diff options
| -rw-r--r-- | .gitmodules | 3 | ||||
| -rw-r--r-- | CMakeLists.txt | 4 | ||||
| m--------- | external/PGM-index | 0 | ||||
| -rw-r--r-- | include/shard/PGM.h | 345 | ||||
| -rw-r--r-- | tests/pgm_tests.cpp | 217 |
5 files changed, 569 insertions, 0 deletions
diff --git a/.gitmodules b/.gitmodules index a80fc10..36b61ae 100644 --- a/.gitmodules +++ b/.gitmodules @@ -1,3 +1,6 @@ [submodule "external/PLEX"] path = external/PLEX url = git@github.com:stoianmihail/PLEX.git +[submodule "external/PGM-index"] + path = external/PGM-index + url = https://github.com/gvinciguerra/PGM-index.git diff --git a/CMakeLists.txt b/CMakeLists.txt index 420c19c..fc41db3 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -60,6 +60,10 @@ if (tests) add_executable(triespline_tests ${CMAKE_CURRENT_SOURCE_DIR}/tests/triespline_tests.cpp) target_link_libraries(triespline_tests PUBLIC gsl check subunit pthread) target_include_directories(triespline_tests PRIVATE include external/PLEX/include) + + add_executable(pgm_tests ${CMAKE_CURRENT_SOURCE_DIR}/tests/pgm_tests.cpp) + target_link_libraries(pgm_tests PUBLIC gsl check subunit pthread) + target_include_directories(pgm_tests PRIVATE include external/PGM-index/include) endif() # Benchmark build instructions diff --git a/external/PGM-index b/external/PGM-index new file mode 160000 +Subproject 4c1bc8e03307b971f31987fa8b01a52da698aa1 diff --git a/include/shard/PGM.h b/include/shard/PGM.h new file mode 100644 index 0000000..9fad8d0 --- /dev/null +++ b/include/shard/PGM.h @@ -0,0 +1,345 @@ +/* + * include/shard/PGM.h + * + * Copyright (C) 2023 Douglas B. Rumbaugh <drumbaugh@psu.edu> + * + * All rights reserved. Published under the Modified BSD License. + * + */ +#pragma once + + +#include <vector> +#include <cassert> +#include <queue> +#include <memory> +#include <concepts> + +#include "pgm/pgm_index.hpp" +#include "ds/PriorityQueue.h" +#include "util/Cursor.h" +#include "ds/BloomFilter.h" +#include "util/bf_config.h" +#include "framework/MutableBuffer.h" +#include "framework/RecordInterface.h" +#include "framework/ShardInterface.h" +#include "framework/QueryInterface.h" + +namespace de { + +template <RecordInterface R> +struct ts_range_query_parms { + decltype(R::key) lower_bound; + decltype(R::key) upper_bound; +}; + +template <RecordInterface R, bool Rejection> +class PGMLookup; + +template <RecordInterface R> +struct PGMState { + size_t start_idx; + size_t stop_idx; +}; + +template <RecordInterface R> +struct PGMBufferState { + size_t cutoff; + Alias* alias; + decltype(R::weight) max_weight; + + ~PGMBufferState() { + delete alias; + } + +}; + +template <RecordInterface R> +class PGM { +private: + typedef decltype(R::key) K; + typedef decltype(R::value) V; + +public: + + // FIXME: there has to be a better way to do this + friend class PGMLookup<R, true>; + friend class PGMLookup<R, false>; + + PGM(MutableBuffer<R>* buffer) + : m_reccnt(0), m_tombstone_cnt(0) { + + size_t alloc_size = (buffer->get_record_count() * sizeof(Wrapped<R>)) + (CACHELINE_SIZE - (buffer->get_record_count() * sizeof(Wrapped<R>)) % CACHELINE_SIZE); + assert(alloc_size % CACHELINE_SIZE == 0); + m_data = (Wrapped<R>*)std::aligned_alloc(CACHELINE_SIZE, alloc_size); + std::vector<K> keys; + + m_bf = new BloomFilter<K>(BF_FPR, buffer->get_tombstone_count(), BF_HASH_FUNCS); + + size_t offset = 0; + m_reccnt = 0; + auto base = buffer->get_data(); + auto stop = base + buffer->get_record_count(); + + std::sort(base, stop, std::less<Wrapped<R>>()); + + K min_key = base->rec.key; + K max_key = (stop - 1)->rec.key; + + while (base < stop) { + if (!(base->is_tombstone()) && (base + 1) < stop) { + if (base->rec == (base + 1)->rec && (base + 1)->is_tombstone()) { + base += 2; + continue; + } + } else if (base->is_deleted()) { + base += 1; + continue; + } + + base->header &= 1; + m_data[m_reccnt++] = *base; + keys.emplace_back(base->rec.key); + + if (m_bf && base->is_tombstone()) { + m_tombstone_cnt++; + m_bf->insert(base->rec.key); + } + + base++; + } + + if (m_reccnt > 0) { + m_pgm = pgm::PGMIndex<K>(keys); + } + } + + PGM(PGM** shards, size_t len) + : m_reccnt(0), m_tombstone_cnt(0) { + std::vector<Cursor<Wrapped<R>>> cursors; + cursors.reserve(len); + + PriorityQueue<Wrapped<R>> pq(len); + + size_t attemp_reccnt = 0; + size_t tombstone_count = 0; + + for (size_t i = 0; i < len; ++i) { + if (shards[i]) { + auto base = shards[i]->get_data(); + cursors.emplace_back(Cursor{base, base + shards[i]->get_record_count(), 0, shards[i]->get_record_count()}); + attemp_reccnt += shards[i]->get_record_count(); + tombstone_count += shards[i]->get_tombstone_count(); + pq.push(cursors[i].ptr, i); + + } else { + cursors.emplace_back(Cursor<Wrapped<R>>{nullptr, nullptr, 0, 0}); + } + } + + m_bf = new BloomFilter<K>(BF_FPR, tombstone_count, BF_HASH_FUNCS); + + size_t alloc_size = (attemp_reccnt * sizeof(Wrapped<R>)) + (CACHELINE_SIZE - (attemp_reccnt * sizeof(Wrapped<R>)) % CACHELINE_SIZE); + assert(alloc_size % CACHELINE_SIZE == 0); + m_data = (Wrapped<R>*)std::aligned_alloc(CACHELINE_SIZE, alloc_size); + + std::vector<K> keys; + + while (pq.size()) { + auto now = pq.peek(); + auto next = pq.size() > 1 ? pq.peek(1) : queue_record<Wrapped<R>>{nullptr, 0}; + if (!now.data->is_tombstone() && next.data != nullptr && + now.data->rec == next.data->rec && next.data->is_tombstone()) { + + pq.pop(); pq.pop(); + auto& cursor1 = cursors[now.version]; + auto& cursor2 = cursors[next.version]; + if (advance_cursor<Wrapped<R>>(cursor1)) pq.push(cursor1.ptr, now.version); + if (advance_cursor<Wrapped<R>>(cursor2)) pq.push(cursor2.ptr, next.version); + } else { + auto& cursor = cursors[now.version]; + if (!cursor.ptr->is_deleted()) { + m_data[m_reccnt++] = *cursor.ptr; + keys.emplace_back(cursor.ptr->rec.key); + if (m_bf && cursor.ptr->is_tombstone()) { + ++m_tombstone_cnt; + if (m_bf) m_bf->insert(cursor.ptr->rec.key); + } + } + pq.pop(); + + if (advance_cursor<Wrapped<R>>(cursor)) pq.push(cursor.ptr, now.version); + } + } + + if (m_reccnt > 0) { + m_pgm = pgm::PGMIndex<K>(keys); + } + } + + ~PGM() { + if (m_data) free(m_data); + if (m_bf) delete m_bf; + + } + + Wrapped<R> *point_lookup(const R &rec, bool filter=false) { + if (filter && !m_bf->lookup(rec.key)) { + return nullptr; + } + + size_t idx = get_lower_bound(rec.key); + if (idx >= m_reccnt) { + return nullptr; + } + + while (idx < m_reccnt && m_data[idx].rec < rec) ++idx; + + if (m_data[idx].rec == rec) { + return m_data + idx; + } + + return nullptr; + } + + Wrapped<R>* get_data() const { + return m_data; + } + + size_t get_record_count() const { + return m_reccnt; + } + + size_t get_tombstone_count() const { + return m_tombstone_cnt; + } + + const Wrapped<R>* get_record_at(size_t idx) const { + if (idx >= m_reccnt) return nullptr; + return m_data + idx; + } + + + size_t get_memory_usage() { + return 0; + } + +private: + + // FIXME: depending upon the size of the returned bound, + // it may be better to switch between binary search and + // linear scan. + size_t get_lower_bound(const K& key) const { + auto bound = m_pgm.search(key); + size_t idx = bound.lo; + + if (idx >= m_reccnt) { + return m_reccnt; + } + + // if the found location _is_ the key, we're done. + if (m_data[idx].rec.key == key) { + return idx; + } + + // if the found location is larger than the key, we need to + // move backwards towards the beginning of the array + if (m_data[idx].rec.key > key) { + for (ssize_t i=idx; i>=0; i--) { + if (m_data[i].rec.key < key) { + return i+1; + } + } + // otherwise, we move forward towards the end + } else { + for (size_t i=idx; i<m_reccnt; i++) { + if (m_data[i].rec.key >= key) { + return i - 1; + } + } + } + + return m_reccnt; + } + + Wrapped<R>* m_data; + size_t m_reccnt; + size_t m_tombstone_cnt; + K m_max_key; + K m_min_key; + pgm::PGMIndex<K> m_pgm; + BloomFilter<K> *m_bf; +}; + + +template <RecordInterface R> +class PGMRangeQuery { +public: + static void *get_query_state(PGM<R> *ts, void *parms) { + auto res = new PGMState<R>(); + auto p = (ts_range_query_parms<R> *) parms; + + res->start_idx = ts->get_lower_bound(p->lower_bound); + res->stop_idx = ts->get_record_count(); + + return res; + } + + static void* get_buffer_query_state(MutableBuffer<R> *buffer, void *parms) { + auto res = new PGMBufferState<R>(); + res.cutoff = buffer->get_record_count(); + + return res; + } + + static std::vector<Wrapped<R>> query(PGM<R> *ts, void *q_state, void *parms) { + std::vector<Wrapped<R>> records; + auto p = (ts_range_query_parms<R> *) parms; + auto s = (PGMState<R> *) q_state; + auto ptr = ts->get_record_at(s->lower_bound); + size_t i = 0; + while (ptr->rec.key <= p->upper_bound && i < s->stop_idx) { + records.emplace_back(ptr[i]); + i++; + } + + return records; + } + + static std::vector<Wrapped<R>> buffer_query(MutableBuffer<R> *buffer, void *state, void *parms) { + auto p = (ts_range_query_parms<R> *) parms; + auto s = (PGMBufferState<R> *) state; + + + + } + + static std::vector<R> merge(std::vector<std::vector<R>> &results) { + std::vector<R> output; + + for (size_t i=0; i<results.size(); i++) { + for (size_t j=0; j<results[i].size(); j++) { + output.emplace_back(results[i][j]); + } + } + + return output; + } + + static void delete_query_state(void *state) { + auto s = (PGMState<R> *) state; + delete s; + } + + static void delete_buffer_query_state(void *state) { + auto s = (PGMBufferState<R> *) state; + delete s; + } + + + //{q.get_buffer_query_state(p, p)}; + //{q.buffer_query(p, p)}; + +}; + +} diff --git a/tests/pgm_tests.cpp b/tests/pgm_tests.cpp new file mode 100644 index 0000000..33979ae --- /dev/null +++ b/tests/pgm_tests.cpp @@ -0,0 +1,217 @@ +/* + * tests/irs_tests.cpp + * + * Unit tests for PGM (Augmented B+Tree) shard + * + * Copyright (C) 2023 Douglas Rumbaugh <drumbaugh@psu.edu> + * Dong Xie <dongx@psu.edu> + * + * All rights reserved. Published under the Modified BSD License. + * + */ + +#include "shard/PGM.h" +#include "testing.h" + +#include <check.h> + +using namespace de; + +typedef PGM<Rec> Shard; + +START_TEST(t_mbuffer_init) +{ + auto buffer = new MutableBuffer<Rec>(1024, true, 1024); + for (uint64_t i = 512; i > 0; i--) { + uint32_t v = i; + buffer->append({i,v, 1}); + } + + for (uint64_t i = 1; i <= 256; ++i) { + uint32_t v = i; + buffer->append({i, v, 1}, true); + } + + for (uint64_t i = 257; i <= 512; ++i) { + uint32_t v = i + 1; + buffer->append({i, v, 1}); + } + + Shard* shard = new Shard(buffer); + ck_assert_uint_eq(shard->get_record_count(), 512); + + delete buffer; + delete shard; +} + + +START_TEST(t_irs_init) +{ + size_t n = 512; + auto mbuffer1 = create_test_mbuffer<Rec>(n); + auto mbuffer2 = create_test_mbuffer<Rec>(n); + auto mbuffer3 = create_test_mbuffer<Rec>(n); + + auto shard1 = new Shard(mbuffer1); + auto shard2 = new Shard(mbuffer2); + auto shard3 = new Shard(mbuffer3); + + Shard* shards[3] = {shard1, shard2, shard3}; + auto shard4 = new Shard(shards, 3); + + ck_assert_int_eq(shard4->get_record_count(), n * 3); + ck_assert_int_eq(shard4->get_tombstone_count(), 0); + + size_t total_cnt = 0; + size_t shard1_idx = 0; + size_t shard2_idx = 0; + size_t shard3_idx = 0; + + for (size_t i = 0; i < shard4->get_record_count(); ++i) { + auto rec1 = shard1->get_record_at(shard1_idx); + auto rec2 = shard2->get_record_at(shard2_idx); + auto rec3 = shard3->get_record_at(shard3_idx); + + auto cur_rec = shard4->get_record_at(i); + + if (shard1_idx < n && cur_rec->rec == rec1->rec) { + ++shard1_idx; + } else if (shard2_idx < n && cur_rec->rec == rec2->rec) { + ++shard2_idx; + } else if (shard3_idx < n && cur_rec->rec == rec3->rec) { + ++shard3_idx; + } else { + assert(false); + } + } + + delete mbuffer1; + delete mbuffer2; + delete mbuffer3; + + delete shard1; + delete shard2; + delete shard3; + delete shard4; +} + +START_TEST(t_point_lookup) +{ + size_t n = 10000; + + auto buffer = create_double_seq_mbuffer<Rec>(n, false); + auto shard = Shard(buffer); + + for (size_t i=0; i<n; i++) { + Rec r; + auto rec = (buffer->get_data() + i); + r.key = rec->rec.key; + r.value = rec->rec.value; + + auto result = shard.point_lookup(r); + ck_assert_ptr_nonnull(result); + ck_assert_int_eq(result->rec.key, r.key); + ck_assert_int_eq(result->rec.value, r.value); + } + + delete buffer; +} +END_TEST + + +START_TEST(t_point_lookup_miss) +{ + size_t n = 10000; + + auto buffer = create_double_seq_mbuffer<Rec>(n, false); + auto isam = Shard(buffer); + + for (size_t i=n + 100; i<2*n; i++) { + Rec r; + r.key = i; + r.value = i; + + auto result = isam.point_lookup(r); + ck_assert_ptr_null(result); + } + + delete buffer; +} + + +START_TEST(t_full_cancelation) +{ + size_t n = 100; + auto buffer = create_double_seq_mbuffer<Rec>(n, false); + auto buffer_ts = create_double_seq_mbuffer<Rec>(n, true); + + Shard* shard = new Shard(buffer); + Shard* shard_ts = new Shard(buffer_ts); + + ck_assert_int_eq(shard->get_record_count(), n); + ck_assert_int_eq(shard->get_tombstone_count(), 0); + ck_assert_int_eq(shard_ts->get_record_count(), n); + ck_assert_int_eq(shard_ts->get_tombstone_count(), n); + + Shard* shards[] = {shard, shard_ts}; + + Shard* merged = new Shard(shards, 2); + + ck_assert_int_eq(merged->get_tombstone_count(), 0); + ck_assert_int_eq(merged->get_record_count(), 0); + + delete buffer; + delete buffer_ts; + delete shard; + delete shard_ts; + delete merged; +} +END_TEST + + +Suite *unit_testing() +{ + Suite *unit = suite_create("PGM Shard Unit Testing"); + + TCase *create = tcase_create("de::PGM constructor Testing"); + tcase_add_test(create, t_mbuffer_init); + tcase_add_test(create, t_irs_init); + tcase_set_timeout(create, 100); + suite_add_tcase(unit, create); + + + TCase *tombstone = tcase_create("de:PGM::tombstone cancellation Testing"); + tcase_add_test(tombstone, t_full_cancelation); + suite_add_tcase(unit, tombstone); + + + TCase *lookup = tcase_create("de:PGM:point_lookup Testing"); + tcase_add_test(lookup, t_point_lookup); + tcase_add_test(lookup, t_point_lookup_miss); + suite_add_tcase(unit, lookup); + + + return unit; +} + + +int shard_unit_tests() +{ + int failed = 0; + Suite *unit = unit_testing(); + SRunner *unit_shardner = srunner_create(unit); + + srunner_run_all(unit_shardner, CK_NORMAL); + failed = srunner_ntests_failed(unit_shardner); + srunner_free(unit_shardner); + + return failed; +} + + +int main() +{ + int unit_failed = shard_unit_tests(); + + return (unit_failed == 0) ? EXIT_SUCCESS : EXIT_FAILURE; +} |