arbo/addbatch.go
addbatch.goRaw
1package arbo
2
3import (
4 "bytes"
5 "fmt"
6 "math"
7 "runtime"
8 "sort"
9 "sync"
10
11 "github.com/iden3/go-merkletree/db"
12)
13
14/*
15
16AddBatch design
17===============
18
19
20CASE A: Empty Tree --> if tree is empty (root==0)
21=================================================
22- Build the full tree from bottom to top (from all the leaf to the root)
23
24
25CASE B: ALMOST CASE A, Almost empty Tree --> if Tree has numLeafs < minLeafsThreshold
26==============================================================================
27- Get the Leafs (key & value) (iterate the tree from the current root getting
28the leafs)
29- Create a new empty Tree
30- Do CASE A for the new Tree, giving the already existing key&values (leafs)
31from the original Tree + the new key&values to be added from the AddBatch call
32
33
34 R R
35 / \ / \
36 A * / \
37 / \ / \
38 B C * *
39 / | / \
40 / | / \
41 / | / \
42 L: A B G D
43 / \
44 / \
45 / \
46 C *
47 / \
48 / \
49 / \
50 ... ... (nLeafs < minLeafsThreshold)
51
52
53CASE C: ALMOST CASE B --> if Tree has few Leafs (but numLeafs>=minLeafsThreshold)
54==============================================================================
55- Use A, B, G, F as Roots of subtrees
56- Do CASE B for each subtree
57- Then go from L to the Root
58
59 R
60 / \
61 / \
62 / \
63 * *
64 / | / \
65 / | / \
66 / | / \
67L: A B G D
68 / \
69 / \
70 / \
71 C *
72 / \
73 / \
74 / \
75 ... ... (nLeafs >= minLeafsThreshold)
76
77
78
79CASE D: Already populated Tree
80==============================
81- Use A, B, C, D as subtree
82- Sort the Keys in Buckets that share the initial part of the path
83- For each subtree add there the new leafs
84
85 R
86 / \
87 / \
88 / \
89 * *
90 / | / \
91 / | / \
92 / | / \
93L: A B C D
94 /\ /\ / \ / \
95 ... ... ... ... ... ...
96
97
98CASE E: Already populated Tree Unbalanced
99=========================================
100- Need to fill M1 and M2, and then will be able to use CASE D
101 - Search for M1 & M2 in the inputed Keys
102 - Add M1 & M2 to the Tree
103 - From here can use CASE D
104
105 R
106 / \
107 / \
108 / \
109 * *
110 | \
111 | \
112 | \
113L: M1 * M2 * (where M1 and M2 are empty)
114 / | /
115 / | /
116 / | /
117 A * *
118 / \ | \
119 / \ | \
120 / \ | \
121 B * * C
122 / \ |\
123 ... ... | \
124 | \
125 D E
126
127
128
129Algorithm decision
130==================
131- if nLeafs==0 (root==0): CASE A
132- if nLeafs<minLeafsThreshold: CASE B
133- if nLeafs>=minLeafsThreshold && (nLeafs/nBuckets) < minLeafsThreshold: CASE C
134- else: CASE D & CASE E
135
136
137- Multiple tree.Add calls: O(n log n)
138 - Used in: cases A, B, C
139- Tree from bottom to top: O(log n)
140 - Used in: cases D, E
141
142*/
143
144const (
145 minLeafsThreshold = 100 // nolint:gomnd // TMP WIP this will be autocalculated
146)
147
148// AddBatchOpt is the WIP implementation of the AddBatch method in a more
149// optimized approach.
150func (t *Tree) AddBatchOpt(keys, values [][]byte) ([]int, error) {
151 t.updateAccessTime()
152 t.Lock()
153 defer t.Unlock()
154
155 // TODO if len(keys) is not a power of 2, add padding of empty
156 // keys&values. Maybe when len(keyvalues) is not a power of 2, cut at
157 // the biggest power of 2 under the len(keys), add those 2**n key-values
158 // using the AddBatch approach, and then add the remaining key-values
159 // using tree.Add.
160
161 kvs, err := t.keysValuesToKvs(keys, values)
162 if err != nil {
163 return nil, err
164 }
165
166 t.tx, err = t.db.NewTx() // TODO add t.tx.Commit()
167 if err != nil {
168 return nil, err
169 }
170
171 nCPU := runtime.NumCPU()
172
173 // CASE A: if nLeafs==0 (root==0)
174 if bytes.Equal(t.root, t.emptyHash) {
175 // if len(kvs) is not a power of 2, cut at the bigger power
176 // of two under len(kvs), build the tree with that, and add
177 // later the excedents
178 kvsP2, kvsNonP2 := cutPowerOfTwo(kvs)
179 invalids, err := t.buildTreeBottomUp(nCPU, kvsP2)
180 if err != nil {
181 return nil, err
182 }
183 for i := 0; i < len(kvsNonP2); i++ {
184 err = t.add(0, kvsNonP2[i].k, kvsNonP2[i].v)
185 if err != nil {
186 invalids = append(invalids, kvsNonP2[i].pos)
187 }
188 }
189 return invalids, nil
190 }
191
192 // CASE B: if nLeafs<nBuckets
193 nLeafs, err := t.GetNLeafs()
194 if err != nil {
195 return nil, err
196 }
197 if nLeafs < minLeafsThreshold { // CASE B
198 invalids, excedents, err := t.caseB(0, kvs)
199 if err != nil {
200 return nil, err
201 }
202 // add the excedents
203 for i := 0; i < len(excedents); i++ {
204 err = t.add(0, excedents[i].k, excedents[i].v)
205 if err != nil {
206 invalids = append(invalids, excedents[i].pos)
207 }
208 }
209 return invalids, nil
210 }
211
212 // CASE C: if nLeafs>=minLeafsThreshold && (nLeafs/nBuckets) < minLeafsThreshold
213 // available parallelization, will need to be a power of 2 (2**n)
214 var excedents []kv
215 l := int(math.Log2(float64(nCPU)))
216 if nLeafs >= minLeafsThreshold && (nLeafs/nCPU) < minLeafsThreshold {
217 // TODO move to own function
218 // 1. go down until level L (L=log2(nBuckets))
219 keysAtL, err := t.getKeysAtLevel(l + 1)
220 if err != nil {
221 return nil, err
222 }
223
224 buckets := splitInBuckets(kvs, nCPU)
225
226 // 2. use keys at level L as roots of the subtrees under each one
227 var subRoots [][]byte
228 // TODO parallelize
229 for i := 0; i < len(keysAtL); i++ {
230 bucketTree := Tree{tx: t.tx, db: t.db, maxLevels: t.maxLevels,
231 hashFunction: t.hashFunction, root: keysAtL[i]}
232
233 // 3. and do CASE B for each
234 _, bucketExcedents, err := bucketTree.caseB(l, buckets[i])
235 if err != nil {
236 return nil, err
237 }
238 excedents = append(excedents, bucketExcedents...)
239 subRoots = append(subRoots, bucketTree.root)
240 }
241 // 4. go upFromKeys from the new roots of the subtrees
242 newRoot, err := t.upFromKeys(subRoots)
243 if err != nil {
244 return nil, err
245 }
246 t.root = newRoot
247
248 var invalids []int
249 for i := 0; i < len(excedents); i++ {
250 // Add until the level L
251 err = t.add(0, excedents[i].k, excedents[i].v)
252 if err != nil {
253 invalids = append(invalids, excedents[i].pos) // TODO WIP
254 }
255 }
256
257 return invalids, nil
258 }
259
260 // TODO store t.root into DB
261 // TODO update NLeafs from DB
262
263 return nil, fmt.Errorf("UNIMPLEMENTED")
264}
265
266func (t *Tree) caseB(l int, kvs []kv) ([]int, []kv, error) {
267 // get already existing keys
268 aKs, aVs, err := t.getLeafs(t.root)
269 if err != nil {
270 return nil, nil, err
271 }
272 aKvs, err := t.keysValuesToKvs(aKs, aVs)
273 if err != nil {
274 return nil, nil, err
275 }
276 // add already existing key-values to the inputted key-values
277 kvs = append(kvs, aKvs...)
278
279 // proceed with CASE A
280 sortKvs(kvs)
281
282 // cutPowerOfTwo, the excedent add it as normal Tree.Add
283 kvsP2, kvsNonP2 := cutPowerOfTwo(kvs)
284 invalids, err := t.buildTreeBottomUpSingleThread(kvsP2)
285 if err != nil {
286 return nil, nil, err
287 }
288 // return the excedents which will be added at the full tree at the end
289 return invalids, kvsNonP2, nil
290}
291
292func splitInBuckets(kvs []kv, nBuckets int) [][]kv {
293 buckets := make([][]kv, nBuckets)
294 // 1. classify the keyvalues into buckets
295 for i := 0; i < len(kvs); i++ {
296 pair := kvs[i]
297
298 bucketnum := keyToBucket(pair.k, nBuckets)
299 buckets[bucketnum] = append(buckets[bucketnum], pair)
300 }
301 return buckets
302}
303
304// TODO rename in a more 'real' name (calculate bucket from/for key)
305func keyToBucket(k []byte, nBuckets int) int {
306 nLevels := int(math.Log2(float64(nBuckets)))
307 b := make([]int, nBuckets)
308 for i := 0; i < nBuckets; i++ {
309 b[i] = i
310 }
311 r := b
312 mid := len(r) / 2 //nolint:gomnd
313 for i := 0; i < nLevels; i++ {
314 if int(k[i/8]&(1<<(i%8))) != 0 {
315 r = r[mid:]
316 mid = len(r) / 2 //nolint:gomnd
317 } else {
318 r = r[:mid]
319 mid = len(r) / 2 //nolint:gomnd
320 }
321 }
322 return r[0]
323}
324
325type kv struct {
326 pos int // original position in the array
327 keyPath []byte
328 k []byte
329 v []byte
330}
331
332// compareBytes compares byte slices where the bytes are compared from left to
333// right and each byte is compared by bit from right to left
334func compareBytes(a, b []byte) bool {
335 // WIP
336 for i := 0; i < len(a); i++ {
337 for j := 0; j < 8; j++ {
338 aBit := a[i] & (1 << j)
339 bBit := b[i] & (1 << j)
340 if aBit > bBit {
341 return false
342 } else if aBit < bBit {
343 return true
344 }
345 }
346 }
347 return false
348}
349
350// sortKvs sorts the kv by path
351func sortKvs(kvs []kv) {
352 sort.Slice(kvs, func(i, j int) bool {
353 return compareBytes(kvs[i].keyPath, kvs[j].keyPath)
354 })
355}
356
357func (t *Tree) keysValuesToKvs(ks, vs [][]byte) ([]kv, error) {
358 if len(ks) != len(vs) {
359 return nil, fmt.Errorf("len(keys)!=len(values) (%d!=%d)",
360 len(ks), len(vs))
361 }
362 kvs := make([]kv, len(ks))
363 for i := 0; i < len(ks); i++ {
364 keyPath := make([]byte, t.hashFunction.Len())
365 copy(keyPath[:], ks[i])
366 kvs[i].pos = i
367 kvs[i].keyPath = ks[i]
368 kvs[i].k = ks[i]
369 kvs[i].v = vs[i]
370 }
371
372 return kvs, nil
373}
374
375/*
376func (t *Tree) kvsToKeysValues(kvs []kv) ([][]byte, [][]byte) {
377 ks := make([][]byte, len(kvs))
378 vs := make([][]byte, len(kvs))
379 for i := 0; i < len(kvs); i++ {
380 ks[i] = kvs[i].k
381 vs[i] = kvs[i].v
382 }
383 return ks, vs
384}
385*/
386
387// buildTreeBottomUp splits the key-values into n Buckets (where n is the number
388// of CPUs), in parallel builds a subtree for each bucket, once all the subtrees
389// are built, uses the subtrees roots as keys for a new tree, which as result
390// will have the complete Tree build from bottom to up, where until the
391// log2(nCPU) level it has been computed in parallel.
392func (t *Tree) buildTreeBottomUp(nCPU int, kvs []kv) ([]int, error) {
393 buckets := splitInBuckets(kvs, nCPU)
394 subRoots := make([][]byte, nCPU)
395 invalidsInBucket := make([][]int, nCPU)
396 txs := make([]db.Tx, nCPU)
397
398 var wg sync.WaitGroup
399 wg.Add(nCPU)
400 for i := 0; i < nCPU; i++ {
401 go func(cpu int) {
402 sortKvs(buckets[cpu])
403
404 var err error
405 txs[cpu], err = t.db.NewTx()
406 if err != nil {
407 panic(err) // TODO
408 }
409 bucketTree := Tree{tx: txs[cpu], db: t.db, maxLevels: t.maxLevels,
410 hashFunction: t.hashFunction, root: t.emptyHash}
411
412 currInvalids, err := bucketTree.buildTreeBottomUpSingleThread(buckets[cpu])
413 if err != nil {
414 panic(err) // TODO
415 }
416 invalidsInBucket[cpu] = currInvalids
417 subRoots[cpu] = bucketTree.root
418 wg.Done()
419 }(i)
420 }
421 wg.Wait()
422
423 newRoot, err := t.upFromKeys(subRoots)
424 if err != nil {
425 return nil, err
426 }
427 t.root = newRoot
428
429 var invalids []int
430 for i := 0; i < len(invalidsInBucket); i++ {
431 invalids = append(invalids, invalidsInBucket[i]...)
432 }
433 return invalids, err
434}
435
436// buildTreeBottomUpSingleThread builds the tree with the given []kv from bottom
437// to the root. keys & values must be sorted by path, and the array ks must be
438// length multiple of 2
439func (t *Tree) buildTreeBottomUpSingleThread(kvs []kv) ([]int, error) {
440 // TODO check that log2(len(leafs)) < t.maxLevels, if not, maxLevels
441 // would be reached and should return error
442
443 var invalids []int
444 // build the leafs
445 leafKeys := make([][]byte, len(kvs))
446 for i := 0; i < len(kvs); i++ {
447 // TODO handle the case where Key&Value == 0
448 leafKey, leafValue, err := newLeafValue(t.hashFunction, kvs[i].k, kvs[i].v)
449 if err != nil {
450 // return nil, err
451 invalids = append(invalids, kvs[i].pos)
452 }
453 // store leafKey & leafValue to db
454 if err := t.tx.Put(leafKey, leafValue); err != nil {
455 // return nil, err
456 invalids = append(invalids, kvs[i].pos)
457 }
458 leafKeys[i] = leafKey
459 }
460 r, err := t.upFromKeys(leafKeys)
461 if err != nil {
462 return invalids, err
463 }
464 t.root = r
465 return invalids, nil
466}
467
468// keys & values must be sorted by path, and the array ks must be length
469// multiple of 2
470func (t *Tree) upFromKeys(ks [][]byte) ([]byte, error) {
471 if len(ks) == 1 {
472 return ks[0], nil
473 }
474
475 var rKs [][]byte
476 for i := 0; i < len(ks); i += 2 {
477 // TODO handle the case where Key&Value == 0
478 k, v, err := newIntermediate(t.hashFunction, ks[i], ks[i+1])
479 if err != nil {
480 return nil, err
481 }
482 // store k-v to db
483 if err = t.tx.Put(k, v); err != nil {
484 return nil, err
485 }
486 rKs = append(rKs, k)
487 }
488 return t.upFromKeys(rKs)
489}
490
491func (t *Tree) getLeafs(root []byte) ([][]byte, [][]byte, error) {
492 var ks, vs [][]byte
493 err := t.iter(root, func(k, v []byte) {
494 if v[0] != PrefixValueLeaf {
495 return
496 }
497 leafK, leafV := readLeafValue(v)
498 ks = append(ks, leafK)
499 vs = append(vs, leafV)
500 })
501 return ks, vs, err
502}
503
504func (t *Tree) getKeysAtLevel(l int) ([][]byte, error) {
505 var keys [][]byte
506 err := t.iterWithStop(t.root, 0, func(currLvl int, k, v []byte) bool {
507 if currLvl == l {
508 keys = append(keys, k)
509 }
510 if currLvl >= l {
511 return true // to stop the iter from going down
512 }
513 return false
514 })
515
516 return keys, err
517}
518
519// cutPowerOfTwo returns []kv of length that is a power of 2, and a second []kv
520// with the extra elements that don't fit in a power of 2 length
521func cutPowerOfTwo(kvs []kv) ([]kv, []kv) {
522 x := len(kvs)
523 if (x & (x - 1)) != 0 {
524 p2 := highestPowerOfTwo(x)
525 return kvs[:p2], kvs[p2:]
526 }
527 return kvs, nil
528}
529
530func highestPowerOfTwo(n int) int {
531 res := 0
532 for i := n; i >= 1; i-- {
533 if (i & (i - 1)) == 0 {
534 res = i
535 break
536 }
537 }
538 return res
539}
540
541// func computeSimpleAddCost(nLeafs int) int {
542// // nLvls 2^nLvls
543// nLvls := int(math.Log2(float64(nLeafs)))
544// return nLvls * int(math.Pow(2, float64(nLvls)))
545// }
546//
547// func computeBottomUpAddCost(nLeafs int) int {
548// // 2^nLvls * 2 - 1
549// nLvls := int(math.Log2(float64(nLeafs)))
550// return (int(math.Pow(2, float64(nLvls))) * 2) - 1
551// }