16 files changed, 710 insertions, 0 deletions
diff --git a/perf/benchmark/CMakeLists.txt b/perf/benchmark/CMakeLists.txt
new file mode 100644
index 0000000..323240b
--- /dev/null
+++ b/perf/benchmark/CMakeLists.txt
@@ -0,0 +1,21 @@
+cmake_minimum_required(VERSION 3.0)
+project(BBQ_BENCHMARK)
+
+# 搜索当前cmake文件所在目录下的c文件
+file(GLOB SRC_LIST "${CMAKE_CURRENT_SOURCE_DIR}/*.c")
+file(GLOB SRC_BBQ_UT_FUNC_LIST "${UT_BBQ_FUNC_DIR}/*.c")
+list(APPEND SRC_LIST ${SRC_BBQ_UT_FUNC_LIST})
+
+# 指定可执行文件输出路径
+set(EXECUTABLE_OUTPUT_PATH ${EXEC_PATH})
+
+add_executable(benchmark ${SRC_LIST}) # 添加可执行程序
+target_link_libraries(benchmark dl iniparser pthread rte_ring rte_eal rte_kvargs rte_telemetry rmind_ringbuf bbq m) # 链接库
+#target_link_libraries(benchmark dl iniparser pthread dpdk rmind_ringbuf bbq m) # 链接库
+
+add_custom_command(  
+    TARGET benchmark POST_BUILD  
+    COMMAND ${CMAKE_COMMAND} -E copy  
+        ${CMAKE_CURRENT_SOURCE_DIR}/benchmark.sh  
+        ${EXECUTABLE_OUTPUT_PATH}/benchmark.sh  
+)  
diff --git a/perf/benchmark/bcm_benchmark.c b/perf/benchmark/bcm_benchmark.c
new file mode 100644
index 0000000..5a7db7a
--- /dev/null
+++ b/perf/benchmark/bcm_benchmark.c
@@ -0,0 +1,153 @@
+/*
+ * @Author: liuyu
+ * @LastEditTime: 2024-07-07 21:19:02
+ * @Email: [email protected]
+ * @Describe: TODO
+ */
+#include "bbq.h"
+#include "bcm_queue.h"
+#include "iniparser.h"
+#include "ut_bbq_func.h"
+#include <pthread.h>
+#include <stdatomic.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/prctl.h>
+#include <unistd.h>
+
+void bcm_report_printf(struct ut_cfg *cfg, struct ut_merge_data *data, struct ut_exit_data **raw_data, uint32_t thread_cnt, enum ut_thread_type ttype) {
+    char name[10] = {0};
+    double latency_ns = 0;
+    double throughput = 0;
+
+    printf("\n---------%s---------\n", ttype == UT_THREAD_PRODUCER ? "生产者" : "消费者");
+    double use_time = ut_clock_time_to_double(&data->use_time);
+    printf("执行时间 : %lf 秒\n", use_time);
+    printf("执行次数 : %lu (burst=%u)\n", data->run_times, cfg->ring.burst_cnt);
+    printf("成功%s : %lu\n", ttype == UT_THREAD_PRODUCER ? "入队" : "出队", data->ok_cnt);
+    printf("数据错误次数 : %lu\n", data->data_error_cnt);
+
+    // 同时有生产者、消费者时才输出
+    if (cfg->ring.producer_cnt > 0 && cfg->ring.consumer_cnt > 0) {
+        throughput = data->ok_cnt / use_time;
+        printf("吞吐 :%.0lf/s (%e/s)\n", throughput, throughput);
+
+        // 多生产者单消费者 或 单生产者多消费才输出
+        if ((cfg->ring.producer_cnt == 1 && cfg->ring.consumer_cnt > 1) ||
+            (cfg->ring.producer_cnt > 1 && cfg->ring.consumer_cnt == 1)) {
+            for (uint32_t i = 0, bbq_head_idx = 1; i < thread_cnt; i++) {
+                if (raw_data[i]->arg->ttype == ttype) {
+                    struct ut_metric tmp_time = ut_clock_time_sub(raw_data[i]->metric_end, raw_data[i]->metric_start);
+                    throughput = raw_data[i]->ok_cnt / ut_clock_time_to_double(&tmp_time);
+                    printf("    %s-%d 吞吐 :%.0lf/s (%e/s)", name, bbq_head_idx, throughput, throughput);
+                    bbq_head_idx++;
+                }
+            }
+        }
+
+        if (ttype == UT_THREAD_CONSUMER && cfg->ring.workload == UT_WORKLOAD_COMPLEX) {
+            latency_ns = data->latency_ns * 1.0 / data->ok_cnt;
+            printf("数据延迟 :%.0lf 纳秒 (%e)\n", latency_ns, latency_ns);
+        }
+
+        latency_ns = data->op_ok_latency_ns * 1.0 / data->ok_cnt;
+        printf("操作延迟 :%.0lf 纳秒 (%e)\n", latency_ns, latency_ns);
+    } else {
+        latency_ns = data->op_err_latency_ns * 1.0 / (data->run_times - data->ok_cnt);
+        if (ttype == UT_THREAD_PRODUCER) {
+            printf("满队入队操作延迟 :%.0lf 纳秒 (%e)\n", latency_ns, latency_ns);
+        } else {
+            printf("空队出队操作延迟 :%.0lf 纳秒 (%e)\n", latency_ns, latency_ns);
+        }
+    }
+}
+
+void bcm_report_generate(struct ut_cfg *cfg, struct ut_exit_data **exit_data, uint32_t thread_cnt) {
+    // ut_report  report;
+
+    struct ut_merge_s merge = {0};
+    ut_merge_all_data(exit_data, thread_cnt, &merge);
+
+    printf("ring类型: %s\n", ut_ring_type_enum2str(cfg->ring.ring_type));
+    printf("简介: %s\n", cfg->base.introduce);
+    printf("配置: %s\n", cfg->base.name);
+    if (cfg->ring.producer_cnt > 0) {
+        bcm_report_printf(cfg, &merge.producer, exit_data, thread_cnt, UT_THREAD_PRODUCER);
+    }
+
+    if (cfg->ring.consumer_cnt > 0) {
+        bcm_report_printf(cfg, &merge.consumer, exit_data, thread_cnt, UT_THREAD_CONSUMER);
+    }
+
+    if (cfg->ring.producer_cnt > 0 && cfg->ring.consumer_cnt > 0) {
+        printf("生产消费个数验证: %s\n", merge.consumer.ok_cnt == merge.producer.ok_cnt ? "相等" : "不等!!!!!!!!!");
+    }
+}
+
+int main(int argc, char *argv[]) {
+    char *config;
+    char *ring_type;
+    uint32_t burst_cnt = 0;
+
+    if (argc == 4) {
+        config = argv[1];
+        ring_type = argv[2];
+        burst_cnt = strtoul(argv[3], NULL, 0);
+        if (burst_cnt <= 0) {
+            burst_cnt = 1;
+        }
+    } else {
+        config = "/root/code/c/bbq/perf/benchmark/config/compare/case1_simple_spsc.ini";
+        ring_type = "dpdk";
+        burst_cnt = 16;
+        UT_ERR_LOG("use default config, ringt_type:%s burst:%u config:%s argc:%d", ring_type, burst_cnt, config, argc);
+    }
+
+    char thread_name[128] = {0};
+    snprintf(thread_name, sizeof(thread_name), "main:%lu", pthread_self());
+    prctl(PR_SET_NAME, thread_name);
+
+    // 加载配置
+    struct ut_info_s info;
+    if (ut_load_config(config, ring_type, burst_cnt, &info.cfg) != 0) {
+        UT_ERR_LOG("load config error");
+        return -1;
+    }
+
+    // 队列初始化
+    int ret = -1;
+    struct ut_queue q;
+    ret = bcm_queue_init(&info.cfg, &q);
+    if (ret != 0) {
+        UT_ERR_LOG("init failed :%d", ret);
+        return ret;
+    }
+
+    // 创建线程
+    pthread_t *threads = ut_threads_create(&info, &q);
+    if (threads == NULL) {
+        UT_ERR_LOG("pthread_arr is NULL");
+        return ret;
+    }
+
+    // 等待所有线程完成，回收数据
+    uint32_t thread_cnt = info.cfg.ring.producer_cnt + info.cfg.ring.consumer_cnt;
+    struct ut_exit_data **exit_data = (struct ut_exit_data **)ut_malloc(UT_MODULE_BCM, sizeof(struct ut_exit_data **) * (thread_cnt));
+    uint32_t i = 0;
+
+    ut_wait_all_threads_exit(&info, thread_cnt, threads, exit_data);
+
+    // 生成benchmark报告
+    bcm_report_generate(&info.cfg, exit_data, thread_cnt);
+
+    // 回收、释放数据
+    for (i = 0; i < thread_cnt; i++) {
+        ut_exit_data_destory(exit_data[i]);
+    }
+    ut_free(UT_MODULE_BCM, exit_data);
+    ut_threads_destory(&info, threads);
+    ut_queue_destory(&q);
+    ut_memory_counter_print();
+
+    return 0;
+}
+\ No newline at end of file
diff --git a/perf/benchmark/bcm_loadconfig.c b/perf/benchmark/bcm_loadconfig.c
new file mode 100644
index 0000000..68f7bcc
--- /dev/null
+++ b/perf/benchmark/bcm_loadconfig.c
@@ -0,0 +1,91 @@
+/*
+ * @Author: liuyu
+ * @LastEditTime: 2024-07-07 22:13:33
+ * @Email: [email protected]
+ * @Describe: TODO
+ */
+#include "bbq.h"
+#include "iniparser.h"
+#include "ut_bbq_func.h"
+
+#include <string.h>
+
+int ut_load_config(const char *config, const char *ring_type, uint32_t burst_cnt, struct ut_cfg *cfg) {
+    int ret = 0;
+    // 加载配置
+    UT_INFO_LOG("load config:%s", config);
+    dictionary *ini = iniparser_load(config);
+    if (ini == NULL) {
+        return -1;
+    }
+
+    strncpy(cfg->base.name, config, sizeof(cfg->base.name) - 1);
+    cfg->base.name[sizeof(cfg->base.name) - 1] = '\0';
+
+    // 获取键值
+    const char *introduce = iniparser_getstring(ini, "base:introduce", "none");
+    strncpy(cfg->base.introduce, introduce, sizeof(cfg->base.introduce) - 1);
+    cfg->base.introduce[sizeof(cfg->base.introduce) - 1] = '\0'; // 手工写上 \0
+    const char *cores = iniparser_getstring(ini, "base:cores", "unknown");
+    ret = sscanf(cores, "%hu-%hu", &cfg->base.core_begin, &cfg->base.core_end);
+    if (ret != 2) {
+        UT_ERR_LOG("cores error: %s", cores);
+        return -1;
+    }
+
+    const char *workload = iniparser_getstring(ini, "ring:workload", "unknown");
+    cfg->ring.workload = ut_workload_str2enum(workload);
+    cfg->ring.entries_cnt = iniparser_getuint64(ini, "ring:entries_cnt", 0);
+    cfg->ring.producer_cnt = iniparser_getint(ini, "ring:producer_cnt", 0);
+    cfg->ring.consumer_cnt = iniparser_getint(ini, "ring:consumer_cnt", 0);
+    cfg->ring.block_count = iniparser_getint(ini, "ring:block_count", 0);
+    cfg->run.run_ok_times = iniparser_getint(ini, "run:run_ok_times", 0);
+    cfg->run.run_time = iniparser_getuint64(ini, "run:run_time", 0);
+
+    // 设置ring_type
+    cfg->ring.ring_type = ut_ring_type_str2enum(ring_type);
+    if (cfg->ring.ring_type >= UT_RING_TYPE_MAX) {
+        UT_ERR_LOG("unknown ring type:%d", cfg->ring.ring_type);
+        return -1;
+    }
+    cfg->ring.burst_cnt = burst_cnt;
+
+    if (cfg->ring.ring_type == UT_RING_TYPE_RMIND) {
+        // rmind仅支持1个消费者，仅支持burst方式
+        if (cfg->ring.consumer_cnt > 1) {
+            UT_ERR_LOG("ring type:%s only support single consumer", UT_RING_TYPE_RMIND_STR);
+            return -1;
+        }
+
+        if (cfg->ring.burst_cnt <= 1) {
+            UT_ERR_LOG("ring type:%s only support burst_cnt > 1 !", UT_RING_TYPE_RMIND_STR);
+            return -1;
+        }
+
+        if (cfg->ring.workload == UT_WORKLOAD_COMPLEX) {
+            UT_ERR_LOG("ring type:%s only support simple workload !", UT_RING_TYPE_RMIND_STR);
+            return -1;
+        }
+    }
+
+    if (cfg->run.run_time == 0 && cfg->run.run_ok_times == 0) {
+        UT_ERR_LOG("At least one of run_time or run_ok_times is not 0");
+        return -1;
+    }
+
+    UT_INFO_LOG("introduce:%s", cfg->base.introduce);
+    UT_INFO_LOG("cores:%u-%u", cfg->base.core_begin, cfg->base.core_end);
+    UT_INFO_LOG("workload:%s(%u)", workload, cfg->ring.workload);
+    UT_INFO_LOG("entries_cnt:%lu", cfg->ring.entries_cnt);
+    UT_INFO_LOG("producer_cnt:%u", cfg->ring.producer_cnt);
+    UT_INFO_LOG("consumer_cnt:%u", cfg->ring.consumer_cnt);
+    UT_INFO_LOG("block_count:%u", cfg->ring.block_count);
+    UT_INFO_LOG("run_ok_times:%lu", cfg->run.run_ok_times);
+    UT_INFO_LOG("run_time:%lu", cfg->run.run_time);
+    UT_INFO_LOG("ring_type:%s(%u)", ring_type, cfg->ring.ring_type);
+    UT_INFO_LOG("burst_cnt:%u", burst_cnt);
+
+    // 释放dictionary对象
+    iniparser_freedict(ini);
+    return 0;
+}
+\ No newline at end of file
diff --git a/perf/benchmark/bcm_queue.c b/perf/benchmark/bcm_queue.c
new file mode 100644
index 0000000..b116f25
--- /dev/null
+++ b/perf/benchmark/bcm_queue.c
@@ -0,0 +1,187 @@
+/*
+ * @Author: liuyu
+ * @LastEditTime: 2024-07-07 20:52:05
+ * @Email: [email protected]
+ * @Describe: TODO
+ */
+#include "bcm_queue.h"
+#include "ringbuf.h"
+
+static __rte_always_inline unsigned int
+bcm_dpdk_ring_enqueue_burst(struct rte_ring *r, void **obj_table, uint32_t n, uint16_t thread_idx, uint32_t *wait_consumed) {
+    UT_AVOID_WARNING(thread_idx);
+    int ret = 0;
+
+    if (wait_consumed) {
+        unsigned int free_space = 0;
+        ret = rte_ring_enqueue_burst(r, (void *const *)obj_table, n, &free_space);
+        *wait_consumed = r->size - free_space - 1;
+    } else {
+        ret = rte_ring_enqueue_burst(r, (void *const *)obj_table, n, NULL);
+    }
+
+    return ret;
+}
+
+int ut_queue_init_dpdk(struct ut_cfg *cfg, struct ut_queue *q) {
+    /* generate eal parameters */
+    const char *eal_args[] = {"bcm_dpdk", "-n", "4", "--proc-type", "auto", "--no-huge", "-m", "2048"};
+    if (rte_eal_init(RTE_DIM(eal_args), (char **)eal_args) < 0) {
+        return -1;
+    }
+
+    q->ring_type = UT_RING_TYPE_DPDK;
+    unsigned int flags = 0;
+    if (cfg->ring.producer_cnt <= 1) {
+        flags |= RING_F_SP_ENQ;
+    } else {
+        flags |= RING_F_MP_RTS_ENQ;
+    }
+
+    if (cfg->ring.consumer_cnt <= 1) {
+        flags |= RING_F_SC_DEQ;
+    } else {
+        flags |= RING_F_MC_RTS_DEQ;
+    }
+
+    q->ring = (void *)rte_ring_create("dpdk_ring", cfg->ring.entries_cnt, rte_socket_id(), flags);
+    if (q->ring == NULL) {
+        return BBQ_ERR_INPUT_NULL;
+    }
+
+    q->ring_free_f = (ut_ring_free_f)rte_ring_free;
+    q->enqueue_f = (ut_ring_enqueue_f)rte_ring_enqueue;
+    q->dequeue_f = (ut_ring_dequeue_f)rte_ring_dequeue;
+    q->enqueue_burst_f = (ut_enqueue_burst_f)bcm_dpdk_ring_enqueue_burst;
+    q->dequeue_burst_f = (ut_dequeue_burst_f)rte_ring_dequeue_burst;
+
+    return BBQ_OK;
+}
+
+unsigned char *rmind_buf;
+uint16_t worker_cnt;
+ringbuf_worker_t **rmind_workers;
+
+void ut_queue_free_rmind(void *ring) {
+    for (uint16_t i = 0; i < worker_cnt; i++) {
+        ringbuf_unregister((ringbuf_t *)ring, rmind_workers[i]);
+    }
+
+    ut_free(UT_MODULE_RMIND, rmind_workers);
+    ut_free(UT_MODULE_RMIND, rmind_buf);
+    ut_free(UT_MODULE_RMIND, ring);
+}
+
+uint32_t ut_enqueue_burst_rmind(void *ring, void **obj_table, uint32_t n, uint16_t thread_idx, uint32_t *wait_consumed) {
+    UT_AVOID_WARNING(wait_consumed);
+    uint32_t cnt = 0;
+    int ret = 0;
+    size_t off = 0;
+    void *obj = NULL;
+    ringbuf_worker_t *w = (ringbuf_worker_t *)rmind_workers[thread_idx];
+    size_t len = sizeof(uintptr_t);
+
+    for (cnt = 0; cnt < n; cnt++) {
+        obj = obj_table[cnt];
+        uintptr_t uptr = (uintptr_t)obj;
+
+        if ((ret = ringbuf_acquire(ring, w, len)) != -1) {
+            off = (size_t)ret;
+            memcpy(&rmind_buf[off], &uptr, len);
+            ringbuf_produce(ring, w);
+        } else {
+            break;
+        }
+    }
+
+    return cnt;
+}
+
+uint32_t ut_dequeue_burst_rmind(void *ring, void *obj_table, uint32_t n, uint32_t *wait_consumed) {
+    UT_AVOID_WARNING(n);
+    UT_AVOID_WARNING(wait_consumed);
+    size_t len = 0;
+    size_t off = 0;
+    size_t per_size = sizeof(void *);
+    void **table = (void **)obj_table;
+
+    if ((len = ringbuf_consume(ring, &off)) != 0) {
+        size_t rem = len;
+        size_t i = 0;
+
+        while (rem) {
+            uintptr_t *data = (uintptr_t *)(&rmind_buf[off]);
+            table[i] = (void *)(*data);
+            i++;
+            off += per_size;
+            rem -= sizeof(void *);
+        }
+        ringbuf_release(ring, len);
+        return i;
+    }
+
+    return 0;
+}
+
+int ut_queue_init_rmind(struct ut_cfg *cfg, struct ut_queue *q) {
+    static size_t ringbuf_obj_size;
+    worker_cnt = cfg->ring.producer_cnt + cfg->ring.consumer_cnt;
+
+    ringbuf_get_sizes(worker_cnt, &ringbuf_obj_size, NULL);
+    ringbuf_t *r = ut_malloc(UT_MODULE_RMIND, ringbuf_obj_size);
+    if (r == NULL) {
+        exit(-1);
+    }
+
+    size_t buf_size = sizeof(void *) * cfg->ring.entries_cnt;
+    rmind_buf = ut_malloc(UT_MODULE_RMIND, buf_size);
+    if (rmind_buf == NULL) {
+        exit(-1);
+    }
+    ringbuf_setup(r, worker_cnt, buf_size);
+
+    rmind_workers = ut_malloc(UT_MODULE_RMIND, sizeof(*rmind_workers) * worker_cnt);
+    if (rmind_workers == NULL) {
+        exit(-1);
+    }
+    for (uint32_t i = 0; i < worker_cnt; i++) {
+        rmind_workers[i] = ringbuf_register(r, i);
+        if (rmind_workers[i] == NULL) {
+            exit(-1);
+        }
+    }
+
+    q->ring = r;
+    q->ring_free_f = (ut_ring_free_f)ut_queue_free_rmind;
+    q->enqueue_f = NULL;
+    q->dequeue_f = NULL;
+    q->enqueue_burst_f = (ut_enqueue_burst_f)ut_enqueue_burst_rmind;
+    q->dequeue_burst_f = (ut_dequeue_burst_f)ut_dequeue_burst_rmind;
+
+    return 0;
+}
+
+int bcm_queue_init(struct ut_cfg *cfg, struct ut_queue *q) {
+    if (cfg == NULL || q == NULL) {
+        return BBQ_ERR_INPUT_NULL;
+    }
+
+    memset(q, 0, sizeof(*q));
+    int ret = -1;
+    q->ring_type = cfg->ring.ring_type;
+    switch (q->ring_type) {
+    case UT_RING_TYPE_DPDK:
+        ret = ut_queue_init_dpdk(cfg, q);
+        break;
+    case UT_RING_TYPE_BBQ:
+        ret = ut_queue_init_bbq(cfg, q);
+        break;
+    case UT_RING_TYPE_RMIND:
+        ret = ut_queue_init_rmind(cfg, q);
+        break;
+    default:
+        return BBQ_ERR;
+    }
+
+    return ret;
+}
diff --git a/perf/benchmark/bcm_queue.h b/perf/benchmark/bcm_queue.h
new file mode 100644
index 0000000..11006e4
--- /dev/null
+++ b/perf/benchmark/bcm_queue.h
@@ -0,0 +1,10 @@
+/*
+ * @Author: liuyu
+ * @LastEditTime: 2024-07-07 21:18:56
+ * @Email: [email protected]
+ * @Describe: TODO
+ */
+#include "rte_ring.h"
+#include "ut_bbq_func.h"
+
+extern int bcm_queue_init(struct ut_cfg *cfg, struct ut_queue *q);
+\ No newline at end of file
diff --git a/perf/benchmark/benchmark.sh b/perf/benchmark/benchmark.sh
new file mode 100755
index 0000000..abde645
--- /dev/null
+++ b/perf/benchmark/benchmark.sh
@@ -0,0 +1,107 @@
+#!/bin/bash
+###
+# @Author: liuyu
+ # @LastEditTime: 2024-06-30 21:43:01
+# @Email: [email protected]
+# @Describe: 运行性能测试的脚本
+###
+
+ring_type_arr=("bbq" "dpdk" "rmind")
+burst_arr=("32" "16" "8" "1")
+
+# 检查参数数量
+if [ "$#" -ne 3 ]; then
+    echo "Usage: $0 <path to benchmark> <path to config file or directory> <ring type>"
+    exit 1
+fi
+
+# 获取参数值
+BENCHMARK_PATH=$1
+CONFIG_DIR=$2
+RING_TYPE=$3
+
+function exec_benchmark_ring_type() {
+    local ini="$1"
+    local ring="$2"
+    local log_file=$3
+    local burst=$4
+
+    # 如果以perf开头的配置文件，还要执行perf统计
+    if [[ $(basename "$ini") == perf* ]]; then
+        echo "skip perf*"
+        # echo perf stat -e L1-dcache-loads,L1-dcache-load-misses "$BENCHMARK_PATH" "$ini" "$ring" "$BURST_CNT"
+        # perf stat -e L1-dcache-loads,L1-dcache-load-misses "$BENCHMARK_PATH" "$ini" "$ring" "$BURST_CNT"
+    else
+        "$BENCHMARK_PATH" "$ini" "$ring" "$burst" 2>&1 | tee -a "$log_file"
+    fi
+}
+
+function exec_benchmark() {
+    # 提取配置文件名（不带路径）
+    local INI_FILE=$1
+    local log_file=$2
+    local burst=$3
+
+    # 执行benchmark命令并传递配置文件作为参数
+    echo "Executing benchmark with $INI_FILE"
+
+    # 使用所有ring_type
+    if [ "$RING_TYPE" == "all" ]; then
+        for ring_type_tmp in "${ring_type_arr[@]}"; do
+            echo start ring:"$ring_type_tmp"
+            exec_benchmark_ring_type "$INI_FILE" "$ring_type_tmp" "$log_file" "$burst"
+        done
+    else
+        exec_benchmark_ring_type "$INI_FILE" "$RING_TYPE" "$log_file" "$burst"
+    fi
+}
+
+# 检查benchmark文件是否存在且可执行
+if [ ! -x "$BENCHMARK_PATH" ]; then
+    echo "Error: Benchmark executable '$BENCHMARK_PATH' does not exist or is not executable."
+    exit 1
+fi
+
+# 检查配置文件目录是否存在
+if [ ! -e "$CONFIG_DIR" ]; then
+    echo "Error: Config directory '$CONFIG_DIR' does not exist."
+    exit 1
+fi
+
+# 创建报告目录
+timestamp=$(date +"%Y%m%d_%H%M%S")
+folder_path="/tmp/bbq/$timestamp"
+rm -rf "$folder_path"
+mkdir -p "$folder_path"
+
+# 开始时间
+start_time=$(date +%s)
+
+for burst in "${burst_arr[@]}"; do
+    burst_dir="$folder_path"/burst_"$burst"
+    mkdir -p "$burst_dir"
+    for i in {1..3}; do
+        echo ======"$i"========
+        report_path="$burst_dir"/report_"$i".txt
+        if [[ -f "$CONFIG_DIR" ]]; then
+            # 如果是文件，直接执行
+            exec_benchmark "$CONFIG_DIR" "$report_path" "$burst"
+        else
+            # 使用 find 命令递归地搜索所有的 .ini 文件，并按文件名排序
+            find "$CONFIG_DIR" -type f -name "*.ini" -print0 | sort -z | while IFS= read -r -d '' INI_FILE; do
+                if [ -f "$INI_FILE" ]; then
+                    exec_benchmark "$INI_FILE" "$report_path" "$burst"
+                fi
+            done
+        fi
+
+        sleep 1
+    done
+done
+
+# 结束时间
+end_time=$(date +%s)
+# 计算时间差（秒）
+runtime=$((end_time - start_time))
+# 输出运行时间
+echo "done, use $runtime second"
diff --git a/perf/benchmark/config/bbq_debug/debug.ini b/perf/benchmark/config/bbq_debug/debug.ini
new file mode 100644
index 0000000..cead690
--- /dev/null
+++ b/perf/benchmark/config/bbq_debug/debug.ini
@@ -0,0 +1,15 @@
+[base]
+    introduce = "bbq简单负载下，多生产者、单消费者 block=0" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "simple" ;负载模式 simple/complex
+    entries_cnt = 8 ;ring初始化时分配entry的个数
+    block_count = 2 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 4 ;生产者个数
+    consumer_cnt = 1 ;消费者个数
+
+[run]
+    run_time = 10 ; 整体运行的秒数，大于0生效
+    enqueue_cnt  = 0 ;每个线程入队次数，大于0生效
+    dequeue_cnt  = 0 ;每个线程出队次数，大于0生效
+\ No newline at end of file
diff --git a/perf/benchmark/config/compare/case1_simple_spsc.ini b/perf/benchmark/config/compare/case1_simple_spsc.ini
new file mode 100644
index 0000000..e1ee087
--- /dev/null
+++ b/perf/benchmark/config/compare/case1_simple_spsc.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case1 简单负载下，单生产者、单消费者 吞吐测试" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "simple" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 1 ;生产者个数
+    consumer_cnt = 1 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
+\ No newline at end of file
diff --git a/perf/benchmark/config/compare/case2_simple_spmc.ini b/perf/benchmark/config/compare/case2_simple_spmc.ini
new file mode 100644
index 0000000..8216381
--- /dev/null
+++ b/perf/benchmark/config/compare/case2_simple_spmc.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case2 简单负载下，单生产者、多消费者 吞吐测试" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "simple" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 1 ;生产者个数
+    consumer_cnt = 4 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
diff --git a/perf/benchmark/config/compare/case3_simple_mpsc.ini b/perf/benchmark/config/compare/case3_simple_mpsc.ini
new file mode 100644
index 0000000..3476774
--- /dev/null
+++ b/perf/benchmark/config/compare/case3_simple_mpsc.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case3 简单负载下，多生产者、单消费者 吞吐测试" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "simple" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 4 ;生产者个数
+    consumer_cnt = 1 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
+\ No newline at end of file
diff --git a/perf/benchmark/config/compare/case4_complex_spmc.ini b/perf/benchmark/config/compare/case4_complex_spmc.ini
new file mode 100644
index 0000000..64e3e41
--- /dev/null
+++ b/perf/benchmark/config/compare/case4_complex_spmc.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case4 复杂负载下，单生产者、多消费者 吞吐测试" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "complex" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 1 ;生产者个数
+    consumer_cnt = 4 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
+\ No newline at end of file
diff --git a/perf/benchmark/config/compare/case5_complex_mpsc.ini b/perf/benchmark/config/compare/case5_complex_mpsc.ini
new file mode 100644
index 0000000..2404649
--- /dev/null
+++ b/perf/benchmark/config/compare/case5_complex_mpsc.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case5 复杂负载下，多生产者、单消费者 吞吐测试" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "complex" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 4 ;生产者个数
+    consumer_cnt = 1 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
+\ No newline at end of file
diff --git a/perf/benchmark/config/compare/case6_simple_mp0c.ini b/perf/benchmark/config/compare/case6_simple_mp0c.ini
new file mode 100644
index 0000000..62325e4
--- /dev/null
+++ b/perf/benchmark/config/compare/case6_simple_mp0c.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case6 简单负载下，多生产者、无消费者，测试满队入队操作时延" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "simple" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 4 ;生产者个数
+    consumer_cnt = 0 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
+\ No newline at end of file
diff --git a/perf/benchmark/config/compare/case7_simple_0pmc.ini b/perf/benchmark/config/compare/case7_simple_0pmc.ini
new file mode 100644
index 0000000..1c5ed2c
--- /dev/null
+++ b/perf/benchmark/config/compare/case7_simple_0pmc.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case7 简单负载下，无生产者、多消费者，测试空队出队操作时延" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "simple" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 0 ;生产者个数
+    consumer_cnt = 4 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
+\ No newline at end of file
diff --git a/perf/benchmark/config/compare/case8_simple_mpmc.ini b/perf/benchmark/config/compare/case8_simple_mpmc.ini
new file mode 100644
index 0000000..b5026bb
--- /dev/null
+++ b/perf/benchmark/config/compare/case8_simple_mpmc.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case8 简单负载下，多生产者、多消费者 吞吐测试" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "simple" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 4 ;生产者个数
+    consumer_cnt = 4 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
+\ No newline at end of file
diff --git a/perf/benchmark/config/compare/case9_simple_mpmc_overcore.ini b/perf/benchmark/config/compare/case9_simple_mpmc_overcore.ini
new file mode 100644
index 0000000..b4b304a
--- /dev/null
+++ b/perf/benchmark/config/compare/case9_simple_mpmc_overcore.ini
@@ -0,0 +1,14 @@
+[base]
+    introduce = "general case9 简单负载下，多生产者、多消费者 线程核心超过已有核心时的吞吐测试" ;测试配置说明
+    cores = "20-27" ;测试用核心起止范围
+
+[ring]
+    workload    = "simple" ;负载模式 simple/complex
+    entries_cnt = 4096 ;ring初始化时分配entry的个数
+    block_count = 0 ;bbq配置，等于0则表示根据entries_cnt自动计算
+    producer_cnt = 16 ;生产者个数
+    consumer_cnt = 16 ;消费者个数
+
+[run]
+    run_time = 15 ; 整体运行的秒数，大于0生效
+    run_ok_times  = 0 ;成功入队/出队次数，大于0生效。
+\ No newline at end of file