From 1ef2e24e9a6a05791b694f4beb7795af343baf78 Mon Sep 17 00:00:00 2001 From: Douglas Rumbaugh Date: Mon, 5 Jun 2023 10:31:18 -0400 Subject: TrieSpline: initial commit of shard --- include/shard/TrieSpline.h | 340 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 340 insertions(+) create mode 100644 include/shard/TrieSpline.h (limited to 'include/shard') diff --git a/include/shard/TrieSpline.h b/include/shard/TrieSpline.h new file mode 100644 index 0000000..5cd01e3 --- /dev/null +++ b/include/shard/TrieSpline.h @@ -0,0 +1,340 @@ +/* + * include/shard/TrieSpline.h + * + * Copyright (C) 2023 Douglas B. Rumbaugh + * + * All rights reserved. Published under the Modified BSD License. + * + */ +#pragma once + + +#include +#include +#include +#include +#include + +#include "PLEX/include/ts/builder.h" +#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 { + +size_t g_max_error = 1024; + +template +struct ts_lookup_parms { + size_t sample_size; + gsl_rng *rng; +}; + +template +class TrieSplineLookup; + +template +struct TrieSplineState { + decltype(R::weight) tot_weight; + + TrieSplineState() { + tot_weight = 0; + } +}; + +template +struct TrieSplineBufferState { + size_t cutoff; + Alias* alias; + decltype(R::weight) max_weight; + + ~TrieSplineBufferState() { + delete alias; + } + +}; + +template +class TrieSpline { +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 TrieSplineLookup; + friend class TrieSplineLookup; + + TrieSpline(MutableBuffer* buffer) + : m_reccnt(0), m_tombstone_cnt(0) { + + size_t alloc_size = (buffer->get_record_count() * sizeof(Wrapped)) + (CACHELINE_SIZE - (buffer->get_record_count() * sizeof(Wrapped)) % CACHELINE_SIZE); + assert(alloc_size % CACHELINE_SIZE == 0); + m_data = (Wrapped*)std::aligned_alloc(CACHELINE_SIZE, alloc_size); + + m_bf = new BloomFilter(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>()); + + K min_key = base->rec.key; + K max_key = stop->rec.key; + + auto bldr = ts::Builder(min_key, max_key, g_max_error); + + 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; + } + + if (m_reccnt == 0) { + m_max_key = m_min_key = base->key; + } else if (base->key > m_max_key) { + m_max_key = base->key; + } else if (base->key < m_min_key) { + m_min_key = base->key; + } + + base->header &= 1; + m_data[m_reccnt++] = *base; + bldr.AddKey(base->key); + + if (m_bf && base->is_tombstone()) { + m_tombstone_cnt++; + m_bf->insert(base->rec.key); + } + + base++; + } + + if (m_reccnt > 0) { + m_ts = bldr.Finalize(); + } + } + + TrieSpline(TrieSpline** shards, size_t len) + : m_reccnt(0), m_tombstone_cnt(0) { + std::vector>> cursors; + cursors.reserve(len); + + PriorityQueue> 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); + + if (i == 0) { + m_max_key = shards[i]->m_max_key; + m_min_key = shards[i]->m_min_key; + } else if (shards[i]->m_max_key > m_max_key) { + m_max_key = shards[i]->m_max_key; + } else if (shards[i]->m_min_key < m_min_key) { + m_min_key = shards[i]->m_min_key; + } + } else { + cursors.emplace_back(Cursor>{nullptr, nullptr, 0, 0}); + } + } + + m_bf = new BloomFilter(BF_FPR, tombstone_count, BF_HASH_FUNCS); + auto bldr = ts::Builder(m_min_key, m_max_key, g_max_error); + + size_t alloc_size = (attemp_reccnt * sizeof(Wrapped)) + (CACHELINE_SIZE - (attemp_reccnt * sizeof(Wrapped)) % CACHELINE_SIZE); + assert(alloc_size % CACHELINE_SIZE == 0); + m_data = (Wrapped*)std::aligned_alloc(CACHELINE_SIZE, alloc_size); + + while (pq.size()) { + auto now = pq.peek(); + auto next = pq.size() > 1 ? pq.peek(1) : queue_record>{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>(cursor1)) pq.push(cursor1.ptr, now.version); + if (advance_cursor>(cursor2)) pq.push(cursor2.ptr, next.version); + } else { + auto& cursor = cursors[now.version]; + if (!cursor.ptr->is_deleted()) { + m_data[m_reccnt++] = *cursor.ptr; + bldr.AddKey(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>(cursor)) pq.push(cursor.ptr, now.version); + } + } + + if (m_reccnt > 0) { + m_ts = bldr.Finalize(); + } + } + + ~TrieSpline() { + if (m_data) free(m_data); + if (m_bf) delete m_bf; + + } + + Wrapped *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* 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* 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_ts.GetSearchBound(key); + size_t idx = bound.begin; + + if (idx >= m_reccnt) { + return m_reccnt; + } + + // if the found location is larger than the key, we need to + // move backwards towards the beginning of the array + if (m_data[idx].key > key) { + for (ssize_t i=idx; i>=0; i--) { + if (m_data[i].key < key) { + return i+1; + } + } + // otherwise, we move forward towards the end + } else { + for (size_t i=idx; i= key) { + return i-1; + } + } + } + + return m_reccnt; + } + + Wrapped* m_data; + size_t m_reccnt; + size_t m_tombstone_cnt; + K m_max_key; + K m_min_key; + ts::TrieSpline m_ts; + BloomFilter *m_bf; +}; + + +template +class TrieSplineLookup { +public: + static void *get_query_state(TrieSpline *wss, void *parms) { + auto res = new TrieSplineState(); + + return res; + } + + static void* get_buffer_query_state(MutableBuffer *buffer, void *parms) { + + } + + static std::vector> query(TrieSpline *wss, void *q_state, void *parms) { + + } + + static std::vector> buffer_query(MutableBuffer *buffer, void *state, void *parms) { + + } + + static std::vector merge(std::vector> &results) { + std::vector output; + + for (size_t i=0; i *) state; + delete s; + } + + static void delete_buffer_query_state(void *state) { + auto s = (TrieSplineBufferState *) state; + delete s; + } + + + //{q.get_buffer_query_state(p, p)}; + //{q.buffer_query(p, p)}; + +}; + +} -- cgit v1.2.3