1%1%10%100%1,0002,0003,0004,0005,0006,0007,0008,000keys insertedabsent keys answered yesone filterm = 19,171, k = 7; dotted: the design sizedashed: the design rate When the algorithm flips a coin

A filter past its design size

A Bloom filter sized for two thousand keys at one per cent answers yes to 15.6% of absent keys at four thousand and 68.1% at eight thousand. Nothing fails and nothing warns. A stack of filters that adds a tighter layer whenever the top one fills holds 2.0% at eight thousand, under a bound it can state in advance — in 2.9 times the bits of one filter sized for eight thousand from the start.

1,00010,000248163264ε = 0.02, α = 0.58ε = 0.01, α = 0.56ε = 0.005, α = 0.54tuples keptshards mergedlog-normal, σ = 1.2 — a latency distribution · 20,000 valuesα 0.58 / 0.56 / 0.54 · worst residual 2.0% One pass, and no room

The tuples a merge does not give back

A merge of thirty-two quantile summaries keeps seven times the tuples of one summary over the same values, and sixty-four keeps ten and a half. Fitted across the sweep the count goes as the shard number to the power 0.56, which answers what it converges to — it does not.

10,000110sampling, bitssteps per occurrence1 in 21 in 41 in 81 in 161 in 321 in 641 in 128r-indexEnglish-like · 16 copies · r = 233pattern " time " The other axis

Every occurrence at the same price

A regular sampling has one dial and it moves two costs together — 1,495 bits at 63 LF steps an occurrence, 131,088 bits at none. A sampling at the run boundaries sits at 10,244 bits and one predecessor query, which is a point the curve reaches only at 69,649.

All threads