From 957bd0b2ed7d41ecba6c558a3f2749803f5a2c02 Mon Sep 17 00:00:00 2001
From: Douglas Rumbaugh <dbr4@psu.edu>
Date: Thu, 27 Jul 2023 18:22:22 -0400
Subject: PGM-based Point Lookup benchmarks

---
 benchmarks/include/bench_utility.h      |   1 +
 benchmarks/include/standalone_utility.h |  20 +++
 benchmarks/pgm_pl_bench.cpp             |  67 ++++++++++
 benchmarks/upgm_pl_bench.cpp            | 212 ++++++++++++++++++++++++++++++++
 4 files changed, 300 insertions(+)
 create mode 100644 benchmarks/pgm_pl_bench.cpp
 create mode 100644 benchmarks/upgm_pl_bench.cpp

(limited to 'benchmarks')

diff --git a/benchmarks/include/bench_utility.h b/benchmarks/include/bench_utility.h
index 2ab1532..e33b93d 100644
--- a/benchmarks/include/bench_utility.h
+++ b/benchmarks/include/bench_utility.h
@@ -46,6 +46,7 @@ typedef de::EuclidPoint<double, W2V_SIZE> Word2VecRec;
 typedef de::DynamicExtension<WRec, de::WSS<WRec>, de::WSSQuery<WRec>> ExtendedWSS;
 typedef de::DynamicExtension<Rec, de::TrieSpline<Rec>, de::TrieSplineRangeQuery<Rec>> ExtendedTSRQ;
 typedef de::DynamicExtension<Rec, de::PGM<Rec>, de::PGMRangeQuery<Rec>> ExtendedPGMRQ;
+typedef de::DynamicExtension<Rec, de::PGM<Rec>, de::PGMPointLookup<Rec>> ExtendedPGM_PL;
 typedef de::DynamicExtension<Rec, de::MemISAM<Rec>, de::IRSQuery<Rec>> ExtendedISAM_IRS;
 typedef de::DynamicExtension<Rec, de::MemISAM<Rec>, de::ISAMRangeQuery<Rec>> ExtendedISAM_RQ;
 typedef de::DynamicExtension<Word2VecRec, de::VPTree<Word2VecRec>, de::KNNQuery<Word2VecRec>> ExtendedVPTree_KNN;
diff --git a/benchmarks/include/standalone_utility.h b/benchmarks/include/standalone_utility.h
index 272a332..9876e84 100644
--- a/benchmarks/include/standalone_utility.h
+++ b/benchmarks/include/standalone_utility.h
@@ -84,6 +84,26 @@ static void delete_bench_env()
     gsl_rng_free(g_rng);
 }
 
+
+template <typename QP>
+static std::vector<QP> read_lookup_queries(std::string fname, double selectivity) {
+    std::vector<QP> queries;
+
+    FILE *qf = fopen(fname.c_str(), "r");
+    size_t start, stop;
+    double sel;
+    while (fscanf(qf, "%zu%zu%lf\n", &start, &stop, &sel) != EOF) {
+        if (start < stop && std::abs(sel - selectivity) < 0.1) {
+            QP q;
+            q.target_key = start;
+            queries.push_back(q);
+        }
+    }
+    fclose(qf);
+
+    return queries;
+}
+
 template <typename QP>
 static std::vector<QP> read_range_queries(std::string fname, double selectivity) {
     std::vector<QP> queries;
diff --git a/benchmarks/pgm_pl_bench.cpp b/benchmarks/pgm_pl_bench.cpp
new file mode 100644
index 0000000..f798861
--- /dev/null
+++ b/benchmarks/pgm_pl_bench.cpp
@@ -0,0 +1,67 @@
+/*
+ * benchmarks/triespline_rq_bench.cpp
+ *
+ * Copyright (C) 2023 Douglas Rumbaugh <drumbaugh@psu.edu> 
+ *
+ * All rights reserved. Published under the Modified BSD License.
+ *
+ */
+#include "include/bench.h"
+
+int main(int argc, char **argv)
+{
+    if (argc < 5) {
+        fprintf(stderr, "Usage: pgm_pl_bench <filename> <record_count> <delete_proportion> <query_file> [osm_data]\n");
+        exit(EXIT_FAILURE);
+    }
+
+    std::string filename = std::string(argv[1]);
+    size_t record_count = atol(argv[2]);
+    size_t buffer_cap = 1000;
+    size_t scale_factor = 6;
+    double delete_prop = atof(argv[3]);
+    double max_delete_prop = (delete_prop > 0) ? delete_prop : 1;
+    std::string query_file = std::string(argv[4]);
+    bool use_osm = (argc == 6) ? atoi(argv[5]) : 0;
+
+    double insert_batch = 0.1; 
+
+    init_bench_env(record_count, true, use_osm);
+
+    auto de = ExtendedPGM_PL(buffer_cap, scale_factor, max_delete_prop);
+    auto queries = read_lookup_queries<de::PGMPointLookupParms<Rec>>(query_file, .0001);
+
+    std::fstream datafile;
+    datafile.open(filename, std::ios::in | std::ios::binary);
+
+    std::vector<Rec> to_delete;
+
+    // warm up the tree with initial_insertions number of initially inserted
+    // records
+    size_t warmup_cnt = insert_batch * record_count;
+    warmup<ExtendedPGM_PL, Rec>(datafile, de, warmup_cnt, delete_prop, to_delete, true, true);
+
+    size_t insert_cnt = record_count - warmup_cnt;
+
+    insert_tput_bench<ExtendedPGM_PL, Rec>(de, datafile, insert_cnt, delete_prop, to_delete, true);
+    fprintf(stdout, "%ld\t", de.get_memory_usage());
+    query_latency_bench<ExtendedPGM_PL, Rec, de::PGMPointLookupParms<Rec>>(de, queries, 1);
+
+    fprintf(stdout, "\n");
+
+    auto ts = de.create_static_structure();
+
+    fprintf(stdout, "%ld\t", ts->get_memory_usage());
+    static_latency_bench<de::PGM<Rec>, Rec, de::PGMPointLookupParms<Rec>, de::PGMPointLookup<Rec>>(
+        ts, queries, 1
+    );
+    fprintf(stdout, "\n");
+
+    delete ts;
+
+    delete_bench_env();
+    fflush(stdout);
+    fflush(stderr);
+
+    exit(EXIT_SUCCESS);
+}
diff --git a/benchmarks/upgm_pl_bench.cpp b/benchmarks/upgm_pl_bench.cpp
new file mode 100644
index 0000000..e0445b2
--- /dev/null
+++ b/benchmarks/upgm_pl_bench.cpp
@@ -0,0 +1,212 @@
+#include "pgm/pgm_index_dynamic.hpp"
+#include "include/standalone_utility.h"
+
+typedef uint64_t key_type;
+typedef uint64_t value_type;
+
+typedef pgm::DynamicPGMIndex<key_type, value_type, pgm::PGMIndex<key_type, 64>> PGM;
+
+struct record {
+    key_type key;
+    value_type value;
+};
+
+struct query {
+    key_type lower_bound;
+    key_type upper_bound;
+};
+
+template <typename R>
+static bool build_insert_vec(std::fstream &file, std::vector<R> &vec, size_t n, 
+                             double delete_prop, std::vector<R> &to_delete, bool binary=false) {
+    vec.clear();
+    for (size_t i=0; i<n; i++) {
+        R rec;
+        if (!next_record(file, rec, binary)) {
+            if (i == 0) {
+                return false;
+            }
+
+            break;
+        }
+
+        vec.emplace_back(rec);
+
+        if (gsl_rng_uniform(g_rng) < delete_prop + (delete_prop * .1)) {
+            to_delete.emplace_back(rec);
+        }
+    }
+
+    return true;
+}
+
+
+static bool warmup(std::fstream &file, PGM &pgm, size_t count, 
+                   double delete_prop, std::vector<record> to_delete, bool progress=true, bool binary=false) {
+    size_t batch = std::min(.1 * count, 25000.0);
+
+    std::vector<record> insert_vec;
+    std::vector<record> delete_vec;
+    insert_vec.reserve(batch);
+    delete_vec.reserve(batch*delete_prop);
+
+    size_t inserted = 0;
+    size_t delete_idx = 0;
+
+    double last_percent = 0;
+    while (inserted < count) {
+        // Build vector of records to insert and potentially delete
+        auto continue_warmup = build_insert_vec<record>(file, insert_vec, batch, delete_prop, to_delete, binary);
+        if (inserted > batch) {
+            build_delete_vec(to_delete, delete_vec, batch*delete_prop);
+            delete_idx = 0;
+        }
+
+        for (size_t i=0; i<insert_vec.size(); i++) {
+            // process a delete if necessary
+            if (delete_idx < delete_vec.size() && gsl_rng_uniform(g_rng) < delete_prop) {
+                pgm.erase(delete_vec[delete_idx++].key);
+            }
+
+            pgm.insert_or_assign(insert_vec[i].key, insert_vec[i].value);
+            inserted++;
+            progress_update((double) inserted / (double) count, "warming up:");
+        }
+    }
+
+    return true;
+}
+
+
+static void pgm_rq_insert(PGM &pgm, std::fstream &file, size_t insert_cnt, double delete_prop, std::vector<record> &to_delete, bool binary=false) {
+    size_t delete_cnt = insert_cnt * delete_prop;
+
+    size_t applied_deletes = 0;
+    size_t applied_inserts = 0;
+
+    size_t BATCH=1000;
+
+    std::vector<record> insert_vec;
+    std::vector<record> delete_vec;
+    insert_vec.reserve(BATCH);
+    delete_vec.reserve(BATCH*delete_prop);
+
+    size_t delete_idx = 0;
+
+    bool continue_benchmark = true;
+
+    size_t total_time = 0;
+
+    while (applied_inserts < insert_cnt && continue_benchmark) { 
+        continue_benchmark = build_insert_vec(file, insert_vec, BATCH, delete_prop, to_delete, binary);
+        progress_update((double) applied_inserts / (double) insert_cnt, "inserting:");
+        if (applied_deletes < delete_cnt) {
+            build_delete_vec(to_delete, delete_vec, BATCH*delete_prop);
+            delete_idx = 0;
+        }
+
+        if (insert_vec.size() == 0) {
+            break;
+        }
+
+        auto insert_start = std::chrono::high_resolution_clock::now();
+        for (size_t i=0; i<insert_vec.size(); i++) {
+            // process a delete if necessary
+            if (applied_deletes < delete_cnt && delete_idx < delete_vec.size() && gsl_rng_uniform(g_rng) < delete_prop) {
+                pgm.erase(delete_vec[delete_idx++].key);
+                applied_deletes++;
+            }
+
+            // insert the record;
+            pgm.insert_or_assign(insert_vec[i].key, insert_vec[i].value);
+            applied_inserts++;
+        }
+        auto insert_stop = std::chrono::high_resolution_clock::now();
+
+        total_time += std::chrono::duration_cast<std::chrono::nanoseconds>(insert_stop - insert_start).count();
+    } 
+
+    progress_update(1.0, "inserting:");
+
+    size_t throughput = (((double) (applied_inserts + applied_deletes) / (double) total_time) * 1e9);
+
+    fprintf(stdout, "%ld\t", throughput);
+}
+
+
+
+static void pgm_pl_bench(PGM &pgm, std::vector<query> queries, size_t trial_cnt=1) 
+{
+    char progbuf[25];
+    sprintf(progbuf, "sampling:");
+
+    size_t batch_size = 100;
+    size_t batches = trial_cnt / batch_size;
+    size_t total_time = 0;
+
+    std::vector<record> result_set;
+
+    for (int i=0; i<trial_cnt; i++) {
+        auto start = std::chrono::high_resolution_clock::now();
+        for (size_t j=0; j<queries.size(); j++) {
+            auto ptr = pgm.find(queries[j].lower_bound);
+            if (ptr != pgm.end() && ptr->first == queries[j].lower_bound) {
+                result_set.push_back({ptr->first, ptr->second});
+            }
+            result_set.clear();
+        }
+        auto stop = std::chrono::high_resolution_clock::now();
+
+        total_time += std::chrono::duration_cast<std::chrono::nanoseconds>(stop - start).count();
+    }
+
+    size_t latency = total_time / (trial_cnt * queries.size());
+
+    fprintf(stdout, "%ld\t", latency);
+}
+
+int main(int argc, char **argv)
+{
+    if (argc < 5) {
+        fprintf(stderr, "Usage: upgm_pl_bench <filename> <record_count> <delete_proportion> <query_file>\n");
+        exit(EXIT_FAILURE);
+    }
+
+    std::string filename = std::string(argv[1]);
+    size_t record_count = atol(argv[2]);
+    double delete_prop = atof(argv[3]);
+    std::string qfilename = std::string(argv[4]);
+
+    double insert_batch = 0.1; 
+
+    init_bench_env(record_count, true);
+    auto queries = read_range_queries<query>(qfilename, .0001);
+
+    std::vector<std::pair<key_type, value_type>> data;
+    PGM pgm(data.begin(), data.end());
+
+    std::fstream datafile;
+    datafile.open(filename, std::ios::in | std::ios::binary);
+
+    std::vector<record> to_delete;
+
+    // warm up the tree with initial_insertions number of initially inserted
+    // records
+    size_t warmup_cnt = insert_batch * record_count;
+    warmup(datafile, pgm, warmup_cnt, delete_prop, to_delete, true, true);
+
+    size_t insert_cnt = record_count - warmup_cnt;
+
+    pgm_rq_insert(pgm, datafile, insert_cnt, delete_prop, to_delete, true);
+    size_t memory_usage = pgm.size_in_bytes();
+    fprintf(stdout, "%ld\t", memory_usage);
+
+    pgm_pl_bench(pgm, queries);
+    fprintf(stdout, "\n");
+
+    delete_bench_env();
+    fflush(stdout);
+    fflush(stderr);
+
+    exit(EXIT_SUCCESS);
+}
-- 
cgit v1.2.3