wbtiger
/
foundationdb
mirror of https://github.com/apple/foundationdb.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
foundationdb/fdbserver/CompactMap.cpp

836 lines
30 KiB
C++
Raw Blame History

/*
* CompactMap.cpp
*
* This source file is part of the FoundationDB open source project
*
* Copyright 2013-2018 Apple Inc. and the FoundationDB project authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#pragma intrinsic(memcmp)
#include "flow/flow.h"
#include "flow/DeterministicRandom.h"
#include "fdbserver/PrefixTree.h"
#include <stdio.h>
static int nextPowerOfTwo(int n) {
int p;
for (p = 1; p < n; p += p);
return p;
}
static int less(StringRef a, StringRef b) {
int al = a.size(), bl = b.size();
int cl = al < bl ? al : bl;
uint8_t const* ap = a.begin();
uint8_t const* bp = b.begin();
for (int i = 0; i < cl; i++) {
if (ap[i] < bp[i])
return 1;
else if (bp[i] < ap[i])
return 0;
}
return al < bl;
}
struct CompactPreOrderTree {
enum { ENABLE_PREFETCH_RIGHT=1 }; // Use rather more memory BW, but hide a little latency when a right branch takes us out of a cache line. Seems to help slightly.
struct Node {
enum { ENABLE_PREFIX=1 }; // Enable or disable key prefix compression within a CompactPreOrderTree
enum { ENABLE_LEFT_PTR = 0 };
// offsets relative to `this`:
enum { KEY_LENGTH_OFFSET = ENABLE_PREFIX*1 };
enum { KEY_DATA_OFFSET = KEY_LENGTH_OFFSET+1 };
// offsets relative to `keyEnd()`:
enum { LPTR_OFFSET = 0 };
enum { RPTR_OFFSET = 2*ENABLE_LEFT_PTR };
enum { END_OFFSET = RPTR_OFFSET + 2 };
enum { IMPLICIT_LPTR_VALUE = END_OFFSET };
static int getMaxOverhead() { return KEY_DATA_OFFSET + END_OFFSET; }
int keyPrefixLength() { if (ENABLE_PREFIX) return *(uint8_t*)this; else return 0; }
int keyLength() { return *((uint8_t*)this + KEY_LENGTH_OFFSET); }
uint8_t const* keyData() { return (uint8_t const*)this + KEY_DATA_OFFSET; }
uint8_t const* keyEnd() { return (uint8_t const*)this + KEY_DATA_OFFSET + keyLength(); }
StringRef key() { return StringRef(keyData(), keyLength()); }
Node* left() {
auto ke = keyEnd();
return (Node*)(ke + (ENABLE_LEFT_PTR ? *(int16_t*)(ke + LPTR_OFFSET) : IMPLICIT_LPTR_VALUE));
}
Node* right() {
auto ke = keyEnd();
return (Node*)(ke + *(uint16_t*)(ke + RPTR_OFFSET));
}
uint8_t* getEnd() { return (uint8_t*)keyEnd() + END_OFFSET; }
void setKeyPrefixLength(int l) {
if (ENABLE_PREFIX) {
ASSERT(l < 256);
*(uint8_t*)this = l;
}
else
ASSERT(!l);
}
void setKeyLength(int l) { ASSERT(l < 256); *((uint8_t*)this + KEY_LENGTH_OFFSET) = l; }
void setLeftPointer(Node* ptr) {
auto ke = keyEnd();
int o = (uint8_t*)ptr - ke;
ASSERT(ENABLE_LEFT_PTR ? (int16_t(o) == o) : o == IMPLICIT_LPTR_VALUE);
if (ENABLE_LEFT_PTR)
*(uint16_t*)(ke + LPTR_OFFSET) = o;
}
void setRightPointer(Node* ptr) {
auto ke = keyEnd();
int o = (uint8_t*)ptr - ke;
ASSERT(-32768 <= o && o < 32767);
*(uint16_t*)(ke + RPTR_OFFSET) = o;
}
};
int nodeCount;
Node root;
int relAddr(Node* n) { return (uint8_t*)n - (uint8_t*)this; }
Node* lastLessOrEqual(StringRef searchKey) {
Node* n = &root; // n is the root of the subtree we are searching
Node* b = 0; // b is the greatest node <= searchKey which is a parent of n
int nBFIndex = 0; // the index of the node n in the entire tree in "breadth first order", i.e. level by level. This is NOT the order the tree is stored in!
int prefixSize = 0; // the number of bytes of searchKey which are equal to the first bytes of the logical key of the parent of n
int dir;
while (nBFIndex < nodeCount) {
int np = n->keyPrefixLength();
if (ENABLE_PREFETCH_RIGHT) _mm_prefetch((const char*)n->right(), _MM_HINT_T0);
if (prefixSize < np) {
// The searchKey differs from this node's logical key in the prefix this node shares with its parent
// So the comparison between this node and searchKey has the same result as the comparison with the parent and searchKey
// (dir is unchanged)
}
else {
// The searchKey is equal to this node's logical key up to the beginning of the compressed key
int al = searchKey.size() - np;
int bl = n->keyLength();
int cl = al < bl ? al : bl;
int prefixLen = commonPrefixLength(searchKey.begin()+np, n->keyData(), cl);
dir = prefixLen == cl ? al<bl : searchKey[np+prefixLen] < n->keyData()[prefixLen];
if (Node::ENABLE_PREFIX) prefixSize = np + prefixLen;
}
nBFIndex = nBFIndex + nBFIndex + 2 - dir;
auto l = n->left(), r = n->right();
b = dir ? b : n;
n = dir ? l : r;
}
return b;
}
static std::pair<Node*, Node*> lastLessOrEqual2(CompactPreOrderTree* this1, CompactPreOrderTree* this2, StringRef searchKey1, StringRef searchKey2) {
// Do two separate lastLessOrEqual operations at once, to make better use of the memory subsystem.
// Don't try to read this code, it is write only (constructed by copy/paste from lastLessOrEqual and adding 1 and 2 to variables as necessary)
Node* n1 = &this1->root; // n is the root of the subtree we are searching
Node* b1 = 0; // b is the greatest node <= searchKey which is a parent of n
int nBFIndex1 = 0; // the index of the node n in the entire tree in "breadth first order", i.e. level by level. This is NOT the order the tree is stored in!
int prefixSize1 = 0; // the number of bytes of searchKey which are equal to the first bytes of the logical key of the parent of n
int dir1;
Node* n2 = &this2->root; // n is the root of the subtree we are searching
Node* b2 = 0; // b is the greatest node <= searchKey which is a parent of n
int nBFIndex2 = 0; // the index of the node n in the entire tree in "breadth first order", i.e. level by level. This is NOT the order the tree is stored in!
int prefixSize2 = 0; // the number of bytes of searchKey which are equal to the first bytes of the logical key of the parent of n
int dir2;
while (nBFIndex1 < this1->nodeCount && nBFIndex2 < this2->nodeCount) {
int np1 = n1->keyPrefixLength();
int np2 = n2->keyPrefixLength();
if (ENABLE_PREFETCH_RIGHT) {
_mm_prefetch((const char*)n1->right(), _MM_HINT_T0);
_mm_prefetch((const char*)n2->right(), _MM_HINT_T0);
}
if (prefixSize1 < np1) {
// The searchKey differs from this node's logical key in the prefix this node shares with its parent
// So the comparison between this node and searchKey has the same result as the comparison with the parent and searchKey
// (dir is unchanged)
}
else {
// The searchKey is equal to this node's logical key up to the beginning of the compressed key
int al1 = searchKey1.size() - np1;
int bl1 = n1->keyLength();
int cl1 = al1 < bl1 ? al1 : bl1;
int prefixLen1 = commonPrefixLength(searchKey1.begin() + np1, n1->keyData(), cl1);
dir1 = prefixLen1 == cl1 ? al1<bl1 : searchKey1[np1 + prefixLen1] < n1->keyData()[prefixLen1];
prefixSize1 = np1 + prefixLen1;
}
if (prefixSize2 < np2) {
// The searchKey differs from this node's logical key in the prefix this node shares with its parent
// So the comparison between this node and searchKey has the same result as the comparison with the parent and searchKey
// (dir is unchanged)
}
else {
// The searchKey is equal to this node's logical key up to the beginning of the compressed key
int al2 = searchKey2.size() - np2;
int bl2 = n2->keyLength();
int cl2 = al2 < bl2 ? al2 : bl2;
int prefixLen2 = commonPrefixLength(searchKey2.begin() + np2, n2->keyData(), cl2);
dir2 = prefixLen2 == cl2 ? al2<bl2 : searchKey2[np2 + prefixLen2] < n2->keyData()[prefixLen2];
prefixSize2 = np2 + prefixLen2;
}
nBFIndex1 = nBFIndex1 + nBFIndex1 + 2 - dir1;
nBFIndex2 = nBFIndex2 + nBFIndex2 + 2 - dir2;
auto l1 = n1->left(), r1 = n1->right();
auto l2 = n2->left(), r2 = n2->right();
b1 = dir1 ? b1 : n1;
b2 = dir2 ? b2 : n2;
n1 = dir1 ? l1 : r1;
n2 = dir2 ? l2 : r2;
}
while (nBFIndex1 < this1->nodeCount) {
int np1 = n1->keyPrefixLength();
if (prefixSize1 < np1) {
// The searchKey differs from this node's logical key in the prefix this node shares with its parent
// So the comparison between this node and searchKey has the same result as the comparison with the parent and searchKey
// (dir is unchanged)
}
else {
// The searchKey is equal to this node's logical key up to the beginning of the compressed key
int al1 = searchKey1.size() - np1;
int bl1 = n1->keyLength();
int cl1 = al1 < bl1 ? al1 : bl1;
int prefixLen1 = commonPrefixLength(searchKey1.begin() + np1, n1->keyData(), cl1);
dir1 = prefixLen1 == cl1 ? al1<bl1 : searchKey1[np1 + prefixLen1] < n1->keyData()[prefixLen1];
prefixSize1 = np1 + prefixLen1;
}
nBFIndex1 = nBFIndex1 + nBFIndex1 + 2 - dir1;
auto l1 = n1->left(), r1 = n1->right();
b1 = dir1 ? b1 : n1;
n1 = dir1 ? l1 : r1;
}
while (nBFIndex2 < this2->nodeCount) {
int np2 = n2->keyPrefixLength();
if (prefixSize2 < np2) {
// The searchKey differs from this node's logical key in the prefix this node shares with its parent
// So the comparison between this node and searchKey has the same result as the comparison with the parent and searchKey
// (dir is unchanged)
}
else {
// The searchKey is equal to this node's logical key up to the beginning of the compressed key
int al2 = searchKey2.size() - np2;
int bl2 = n2->keyLength();
int cl2 = al2 < bl2 ? al2 : bl2;
int prefixLen2 = commonPrefixLength(searchKey2.begin() + np2, n2->keyData(), cl2);
dir2 = prefixLen2 == cl2 ? al2<bl2 : searchKey2[np2 + prefixLen2] < n2->keyData()[prefixLen2];
prefixSize2 = np2 + prefixLen2;
}
nBFIndex2 = nBFIndex2 + nBFIndex2 + 2 - dir2;
auto l2 = n2->left(), r2 = n2->right();
b2 = dir2 ? b2 : n2;
n2 = dir2 ? l2 : r2;
}
return std::make_pair(b1,b2);
}
#if 0
enum { ENABLE_FANCY_BUILD=1 };
struct BuildInfo {
Node* parent;
bool rightChild;
std::string const& prefix;
std::string* begin;
std::string* end;
BuildInfo(Node* parent, bool rightChild, std::string const& prefix, std::string* begin, std::string* end)
: parent(parent), rightChild(rightChild), prefix(prefix), begin(begin), end(end) {}
};
int build(std::vector<std::string>& input, std::string const& prefix = std::string()) {
nodeCount = input.size();
Deque< BuildInfo > queue;
Deque< BuildInfo > deferred;
queue.push_back(BuildInfo(NULL, false, prefix, &input[0], &input[0] + input.size()));
Node* node = &root;
uint8_t* cacheLineEnd = (uint8_t*)node + 64;
while (queue.size() || deferred.size()) {
if (!queue.size()) {
for (int i = 0; i < deferred.size(); i++)
queue.push_back( deferred[i] );
deferred.clear();
}
BuildInfo bi = queue.front();
queue.pop_front();
int mid = perfectSubtreeSplitPoint(bi.end - bi.begin);
std::string& s = bi.begin[mid];
int prefixLen = Node::ENABLE_PREFIX ? commonPrefixLength((uint8_t*)&bi.prefix[0], (uint8_t*)&s[0], std::min(bi.prefix.size(), s.size())) : 0;
node->setKeyPrefixLength(prefixLen);
node->setKeyLength(s.size() - prefixLen);
memcpy((uint8_t*)node->key().begin(), &s[prefixLen], s.size() - prefixLen);
if (bi.parent) {
if (bi.rightChild)
bi.parent->setRightPointer(node);
else
bi.parent->setLeftPointer(node);
}
if ((uint8_t*)node->getEnd() > cacheLineEnd) {
cacheLineEnd = (uint8_t*)((intptr_t)node->getEnd() &~63) + 64;
for (int i = 0; i < queue.size(); i++)
deferred.push_back(queue[i]);
queue.clear();
}
if (bi.begin != bi.begin + mid)
queue.push_back(BuildInfo(node, false, s, bi.begin, bi.begin + mid));
else if (Node::ENABLE_LEFT_PTR)
node->setLeftPointer(node);
if (bi.begin + mid + 1 != bi.end)
queue.push_back(BuildInfo(node, true, s, bi.begin + mid + 1, bi.end));
else
node->setRightPointer(node);
node = (Node*)node->getEnd();
}
return (uint8_t*)node - (uint8_t*)this;
}
#else
enum { ENABLE_FANCY_BUILD=0 };
int build(std::vector<std::string>& input, std::string const& prefix = std::string()) {
nodeCount = input.size();
return (uint8_t*)build(root, prefix, &input[0], &input[0]+input.size()) - (uint8_t*)this;
}
Node* build(Node& node, std::string const& prefix, std::string* begin, std::string* end) {
if (begin == end) return &node;
int mid = perfectSubtreeSplitPoint(end-begin);
std::string& s = begin[mid];
int prefixLen = Node::ENABLE_PREFIX ? commonPrefixLength((uint8_t*)&prefix[0], (uint8_t*)&s[0], std::min(prefix.size(), s.size())) : 0;
//printf("Node: %s at %d, subtree size %d, mid=%d, prefix %d\n", s.c_str(), relAddr(&node), end-begin, mid, prefixLen);
node.setKeyPrefixLength(prefixLen);
node.setKeyLength(s.size()-prefixLen);
memcpy((uint8_t*)node.key().begin(), &s[prefixLen], s.size()-prefixLen);
Node* next = (Node*)node.getEnd();
if (begin != begin + mid) {
node.setLeftPointer(next);
next = build(*node.left(), s, begin, begin + mid);
}
else if (Node::ENABLE_LEFT_PTR)
node.setLeftPointer(&node);
if (begin + mid + 1 != end) {
node.setRightPointer(next);
next = build(*node.right(), s, begin + mid + 1, end);
}
else
node.setRightPointer(&node);
return next;
}
#endif
};
void compactMapTests(std::vector<std::string> testData, std::vector<std::string> sampleQueries, std::string prefixTreeDOTFile = "") {
double t1, t2;
int r = 0;
std::sort(testData.begin(), testData.end());
/*for (int i = 0; i < testData.size() - 1; i++) {
ASSERT(testData[i + 1].substr(0, 4) != testData[i].substr(0, 4));
ASSERT(_byteswap_ulong(*(uint32_t*)&testData[i][0]) < _byteswap_ulong(*(uint32_t*)&testData[i + 1][0]));
}*/
int totalKeyBytes = 0;
for (auto& s : testData)
totalKeyBytes += s.size();
printf("%d bytes in %lu keys\n", totalKeyBytes, testData.size());
for (int i = 0; i < 5; i++)
printf(" '%s'\n", printable(StringRef(testData[i])).c_str());
CompactPreOrderTree* t = (CompactPreOrderTree*)new uint8_t[sizeof(CompactPreOrderTree) + totalKeyBytes + CompactPreOrderTree::Node::getMaxOverhead() * testData.size()];
t1 = timer_monotonic();
int compactTreeBytes = t->build(testData);
t2 = timer_monotonic();
printf("Compact tree is %d bytes\n", compactTreeBytes);
printf("Build time %0.0f us (%0.2f M/sec)\n", (t2 - t1)*1e6, 1 / (t2 - t1) / 1e6);
t1 = timer_monotonic();
const int nBuild = 20000;
for (int i = 0; i < nBuild; i++)
r += t->build(testData);
t2 = timer_monotonic();
printf("Build time %0.0f us (%0.2f M/sec)\n", (t2 - t1)/nBuild*1e6, nBuild / (t2 - t1) / 1e6);
PrefixTree *pt = (PrefixTree *)new uint8_t[sizeof(PrefixTree) + totalKeyBytes + testData.size() * PrefixTree::Node::getMaxOverhead(1, 256, 256)];
std::vector<PrefixTree::EntryRef> keys;
for(auto &k : testData) {
keys.emplace_back(k, StringRef());
}
t1 = timer_monotonic();
int prefixTreeBytes = pt->build(&*keys.begin(), &*keys.end(), StringRef(), StringRef());
t2 = timer_monotonic();
if(!prefixTreeDOTFile.empty()) {
FILE *fout = fopen(prefixTreeDOTFile.c_str(), "w");
fprintf(fout, "%s\n", pt->toDOT(StringRef(), StringRef()).c_str());
fclose(fout);
}
// Calculate perfect prefix-compressed size
int perfectSize = testData.front().size();
for(int i = 1; i < testData.size(); ++i) {
int common = commonPrefixLength(StringRef(testData[i]), StringRef(testData[i - 1]));
perfectSize += (testData[i].size() - common);
}
printf("PrefixTree tree is %d bytes\n", prefixTreeBytes);
printf("Perfect compressed size with no overhead is %d, average PrefixTree overhead is %.2f per item\n", perfectSize, double(prefixTreeBytes - perfectSize) / testData.size());
printf("PrefixTree Build time %0.0f us (%0.2f M/sec)\n", (t2 - t1)*1e6, 1 / (t2 - t1) / 1e6);
// Test cursor forward iteration
auto c = pt->getCursor(StringRef(), StringRef());
ASSERT(c.moveFirst());
bool end = false;
for(int i = 0; i < keys.size(); ++i) {
ASSERT(c.getKeyRef() == keys[i].key);
end = !c.moveNext();
}
ASSERT(end);
printf("PrefixTree forward scan passed\n");
// Test cursor backward iteration
ASSERT(c.moveLast());
for(int i = keys.size() - 1; i >= 0; --i) {
ASSERT(c.getKeyRef() == keys[i].key);
end = !c.movePrev();
}
ASSERT(end);
printf("PrefixTree reverse scan passed\n");
t1 = timer_monotonic();
for (int i = 0; i < nBuild; i++)
r += pt->build(&*keys.begin(), &*keys.end(), StringRef(), StringRef());
t2 = timer_monotonic();
printf("PrefixTree Build time %0.0f us (%0.2f M/sec)\n", (t2 - t1)/nBuild*1e6, nBuild / (t2 - t1) / 1e6);
t->lastLessOrEqual(LiteralStringRef("8f9fad2e5e2af980a"));
{
std::string s, s1;
CompactPreOrderTree::Node *n;
for (int i = 0; i < testData.size(); i++) {
s = testData[i];
auto s1 = s;// s.substr(0, s.size() - 1);
if (!s1.back())
s1 = s1.substr(0, s1.size() - 1);
else {
s1.back()--;
s1 += "\xff\xff\xff\xff\xff\xff";
}
auto n = t->lastLessOrEqual(s1);
//printf("lastLessOrEqual(%s) = %s\n", s1.c_str(), n ? n->key().toString().c_str() : "(null)");
ASSERT(i ? testData[i - 1].substr(n->keyPrefixLength()) == n->key() : !n);
n = t->lastLessOrEqual(s);
//printf("lastLessOrEqual(%s) = %s\n", s.c_str(), n ? n->key().toString().c_str() : "(null)");
ASSERT(n->key() == s.substr(n->keyPrefixLength()));
s1 = s + "a";
auto n1 = t->lastLessOrEqual(s1);
//printf("lastLessOrEqual(%s) = %s\n", s1.c_str(), n ? n->key().toString().c_str() : "(null)");
ASSERT(n1->key() == s.substr(n1->keyPrefixLength()));
ASSERT(CompactPreOrderTree::lastLessOrEqual2(t, t, s, s1) == std::make_pair(n, n1));
}
printf("compactMap lastLessOrEqual tests passed\n");
}
{
auto cur = pt->getCursor(StringRef(), StringRef());
for (int i = 0; i < keys.size(); i++) {
StringRef s = keys[i].key;
ASSERT(cur.seekLessThanOrEqual(s));
ASSERT(cur.valid());
ASSERT(cur.getKey() == s);
StringRef shortString = s.substr(0, s.size() - 1);
bool shorter = cur.seekLessThanOrEqual(shortString);
if(i > 0) {
if(shortString >= keys[i - 1].key) {
ASSERT(shorter);
ASSERT(cur.valid());
ASSERT(cur.getKey() == keys[i - 1].key);
}
}
else {
ASSERT(!shorter);
}
ASSERT(cur.seekLessThanOrEqual(s.toString() + '\0'));
ASSERT(cur.valid());
ASSERT(cur.getKey() == s);
}
printf("PrefixTree lastLessOrEqual tests passed\n");
}
printf("Making %lu copies:\n", 2*sampleQueries.size());
std::vector<CompactPreOrderTree*> copies;
for (int i = 0; i < 2*sampleQueries.size(); i++) {
copies.push_back((CompactPreOrderTree*) new uint8_t[compactTreeBytes]);
memcpy(copies.back(), t, compactTreeBytes);
}
deterministicRandom()->randomShuffle(copies);
std::vector<PrefixTree*> prefixTreeCopies;
for (int i = 0; i < 2*sampleQueries.size(); i++) {
prefixTreeCopies.push_back((PrefixTree *) new uint8_t[prefixTreeBytes]);
memcpy(prefixTreeCopies.back(), pt, prefixTreeBytes);
}
deterministicRandom()->randomShuffle(prefixTreeCopies);
std::vector<std::vector<std::string>> array_copies;
for (int i = 0; i < sampleQueries.size(); i++) {
array_copies.push_back(testData);
}
deterministicRandom()->randomShuffle(array_copies);
printf("shuffled\n");
t1 = timer_monotonic();
for (auto& q : sampleQueries)
r += (intptr_t)t->lastLessOrEqual(q);
t2 = timer_monotonic();
printf("compactmap, in cache: %d queries in %0.3f sec: %0.3f M/sec\n", (int)sampleQueries.size(), t2 - t1, sampleQueries.size() / (t2 - t1) / 1e6);
auto cur = pt->getCursor(StringRef(), StringRef());
t1 = timer_monotonic();
for (auto& q : sampleQueries)
r += cur.seekLessThanOrEqual(StringRef(q)) ? 1 : 0;
t2 = timer_monotonic();
printf("prefixtree, in cache: %d queries in %0.3f sec: %0.3f M/sec\n", (int)sampleQueries.size(), t2 - t1, sampleQueries.size() / (t2 - t1) / 1e6);
/* t1 = timer_monotonic();
for (int q = 0; q < sampleQueries.size(); q += 2) {
auto x = CompactPreOrderTree::lastLessOrEqual2(t, t, sampleQueries[q], sampleQueries[q + 1]);
r += (intptr_t)x.first + (intptr_t)x.second;
}
t2 = timer_monotonic();
printf("in cache (2x interleaved): %d queries in %0.3f sec: %0.3f M/sec\n", (int)sampleQueries.size(), t2 - t1, sampleQueries.size() / (t2 - t1) / 1e6);
*/
t1 = timer_monotonic();
for (int q = 0; q < sampleQueries.size(); q++)
r += (intptr_t)copies[q]->lastLessOrEqual(sampleQueries[q]);
t2 = timer_monotonic();
printf("compactmap, out of cache: %d queries in %0.3f sec: %0.3f M/sec\n", (int)sampleQueries.size(), t2 - t1, sampleQueries.size() / (t2 - t1) / 1e6);
std::vector<PrefixTree::Cursor> cursors;
for (int q = 0; q < sampleQueries.size(); q++)
cursors.push_back(prefixTreeCopies[q]->getCursor(StringRef(), StringRef()));
t1 = timer_monotonic();
for (int q = 0; q < sampleQueries.size(); q++)
r += cursors[q].seekLessThanOrEqual(sampleQueries[q]) ? 1 : 0;
t2 = timer_monotonic();
printf("prefixtree, out of cache: %d queries in %0.3f sec: %0.3f M/sec\n", (int)sampleQueries.size(), t2 - t1, sampleQueries.size() / (t2 - t1) / 1e6);
/*
t1 = timer_monotonic();
for (int q = 0; q < sampleQueries.size(); q += 2) {
auto x = CompactPreOrderTree::lastLessOrEqual2(copies[q + sampleQueries.size()], copies[q + sampleQueries.size() + 1], sampleQueries[q], sampleQueries[q + 1]);
r += (intptr_t)x.first + (intptr_t)x.second;
}
t2 = timer_monotonic();
printf("out of cache (2x interleaved): %d queries in %0.3f sec: %0.3f M/sec\n", (int)sampleQueries.size(), t2 - t1, sampleQueries.size() / (t2 - t1) / 1e6);
*/
t1 = timer_monotonic();
for (int q = 0; q < sampleQueries.size(); q++)
r += (intptr_t)(std::lower_bound(array_copies[q].begin(), array_copies[q].end(), sampleQueries[q]) - testData.begin());
t2 = timer_monotonic();
printf("std::lower_bound: %d queries in %0.3f sec: %0.3f M/sec\n", (int)sampleQueries.size(), t2 - t1, sampleQueries.size() / (t2 - t1) / 1e6);
}
std::vector<std::string> sampleDocuments(int N) {
std::vector<std::string> testData;
std::string p = "pre";
std::string n = "\x01""name\x00\x00";
std::string a = "\x01""address\x00\x00";
std::string o = "\x01""orders\x00\x00";
std::string oi = "\x01""id\x00\x00";
std::string oa = "\x01""amount\x00\x00";
std::string dbl = "\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00";
for (int i = 0; i < N; i++) {
std::string id = BinaryWriter::toValue(deterministicRandom()->randomUniqueID(), Unversioned()).substr(12).toString();
testData.push_back(p + id + n);
testData.push_back(p + id + a);
for (int j = 0; j < 5; j++) {
std::string okey = p + id + o + dbl + (char)j;
testData.push_back(okey + oi);
testData.push_back(okey + oa);
}
}
return testData;
}
StringRef shortestKeyBetween(StringRef a, StringRef b) {
int p = commonPrefixLength(a.begin(), b.begin(), std::min(a.size(), b.size()));
ASSERT(p < b.size());
return b.substr(0, p + 1);
}
std::vector<std::string> sampleBPlusTreeSeparators( std::vector<std::string> rawDocs, int prefixToStrip ) {
// In the middle of a B+Tree, we won't have adjacent document keys but separators between
// pages. These need only contain as many bytes as necessary to distinguish the last item
// in the previous page and the first item in the next page ("suffix compression"), and when
// balancing the tree we can move a few keys left or right if it makes a big difference in the
// suffix size ("split interval")
// The B+Tree will presumably also do its own prefix compression, so we trim off the "obvious"
// common prefix for this imaginary middle node
std::vector<std::string> testData;
std::sort(rawDocs.begin(), rawDocs.end());
for (int i = 0; i + 1 < rawDocs.size(); i += 1000) {
StringRef bestSplitPoint = shortestKeyBetween(rawDocs[i], rawDocs[i + 1]);
for (int j = i + 1; j < i + 11; j++) {
StringRef s = shortestKeyBetween(rawDocs[j], rawDocs[j + 1]);
if (s.size() < bestSplitPoint.size())
bestSplitPoint = s;
}
testData.push_back(bestSplitPoint.substr(prefixToStrip).toString());
}
return testData;
}
struct Page {
Page() : tree(nullptr), size(0), sizeBuilt(0), unsortedKeys(0) {}
std::vector<PrefixTree::EntryRef> keys;
PrefixTree *tree;
std::string treeBuffer;
int size;
int sizeBuilt;
int unsortedKeys;
void add(StringRef k) {
keys.emplace_back(k, StringRef());
size += k.size();
++unsortedKeys;
}
void sort() {
static auto cmp = [=](const PrefixTree::EntryRef &a, const PrefixTree::EntryRef &b) { return a.key < b.key; };
if(unsortedKeys > 0) {
// sort newest elements, then merge
std::sort(keys.end() - unsortedKeys, keys.end(), cmp);
std::inplace_merge(keys.begin(), keys.end() - unsortedKeys, keys.end(), cmp);
unsortedKeys = 0;
}
}
int build() {
if(sizeBuilt != size) {
sort();
treeBuffer.reserve(keys.size() * PrefixTree::Node::getMaxOverhead(1, 256, 256) + size);
tree = (PrefixTree *)treeBuffer.data();
int b = tree->build(&*keys.begin(), &*keys.end(), StringRef(), StringRef());
sizeBuilt = size;
return b;
}
return 0;
}
};
void ingestBenchmark() {
std::vector<StringRef> keys_generated;
Arena arena;
std::set<StringRef> testmap;
for(int i = 0; i < 1000000; ++i) {
keys_generated.push_back(StringRef(arena, format("........%02X......%02X.....%02X........%02X",
deterministicRandom()->randomInt(0, 100),
deterministicRandom()->randomInt(0, 100),
deterministicRandom()->randomInt(0, 100),
deterministicRandom()->randomInt(0, 100)
)));
}
double t1 = timer_monotonic();
for(const auto &k : keys_generated)
testmap.insert(k);
double t2 = timer_monotonic();
printf("Ingested %d elements into map, Speed %f M/s\n",
(int)keys_generated.size(), keys_generated.size() / (t2 - t1) / 1e6);
// sort a group after k elements were added
for(int k = 5; k <= 20; k += 5) {
// g is average page delta size
for(int g = 10; g <= 150; g += 10) {
// rebuild page after r bytes added
for(int r = 500; r <= 4000; r += 500) {
double elapsed = timer_monotonic();
int builds = 0;
int buildbytes = 0;
int keybytes = 0;
std::vector<Page *> pages;
int pageCount = keys_generated.size() / g;
pages.resize(pageCount);
for(auto &key : keys_generated) {
int p = deterministicRandom()->randomInt(0, pageCount);
Page *&pPage = pages[p];
if(pPage == nullptr)
pPage = new Page();
Page &page = *pPage;
page.add(key);
keybytes += key.size();
if(page.keys.size() % k == 0) {
page.sort();
}
// Rebuild page after r bytes added
if(page.size - page.sizeBuilt > r) {
int b = page.build();
if(b > 0) {
++builds;
buildbytes += b;
}
}
}
for(auto p : pages) {
if(p) {
int b = p->build();
if(b > 0) {
++builds;
buildbytes += b;
}
}
}
elapsed = timer_monotonic() - elapsed;
printf("%6d keys %6d pages %3f builds/page %6d builds/s %6d pages/s %5d avg keys/page sort every %d deltas rebuild every %5d bytes %7d keys/s %8d keybytes/s\n",
(int)keys_generated.size(), pageCount, (double)builds / pageCount, int(builds / elapsed), int(pageCount/elapsed), g, k, r, int(keys_generated.size() / elapsed), int(keybytes / elapsed));
for(auto p : pages) {
delete p;
}
}
}
}
}
int main() {
printf("CompactMap test\n");
#ifndef NDEBUG
printf("Compiler optimization is OFF\n");
#endif
printf("Key prefix compression is %s\n", CompactPreOrderTree::Node::ENABLE_PREFIX ? "ON" : "OFF");
printf("Right subtree prefetch is %s\n", CompactPreOrderTree::ENABLE_PREFETCH_RIGHT ? "ON" : "OFF");
printf("Left pointer is %s\n", CompactPreOrderTree::Node::ENABLE_LEFT_PTR ? "ON" : "OFF");
printf("Fancy build is %s\n", CompactPreOrderTree::ENABLE_FANCY_BUILD ? "ON" : "OFF");
setThreadLocalDeterministicRandomSeed(1);
//ingestBenchmark();
/*for (int subtree_size = 1; subtree_size < 20; subtree_size++) {
printf("Subtree of size %d:\n", subtree_size);
int s = lessOrEqualPowerOfTwo((subtree_size - 1) / 2 + 1) - 1;
printf(" s=%d\n", s);
printf(" 1 + s + s=%d\n", 1 + s + s);
printf(" left: %d\n", subtree_size - 1 - 2 * s);
printf(" s*2+1: %d %d\n", s * 2 + 1, subtree_size - (s * 2 + 1) - 1);
printf(" n-s-1: %d %d\n", subtree_size-s-1, s);
printf(" min: %d %d\n", std::min(s * 2 + 1, subtree_size - s - 1), subtree_size - std::min(s * 2 + 1, subtree_size - s - 1) - 1);
}*/
printf("\n16 byte hexadecimal random keys\n");
std::vector<std::string> testData;
for (int i = 0; i < 200; i++) {
testData.push_back(deterministicRandom()->randomUniqueID().shortString());
}
std::vector<std::string> sampleQueries;
for (int i = 0; i < 10000; i++) {
sampleQueries.push_back(deterministicRandom()->randomUniqueID().shortString().substr(0, deterministicRandom()->randomInt(0, 16)));
}
compactMapTests(testData, sampleQueries);
printf("\nRaw index keys\n");
testData.clear(); sampleQueries.clear();
for (int i = 0; i < 100; i++) {
testData.push_back(format("%d Main Street #%d, New York NY 12345, United States of America|", 1234 * (i / 100), (i/10) % 10 + 1000) + deterministicRandom()->randomUniqueID().shortString());
}
for (int i = 0; i < 10000; i++)
sampleQueries.push_back(format("%d Main Street", deterministicRandom()->randomInt(1000, 10000)));
compactMapTests(testData, sampleQueries, "graph_addresses.dot");
printf("\nb+tree separators for index keys\n");
testData.clear();
for (int i = 0; i < 100000; i++) {
testData.push_back(format("%d Main Street #%d, New York NY 12345, United States of America|", 12 * (i / 100), (i/10) % 10 + 1000) + deterministicRandom()->randomUniqueID().shortString());
}
testData = sampleBPlusTreeSeparators(testData, 0);
compactMapTests(testData, sampleQueries);
printf("\nraw document keys\n");
testData = sampleDocuments(20);
sampleQueries.clear();
std::string p = "pre";
for (int i = 0; i < 10000; i++)
sampleQueries.push_back(p + BinaryWriter::toValue(deterministicRandom()->randomUniqueID(), Unversioned()).substr(12).toString());
compactMapTests(testData, sampleQueries);
printf("\nb+tree split keys for documents\n");
testData = sampleBPlusTreeSeparators(sampleDocuments(30000), p.size());
compactMapTests(testData, sampleQueries);
return 0;
}
Reference in New Issue View Git Blame Copy Permalink