From 27d36dd9a68e4cf454be2ca7877ece0a34c3e929 Mon Sep 17 00:00:00 2001
From: Douglas Rumbaugh <dbr4@psu.edu>
Date: Wed, 31 Jan 2024 15:48:21 -0500
Subject: Insert throughput benchmark

---
 benchmarks/insert_query_tput.cpp            |  96 ++++++++--------
 benchmarks/query_workload_bench.cpp         | 168 ++++++++++++++++++++++++++++
 include/framework/structure/MutableBuffer.h |   2 +-
 3 files changed, 220 insertions(+), 46 deletions(-)
 create mode 100644 benchmarks/query_workload_bench.cpp

diff --git a/benchmarks/insert_query_tput.cpp b/benchmarks/insert_query_tput.cpp
index 8844d04..ed5bfe9 100644
--- a/benchmarks/insert_query_tput.cpp
+++ b/benchmarks/insert_query_tput.cpp
@@ -23,79 +23,85 @@ typedef de::DynamicExtension<Rec, ISAM, Q> Ext;
 
 std::atomic<bool> inserts_done = false;
 
+void query_thread(Ext *extension, size_t n) {
+    gsl_rng *rng = gsl_rng_alloc(gsl_rng_mt19937);
+    size_t range = n*.0001;
 
-void query_thread(Ext *extension, double selectivity, size_t k, gsl_rng *rng) {
-    TIMER_INIT();
-
-    size_t reccnt = extension->get_record_count();
-
-    size_t range = reccnt * selectivity;
-
-    auto q = new de::rc::Parms<Rec>();
-
-    TIMER_START();
-    for (int64_t i=0; i<k; i++) {
-        size_t start = gsl_rng_uniform_int(rng, reccnt - range);
-
+    de::rc::Parms<Rec> *q = new de::rc::Parms<Rec>();
+    while (!inserts_done.load()) {
+        size_t start = gsl_rng_uniform_int(rng, n - range);
         q->lower_bound = start;
         q->upper_bound = start + range;
         auto res = extension->query(q);
         auto r = res.get();
+        usleep(100);
     }
-    TIMER_STOP();
-    auto query_lat = TIMER_RESULT();
-    fprintf(stdout, "Q\t%ld\t%ld\t%ld\n", reccnt, query_lat, k);
+
+    gsl_rng_free(rng);
     delete q;
 }
 
-void insert_thread(Ext *extension, size_t n, size_t k, gsl_rng *rng) {
-    TIMER_INIT();
-
+void insert_thread(Ext *extension, size_t n, gsl_rng *rng) {
     size_t reccnt = 0;
     Rec r;
-    while (reccnt < n) {
-        auto old_reccnt = reccnt;
-
-        TIMER_START();
-        for (size_t i=0; i<k; i++) {
-            r.key = reccnt;
-            r.value = reccnt;
-
-            if (extension->insert(r)) {
-                reccnt++;
-            }
-        }
-        TIMER_STOP();
-        auto insert_lat = TIMER_RESULT();
-
-        fprintf(stdout, "I\t%ld\t%ld\t%ld\n", reccnt, insert_lat, reccnt - old_reccnt);
+    for (size_t i=0; i<n; i++) {
+        r.key = gsl_rng_uniform_int(rng, n);
+        r.value = gsl_rng_uniform_int(rng, n);
 
-        if (reccnt % 100000 == 0 && reccnt != n)  {
-            auto a = std::thread(query_thread, extension, .01, 20, rng);
-            a.detach();
+        while (!extension->insert(r)) {
+            usleep(1);
         }
     }
+
+    inserts_done.store(true);
 }
 
 int main(int argc, char **argv) {
 
-    /* the closeout routine takes _forever_ ... so we'll just leak the memory */
+    if (argc < 3) {
+        fprintf(stderr, "insert_query_tput reccnt query_threads\n");
+        exit(EXIT_FAILURE);
+    }
+
+    size_t n = atol(argv[1]);
+    size_t qthread_cnt = atol(argv[2]);
+
     auto extension = new Ext(1000, 12000, 8);
-    size_t n = 10000000;
-    size_t per_trial = 1000;
-    double selectivity = .001;
+    gsl_rng * rng = gsl_rng_alloc(gsl_rng_mt19937);
+
+    /* warmup structure w/ 10% of records */
+    size_t warmup = .1 * n;
+    Rec r;
+    for (size_t i=0; i<warmup; i++) {
+        r.key = gsl_rng_uniform_int(rng, n);
+        r.value = gsl_rng_uniform_int(rng, n);
+
+        while (!extension->insert(r)) {
+            usleep(1);
+        }
+    }
+
+    extension->await_next_epoch();
 
     TIMER_INIT();
 
-    gsl_rng * rng = gsl_rng_alloc(gsl_rng_mt19937);
+    std::vector<std::thread> qthreads(qthread_cnt);
 
     TIMER_START();
-    std::thread i_thrd(insert_thread, extension, n, per_trial, rng);
+    std::thread i_thrd(insert_thread, extension, n - warmup, rng);
+    for (size_t i=0; i<qthread_cnt; i++) {
+        qthreads[i] = std::thread(query_thread, extension, n);
+    }
     i_thrd.join();
     TIMER_STOP();
 
+    for (size_t i=0; i<qthread_cnt; i++) {
+        qthreads[i].join();
+    }
+
     auto total_latency = TIMER_RESULT();
-    fprintf(stdout, "T\t%ld\n", total_latency);
+    size_t throughput = (size_t) ((double) (n - warmup) / (double) total_latency * 1e9);
+    fprintf(stdout, "T\t%ld\t%ld\n", total_latency, throughput);
 
     gsl_rng_free(rng);
     delete extension;
diff --git a/benchmarks/query_workload_bench.cpp b/benchmarks/query_workload_bench.cpp
new file mode 100644
index 0000000..114f780
--- /dev/null
+++ b/benchmarks/query_workload_bench.cpp
@@ -0,0 +1,168 @@
+/*
+ *
+ */
+
+#define ENABLE_TIMER
+
+#include <thread>
+
+#include "framework/DynamicExtension.h"
+#include "shard/ISAMTree.h"
+#include "query/rangecount.h"
+#include "framework/interface/Record.h"
+
+#include <gsl/gsl_rng.h>
+
+#include "psu-util/timer.h"
+
+
+typedef de::Record<int64_t, int64_t> Rec;
+typedef de::ISAMTree<Rec> ISAM;
+typedef de::rc::Query<ISAM, Rec> Q;
+typedef de::DynamicExtension<Rec, ISAM, Q> Ext;
+
+size_t g_insert_size = 50000;
+size_t g_insert_frequency = 1000;
+size_t g_query_count = 5000;
+
+void query_thread(Ext *extension, gsl_rng *rng, size_t n, bool parallel=true) {
+    TIMER_INIT();
+    double selectivity = .001;
+    size_t k = 100;
+    size_t range = n * selectivity;
+
+    size_t total_result = 0;
+
+    auto q = new de::rc::Parms<Rec>();
+
+    std::vector<std::future<std::vector<Rec>>> results(k);
+
+    TIMER_START();
+    for (int64_t i=0; i<k; i++) {
+        size_t start = gsl_rng_uniform_int(rng, n - range);
+
+        q->lower_bound = start;
+        q->upper_bound = start + range;
+        results[i] = extension->query(q);
+        if (!parallel) {
+            auto x = results[i].get();
+            total_result += x[0].key;
+        }
+    }
+
+    if (parallel) {
+        for (size_t i=0; i<k; i++) {
+            auto x = results[i].get();
+            total_result += x[0].key;
+        }
+    }
+
+    TIMER_STOP();
+    auto query_lat = TIMER_RESULT();
+    fprintf(stdout, "Q\t%ld\t%ld\t%ld\n", extension->get_record_count(), query_lat, k);
+    fprintf(stderr, "Q Total: %ld\n", total_result);
+    delete q;
+}
+
+void insert_thread(Ext *extension, size_t n) {
+    gsl_rng *rng = gsl_rng_alloc(gsl_rng_mt19937);
+
+    TIMER_INIT();
+    size_t k=1000;
+
+    Rec r;
+    for (size_t i=0; i<g_insert_size; i+=k) {
+        TIMER_START();
+        for (size_t j=0; j<k; j++) {
+            r.key = gsl_rng_uniform_int(rng, n);
+            r.value = gsl_rng_uniform_int(rng, n);
+
+            while (!extension->insert(r)) {
+                _mm_pause();
+            }
+        }
+        TIMER_STOP();
+
+        auto insert_lat = TIMER_RESULT();
+        fprintf(stdout, "I\t%ld\t%ld\t%ld\n", extension->get_record_count(), insert_lat, k);
+    }
+}
+
+void parallel_bench(Ext *extension, gsl_rng *rng, size_t n) {
+    TIMER_INIT();
+
+    TIMER_START();
+    for (size_t i=0; i < g_query_count; i+=100) {
+        query_thread(extension, rng, n);
+        if (i % g_insert_frequency == 0) {
+            auto x = std::thread(insert_thread, extension, n);
+            x.detach();
+        }
+    }
+    TIMER_STOP();
+
+    auto workload_duration = TIMER_RESULT();
+    fprintf(stdout, "W\t%ld\n", workload_duration);
+}
+
+
+void serial_bench(Ext *extension, gsl_rng *rng, size_t n) {
+    TIMER_INIT();
+    TIMER_START();
+    for (size_t i=0; i < g_query_count; i+=100) {
+        query_thread(extension, rng, n, false);
+        if (i % g_insert_frequency == 0) {
+            auto x = std::thread(insert_thread, extension, n);
+            x.join();
+        }
+    }
+    TIMER_STOP();
+
+    auto workload_duration = TIMER_RESULT();
+    fprintf(stdout, "W\t%ld\n", workload_duration);
+}
+
+int main(int argc, char **argv) {
+
+    if (argc < 5) {
+        fprintf(stderr, "query_workload_bench reccnt lwm hwm parallel\n");
+        exit(EXIT_FAILURE);
+    }
+
+    size_t n = atol(argv[1]);
+    size_t lwm = atol(argv[2]);
+    size_t hwm = atol(argv[3]);
+    bool parallel = atoi(argv[4]);
+
+    size_t scale_factor = 8;
+
+    auto extension = new Ext(lwm, hwm, scale_factor);
+    size_t per_trial = 1000;
+    double selectivity = .001;
+
+    gsl_rng * rng = gsl_rng_alloc(gsl_rng_mt19937);
+
+    /* build initial structure */
+    size_t reccnt = 0;
+    Rec r;
+    for (size_t i=0; i<n; i++) {
+        r.key = gsl_rng_uniform_int(rng, n);
+        r.value = gsl_rng_uniform_int(rng, n);
+
+        while (!extension->insert(r)) {
+            _mm_pause();
+        }
+    }
+
+    if (parallel) {
+        parallel_bench(extension, rng, n);
+    } else {
+        serial_bench(extension, rng, n);
+    }
+
+    gsl_rng_free(rng);
+    delete extension;
+    fflush(stderr);
+    fflush(stdout);
+}
+
diff --git a/include/framework/structure/MutableBuffer.h b/include/framework/structure/MutableBuffer.h
index 7edde2f..94a9c41 100644
--- a/include/framework/structure/MutableBuffer.h
+++ b/include/framework/structure/MutableBuffer.h
@@ -158,7 +158,7 @@ public:
 
         /* refuse to advance head while there is an old with one references */
         if (m_old_head.load().refcnt > 0) {
-            fprintf(stderr, "[W]: Refusing to advance head due to remaining reference counts");
+            //fprintf(stderr, "[W]: Refusing to advance head due to remaining reference counts\n");
             return false;
         }
 
-- 
cgit v1.2.3