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 // TODO if nCPU is not a power of two, cut at the highest power of two
172 // under nCPU
173 nCPU := runtime.NumCPU()
174 l := int(math.Log2(float64(nCPU)))
175
176 // CASE A: if nLeafs==0 (root==0)
177 if bytes.Equal(t.root, t.emptyHash) {
178 // if len(kvs) is not a power of 2, cut at the bigger power
179 // of two under len(kvs), build the tree with that, and add
180 // later the excedents
181 kvsP2, kvsNonP2 := cutPowerOfTwo(kvs)
182 invalids, err := t.buildTreeBottomUp(nCPU, kvsP2)
183 if err != nil {
184 return nil, err
185 }
186 for i := 0; i < len(kvsNonP2); i++ {
187 err = t.add(0, kvsNonP2[i].k, kvsNonP2[i].v)
188 if err != nil {
189 invalids = append(invalids, kvsNonP2[i].pos)
190 }
191 }
192 return invalids, nil
193 }
194
195 // CASE B: if nLeafs<nBuckets
196 nLeafs, err := t.GetNLeafs()
197 if err != nil {
198 return nil, err
199 }
200 if nLeafs < minLeafsThreshold { // CASE B
201 invalids, excedents, err := t.caseB(0, kvs)
202 if err != nil {
203 return nil, err
204 }
205 // add the excedents
206 for i := 0; i < len(excedents); i++ {
207 err = t.add(0, excedents[i].k, excedents[i].v)
208 if err != nil {
209 invalids = append(invalids, excedents[i].pos)
210 }
211 }
212 return invalids, nil
213 }
214
215 // CASE C: if nLeafs>=minLeafsThreshold && (nLeafs/nBuckets) < minLeafsThreshold
216 // available parallelization, will need to be a power of 2 (2**n)
217 var excedents []kv
218 if nLeafs >= minLeafsThreshold && (nLeafs/nCPU) < minLeafsThreshold {
219 // TODO move to own function
220 // 1. go down until level L (L=log2(nBuckets))
221 keysAtL, err := t.getKeysAtLevel(l + 1)
222 if err != nil {
223 return nil, err
224 }
225
226 buckets := splitInBuckets(kvs, nCPU)
227
228 // 2. use keys at level L as roots of the subtrees under each one
229 var subRoots [][]byte
230 // TODO parallelize
231 for i := 0; i < len(keysAtL); i++ {
232 bucketTree := Tree{tx: t.tx, db: t.db, maxLevels: t.maxLevels,
233 hashFunction: t.hashFunction, root: keysAtL[i]}
234
235 // 3. and do CASE B for each
236 _, bucketExcedents, err := bucketTree.caseB(l, buckets[i])
237 if err != nil {
238 return nil, err
239 }
240 excedents = append(excedents, bucketExcedents...)
241 subRoots = append(subRoots, bucketTree.root)
242 }
243 // 4. go upFromKeys from the new roots of the subtrees
244 newRoot, err := t.upFromKeys(subRoots)
245 if err != nil {
246 return nil, err
247 }
248 t.root = newRoot
249
250 var invalids []int
251 for i := 0; i < len(excedents); i++ {
252 // Add until the level L
253 err = t.add(0, excedents[i].k, excedents[i].v)
254 if err != nil {
255 invalids = append(invalids, excedents[i].pos) // TODO WIP
256 }
257 }
258
259 return invalids, nil
260 }
261
262 // CASE D
263 if true { // TODO enter in CASE D if len(keysAtL)=nCPU, if not, CASE E
264 return t.caseD(nCPU, l, kvs)
265 }
266
267 // TODO store t.root into DB
268 // TODO update NLeafs from DB
269
270 return nil, fmt.Errorf("UNIMPLEMENTED")
271}
272
273func (t *Tree) caseB(l int, kvs []kv) ([]int, []kv, error) {
274 // get already existing keys
275 aKs, aVs, err := t.getLeafs(t.root)
276 if err != nil {
277 return nil, nil, err
278 }
279 aKvs, err := t.keysValuesToKvs(aKs, aVs)
280 if err != nil {
281 return nil, nil, err
282 }
283 // add already existing key-values to the inputted key-values
284 kvs = append(kvs, aKvs...)
285
286 // proceed with CASE A
287 sortKvs(kvs)
288
289 // cutPowerOfTwo, the excedent add it as normal Tree.Add
290 kvsP2, kvsNonP2 := cutPowerOfTwo(kvs)
291 invalids, err := t.buildTreeBottomUpSingleThread(kvsP2)
292 if err != nil {
293 return nil, nil, err
294 }
295 // return the excedents which will be added at the full tree at the end
296 return invalids, kvsNonP2, nil
297}
298
299func (t *Tree) caseD(nCPU, l int, kvs []kv) ([]int, error) {
300 fmt.Println("CASE D", nCPU)
301 keysAtL, err := t.getKeysAtLevel(l + 1)
302 if err != nil {
303 return nil, err
304 }
305 buckets := splitInBuckets(kvs, nCPU)
306
307 var subRoots [][]byte
308 var invalids []int
309 for i := 0; i < len(keysAtL); i++ {
310 bucketTree := Tree{tx: t.tx, db: t.db, maxLevels: t.maxLevels, // maxLevels-l
311 hashFunction: t.hashFunction, root: keysAtL[i]}
312
313 for j := 0; j < len(buckets[i]); j++ {
314 if err = bucketTree.add(l, buckets[i][j].k, buckets[i][j].v); err != nil {
315 fmt.Println("failed", buckets[i][j].k[:4])
316
317 panic(err)
318 // invalids = append(invalids, buckets[i][j].pos)
319 }
320 }
321 subRoots = append(subRoots, bucketTree.root)
322 }
323 newRoot, err := t.upFromKeys(subRoots)
324 if err != nil {
325 return nil, err
326 }
327 t.root = newRoot
328
329 return invalids, nil
330}
331
332func splitInBuckets(kvs []kv, nBuckets int) [][]kv {
333 buckets := make([][]kv, nBuckets)
334 // 1. classify the keyvalues into buckets
335 for i := 0; i < len(kvs); i++ {
336 pair := kvs[i]
337
338 // bucketnum := keyToBucket(pair.k, nBuckets)
339 bucketnum := keyToBucket(pair.keyPath, nBuckets)
340 buckets[bucketnum] = append(buckets[bucketnum], pair)
341 }
342 return buckets
343}
344
345// TODO rename in a more 'real' name (calculate bucket from/for key)
346func keyToBucket(k []byte, nBuckets int) int {
347 nLevels := int(math.Log2(float64(nBuckets)))
348 b := make([]int, nBuckets)
349 for i := 0; i < nBuckets; i++ {
350 b[i] = i
351 }
352 r := b
353 mid := len(r) / 2 //nolint:gomnd
354 for i := 0; i < nLevels; i++ {
355 if int(k[i/8]&(1<<(i%8))) != 0 {
356 r = r[mid:]
357 mid = len(r) / 2 //nolint:gomnd
358 } else {
359 r = r[:mid]
360 mid = len(r) / 2 //nolint:gomnd
361 }
362 }
363 return r[0]
364}
365
366type kv struct {
367 pos int // original position in the array
368 keyPath []byte
369 k []byte
370 v []byte
371}
372
373// compareBytes compares byte slices where the bytes are compared from left to
374// right and each byte is compared by bit from right to left
375func compareBytes(a, b []byte) bool {
376 // WIP
377 for i := 0; i < len(a); i++ {
378 for j := 0; j < 8; j++ {
379 aBit := a[i] & (1 << j)
380 bBit := b[i] & (1 << j)
381 if aBit > bBit {
382 return false
383 } else if aBit < bBit {
384 return true
385 }
386 }
387 }
388 return false
389}
390
391// sortKvs sorts the kv by path
392func sortKvs(kvs []kv) {
393 sort.Slice(kvs, func(i, j int) bool {
394 return compareBytes(kvs[i].keyPath, kvs[j].keyPath)
395 })
396}
397
398func (t *Tree) keysValuesToKvs(ks, vs [][]byte) ([]kv, error) {
399 if len(ks) != len(vs) {
400 return nil, fmt.Errorf("len(keys)!=len(values) (%d!=%d)",
401 len(ks), len(vs))
402 }
403 kvs := make([]kv, len(ks))
404 for i := 0; i < len(ks); i++ {
405 keyPath := make([]byte, t.hashFunction.Len())
406 copy(keyPath[:], ks[i])
407 kvs[i].pos = i
408 kvs[i].keyPath = ks[i]
409 kvs[i].k = ks[i]
410 kvs[i].v = vs[i]
411 }
412
413 return kvs, nil
414}
415
416/*
417func (t *Tree) kvsToKeysValues(kvs []kv) ([][]byte, [][]byte) {
418 ks := make([][]byte, len(kvs))
419 vs := make([][]byte, len(kvs))
420 for i := 0; i < len(kvs); i++ {
421 ks[i] = kvs[i].k
422 vs[i] = kvs[i].v
423 }
424 return ks, vs
425}
426*/
427
428// buildTreeBottomUp splits the key-values into n Buckets (where n is the number
429// of CPUs), in parallel builds a subtree for each bucket, once all the subtrees
430// are built, uses the subtrees roots as keys for a new tree, which as result
431// will have the complete Tree build from bottom to up, where until the
432// log2(nCPU) level it has been computed in parallel.
433func (t *Tree) buildTreeBottomUp(nCPU int, kvs []kv) ([]int, error) {
434 buckets := splitInBuckets(kvs, nCPU)
435 subRoots := make([][]byte, nCPU)
436 invalidsInBucket := make([][]int, nCPU)
437 txs := make([]db.Tx, nCPU)
438
439 var wg sync.WaitGroup
440 wg.Add(nCPU)
441 for i := 0; i < nCPU; i++ {
442 go func(cpu int) {
443 sortKvs(buckets[cpu])
444
445 var err error
446 txs[cpu], err = t.db.NewTx()
447 if err != nil {
448 panic(err) // TODO
449 }
450 bucketTree := Tree{tx: txs[cpu], db: t.db, maxLevels: t.maxLevels,
451 hashFunction: t.hashFunction, root: t.emptyHash}
452
453 currInvalids, err := bucketTree.buildTreeBottomUpSingleThread(buckets[cpu])
454 if err != nil {
455 panic(err) // TODO
456 }
457 invalidsInBucket[cpu] = currInvalids
458 subRoots[cpu] = bucketTree.root
459 wg.Done()
460 }(i)
461 }
462 wg.Wait()
463
464 newRoot, err := t.upFromKeys(subRoots)
465 if err != nil {
466 return nil, err
467 }
468 t.root = newRoot
469
470 var invalids []int
471 for i := 0; i < len(invalidsInBucket); i++ {
472 invalids = append(invalids, invalidsInBucket[i]...)
473 }
474 return invalids, err
475}
476
477// buildTreeBottomUpSingleThread builds the tree with the given []kv from bottom
478// to the root. keys & values must be sorted by path, and the array ks must be
479// length multiple of 2
480func (t *Tree) buildTreeBottomUpSingleThread(kvs []kv) ([]int, error) {
481 // TODO check that log2(len(leafs)) < t.maxLevels, if not, maxLevels
482 // would be reached and should return error
483
484 var invalids []int
485 // build the leafs
486 leafKeys := make([][]byte, len(kvs))
487 for i := 0; i < len(kvs); i++ {
488 // TODO handle the case where Key&Value == 0
489 leafKey, leafValue, err := newLeafValue(t.hashFunction, kvs[i].k, kvs[i].v)
490 if err != nil {
491 // return nil, err
492 invalids = append(invalids, kvs[i].pos)
493 }
494 // store leafKey & leafValue to db
495 if err := t.tx.Put(leafKey, leafValue); err != nil {
496 // return nil, err
497 invalids = append(invalids, kvs[i].pos)
498 }
499 leafKeys[i] = leafKey
500 }
501 r, err := t.upFromKeys(leafKeys)
502 if err != nil {
503 return invalids, err
504 }
505 t.root = r
506 return invalids, nil
507}
508
509// keys & values must be sorted by path, and the array ks must be length
510// multiple of 2
511func (t *Tree) upFromKeys(ks [][]byte) ([]byte, error) {
512 if len(ks) == 1 {
513 return ks[0], nil
514 }
515
516 var rKs [][]byte
517 for i := 0; i < len(ks); i += 2 {
518 // TODO handle the case where Key&Value == 0
519 k, v, err := newIntermediate(t.hashFunction, ks[i], ks[i+1])
520 if err != nil {
521 return nil, err
522 }
523 // store k-v to db
524 if err = t.tx.Put(k, v); err != nil {
525 return nil, err
526 }
527 rKs = append(rKs, k)
528 }
529 return t.upFromKeys(rKs)
530}
531
532func (t *Tree) getLeafs(root []byte) ([][]byte, [][]byte, error) {
533 var ks, vs [][]byte
534 err := t.iter(root, func(k, v []byte) {
535 if v[0] != PrefixValueLeaf {
536 return
537 }
538 leafK, leafV := readLeafValue(v)
539 ks = append(ks, leafK)
540 vs = append(vs, leafV)
541 })
542 return ks, vs, err
543}
544
545func (t *Tree) getKeysAtLevel(l int) ([][]byte, error) {
546 var keys [][]byte
547 err := t.iterWithStop(t.root, 0, func(currLvl int, k, v []byte) bool {
548 if currLvl == l {
549 keys = append(keys, k)
550 }
551 if currLvl >= l {
552 return true // to stop the iter from going down
553 }
554 return false
555 })
556
557 return keys, err
558}
559
560// cutPowerOfTwo returns []kv of length that is a power of 2, and a second []kv
561// with the extra elements that don't fit in a power of 2 length
562func cutPowerOfTwo(kvs []kv) ([]kv, []kv) {
563 x := len(kvs)
564 if (x & (x - 1)) != 0 {
565 p2 := highestPowerOfTwo(x)
566 return kvs[:p2], kvs[p2:]
567 }
568 return kvs, nil
569}
570
571func highestPowerOfTwo(n int) int {
572 res := 0
573 for i := n; i >= 1; i-- {
574 if (i & (i - 1)) == 0 {
575 res = i
576 break
577 }
578 }
579 return res
580}
581
582// func computeSimpleAddCost(nLeafs int) int {
583// // nLvls 2^nLvls
584// nLvls := int(math.Log2(float64(nLeafs)))
585// return nLvls * int(math.Pow(2, float64(nLvls)))
586// }
587//
588// func computeBottomUpAddCost(nLeafs int) int {
589// // 2^nLvls * 2 - 1
590// nLvls := int(math.Log2(float64(nLeafs)))
591// return (int(math.Pow(2, float64(nLvls))) * 2) - 1
592// }