bench: add tcp accept benchmarks

test/benchmark-list.h

@@ -33,6 +33,10 @@ BENCHMARK_DECLARE (tcp_pump1_client)
 BENCHMARK_DECLARE (pipe_pump100_client)
 BENCHMARK_DECLARE (pipe_pump1_client)
 
+BENCHMARK_DECLARE (tcp_multi_accept2)
+BENCHMARK_DECLARE (tcp_multi_accept4)
+BENCHMARK_DECLARE (tcp_multi_accept8)
+
 /* Run until X packets have been sent/received. */
 BENCHMARK_DECLARE (udp_pummel_1v1)
 BENCHMARK_DECLARE (udp_pummel_1v10)
@@ -112,6 +116,10 @@ TASK_LIST_START
   BENCHMARK_ENTRY  (pipe_pound_1000)
   BENCHMARK_HELPER (pipe_pound_1000, pipe_echo_server)
 
+  BENCHMARK_ENTRY  (tcp_multi_accept2)
+  BENCHMARK_ENTRY  (tcp_multi_accept4)
+  BENCHMARK_ENTRY  (tcp_multi_accept8)
+
   BENCHMARK_ENTRY  (udp_pummel_1v1)
   BENCHMARK_ENTRY  (udp_pummel_1v10)
   BENCHMARK_ENTRY  (udp_pummel_1v100)

test/benchmark-multi-accept.c

@@ -0,0 +1,409 @@
+/* Copyright Joyent, Inc. and other Node contributors. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "task.h"
+#include "uv.h"
+
+#define IPC_PIPE_NAME TEST_PIPENAME
+#define NUM_CONNECTS  (250 * 1000)
+
+union stream_handle {
+  uv_pipe_t pipe;
+  uv_tcp_t tcp;
+};
+
+/* Use as (uv_stream_t *) &handle_storage -- it's kind of clunky but it
+ * avoids aliasing warnings.
+ */
+typedef unsigned char handle_storage_t[sizeof(union stream_handle)];
+
+/* Used for passing around the listen handle, not part of the benchmark proper.
+ * We have an overabundance of server types here. It works like this:
+ *
+ *  1. The main thread starts an IPC pipe server.
+ *  2. The worker threads connect to the IPC server and obtain a listen handle.
+ *  3. The worker threads start accepting requests on the listen handle.
+ *  4. The main thread starts connecting repeatedly.
+ *
+ * Step #4 should perhaps be farmed out over several threads.
+ */
+struct ipc_server_ctx {
+  handle_storage_t server_handle;
+  unsigned int num_connects;
+  uv_pipe_t ipc_pipe;
+};
+
+struct ipc_peer_ctx {
+  handle_storage_t peer_handle;
+  uv_write_t write_req;
+};
+
+struct ipc_client_ctx {
+  uv_connect_t connect_req;
+  uv_stream_t* server_handle;
+  uv_pipe_t ipc_pipe;
+  char scratch[16];
+};
+
+/* Used in the actual benchmark. */
+struct server_ctx {
+  handle_storage_t server_handle;
+  unsigned int num_connects;
+  uv_async_t async_handle;
+  uv_thread_t thread_id;
+  uv_sem_t semaphore;
+};
+
+struct client_ctx {
+  handle_storage_t client_handle;
+  unsigned int num_connects;
+  uv_connect_t connect_req;
+};
+
+static void ipc_connection_cb(uv_stream_t* ipc_pipe, int status);
+static void ipc_write_cb(uv_write_t* req, int status);
+static void ipc_close_cb(uv_handle_t* handle);
+static void ipc_connect_cb(uv_connect_t* req, int status);
+static void ipc_read2_cb(uv_pipe_t* ipc_pipe,
+                         ssize_t nread,
+                         uv_buf_t buf,
+                         uv_handle_type type);
+static uv_buf_t ipc_alloc_cb(uv_handle_t* handle, size_t suggested_size);
+
+static void sv_async_cb(uv_async_t* handle, int status);
+static void sv_connection_cb(uv_stream_t* server_handle, int status);
+
+static void cl_connect_cb(uv_connect_t* req, int status);
+static void cl_close_cb(uv_handle_t* handle);
+
+static struct sockaddr_in listen_addr;
+
+
+static void ipc_connection_cb(uv_stream_t* ipc_pipe, int status) {
+  struct ipc_server_ctx* sc;
+  struct ipc_peer_ctx* pc;
+  uv_loop_t* loop;
+  uv_buf_t buf;
+
+  loop = ipc_pipe->loop;
+  buf = uv_buf_init("PING", 4);
+  sc = container_of(ipc_pipe, struct ipc_server_ctx, ipc_pipe);
+  pc = calloc(1, sizeof(*pc));
+  ASSERT(pc != NULL);
+
+  if (ipc_pipe->type == UV_TCP)
+    ASSERT(0 == uv_tcp_init(loop, (uv_tcp_t*) &pc->peer_handle));
+  else if (ipc_pipe->type == UV_NAMED_PIPE)
+    ASSERT(0 == uv_pipe_init(loop, (uv_pipe_t*) &pc->peer_handle, 1));
+  else
+    ASSERT(0);
+
+  ASSERT(0 == uv_accept(ipc_pipe, (uv_stream_t*) &pc->peer_handle));
+  ASSERT(0 == uv_write2(&pc->write_req,
+                        (uv_stream_t*) &pc->peer_handle,
+                        &buf,
+                        1,
+                        (uv_stream_t*) &sc->server_handle,
+                        ipc_write_cb));
+
+  if (--sc->num_connects == 0)
+    uv_close((uv_handle_t*) ipc_pipe, NULL);
+}
+
+
+static void ipc_write_cb(uv_write_t* req, int status) {
+  struct ipc_peer_ctx* ctx;
+  ctx = container_of(req, struct ipc_peer_ctx, write_req);
+  uv_close((uv_handle_t*) &ctx->peer_handle, ipc_close_cb);
+}
+
+
+static void ipc_close_cb(uv_handle_t* handle) {
+  struct ipc_peer_ctx* ctx;
+  ctx = container_of(handle, struct ipc_peer_ctx, peer_handle);
+  free(ctx);
+}
+
+
+static void ipc_connect_cb(uv_connect_t* req, int status) {
+  struct ipc_client_ctx* ctx;
+  ctx = container_of(req, struct ipc_client_ctx, connect_req);
+  ASSERT(0 == status);
+  ASSERT(0 == uv_read2_start((uv_stream_t*) &ctx->ipc_pipe,
+                             ipc_alloc_cb,
+                             ipc_read2_cb));
+}
+
+
+static uv_buf_t ipc_alloc_cb(uv_handle_t* handle, size_t suggested_size) {
+  struct ipc_client_ctx* ctx;
+  ctx = container_of(handle, struct ipc_client_ctx, ipc_pipe);
+  return uv_buf_init(ctx->scratch, sizeof(ctx->scratch));
+}
+
+
+static void ipc_read2_cb(uv_pipe_t* ipc_pipe,
+                         ssize_t nread,
+                         uv_buf_t buf,
+                         uv_handle_type type) {
+  struct ipc_client_ctx* ctx;
+  uv_loop_t* loop;
+
+  ctx = container_of(ipc_pipe, struct ipc_client_ctx, ipc_pipe);
+  loop = ipc_pipe->loop;
+
+  if (type == UV_TCP)
+    ASSERT(0 == uv_tcp_init(loop, (uv_tcp_t*) ctx->server_handle));
+  else if (type == UV_NAMED_PIPE)
+    ASSERT(0 == uv_pipe_init(loop, (uv_pipe_t*) ctx->server_handle, 0));
+  else
+    ASSERT(0);
+
+  ASSERT(0 == uv_accept((uv_stream_t*) &ctx->ipc_pipe, ctx->server_handle));
+  uv_close((uv_handle_t*) &ctx->ipc_pipe, NULL);
+}
+
+
+/* Set up an IPC pipe server that hands out listen sockets to the worker
+ * threads. It's kind of cumbersome for such a simple operation, maybe we
+ * should revive uv_import() and uv_export().
+ */
+static void send_listen_handles(uv_handle_type type,
+                                unsigned int num_servers,
+                                struct server_ctx* servers) {
+  struct ipc_server_ctx ctx;
+  uv_loop_t* loop;
+  unsigned int i;
+
+  loop = uv_default_loop();
+  ctx.num_connects = num_servers;
+
+  if (type == UV_TCP) {
+    ASSERT(0 == uv_tcp_init(loop, (uv_tcp_t*) &ctx.server_handle));
+    ASSERT(0 == uv_tcp_bind((uv_tcp_t*) &ctx.server_handle, listen_addr));
+  }
+  else if (type == UV_NAMED_PIPE) {
+    ASSERT(0 == uv_pipe_init(loop, (uv_pipe_t*) &ctx.server_handle, 0));
+    ASSERT(0 == uv_pipe_bind((uv_pipe_t*) &ctx.server_handle, IPC_PIPE_NAME));
+  }
+  else
+    ASSERT(0);
+
+  ASSERT(0 == uv_pipe_init(loop, &ctx.ipc_pipe, 1));
+  ASSERT(0 == uv_pipe_bind(&ctx.ipc_pipe, IPC_PIPE_NAME));
+  ASSERT(0 == uv_listen((uv_stream_t*) &ctx.ipc_pipe, 128, ipc_connection_cb));
+
+  for (i = 0; i < num_servers; i++)
+    uv_sem_post(&servers[i].semaphore);
+
+  ASSERT(0 == uv_run(loop));
+  uv_close((uv_handle_t*) &ctx.server_handle, NULL);
+  ASSERT(0 == uv_run(loop));
+
+  for (i = 0; i < num_servers; i++)
+    uv_sem_wait(&servers[i].semaphore);
+}
+
+
+static void get_listen_handle(uv_loop_t* loop, uv_stream_t* server_handle) {
+  struct ipc_client_ctx ctx;
+
+  ctx.server_handle = server_handle;
+  ctx.server_handle->data = "server handle";
+
+  ASSERT(0 == uv_pipe_init(loop, &ctx.ipc_pipe, 1));
+  uv_pipe_connect(&ctx.connect_req,
+                  &ctx.ipc_pipe,
+                  IPC_PIPE_NAME,
+                  ipc_connect_cb);
+  ASSERT(0 == uv_run(loop));
+}
+
+
+static void server_cb(void *arg) {
+  struct server_ctx *ctx;
+  uv_loop_t* loop;
+
+  ctx = arg;
+  loop = uv_loop_new();
+  ASSERT(loop != NULL);
+
+  ASSERT(0 == uv_async_init(loop, &ctx->async_handle, sv_async_cb));
+  uv_unref((uv_handle_t*) &ctx->async_handle);
+
+  /* Wait until the main thread is ready. */
+  uv_sem_wait(&ctx->semaphore);
+  get_listen_handle(loop, (uv_stream_t*) &ctx->server_handle);
+  uv_sem_post(&ctx->semaphore);
+
+  /* Now start the actual benchmark. */
+  ASSERT(0 == uv_listen((uv_stream_t*) &ctx->server_handle,
+                        128,
+                        sv_connection_cb));
+  ASSERT(0 == uv_run(loop));
+
+  uv_loop_delete(loop);
+}
+
+
+static void sv_async_cb(uv_async_t* handle, int status) {
+  struct server_ctx* ctx;
+  ctx = container_of(handle, struct server_ctx, async_handle);
+  uv_close((uv_handle_t*) &ctx->server_handle, NULL);
+  uv_close((uv_handle_t*) &ctx->async_handle, NULL);
+}
+
+
+static void sv_connection_cb(uv_stream_t* server_handle, int status) {
+  handle_storage_t* storage;
+  struct server_ctx* ctx;
+
+  ctx = container_of(server_handle, struct server_ctx, server_handle);
+  ASSERT(status == 0);
+
+  storage = malloc(sizeof(*storage));
+  ASSERT(storage != NULL);
+
+  if (server_handle->type == UV_TCP)
+    ASSERT(0 == uv_tcp_init(server_handle->loop, (uv_tcp_t*) storage));
+  else if (server_handle->type == UV_NAMED_PIPE)
+    ASSERT(0 == uv_pipe_init(server_handle->loop, (uv_pipe_t*) storage, 0));
+  else
+    ASSERT(0);
+
+  ASSERT(0 == uv_accept(server_handle, (uv_stream_t*) storage));
+  uv_close((uv_handle_t*) storage, (uv_close_cb) free);
+  ctx->num_connects++;
+}
+
+
+static void cl_connect_cb(uv_connect_t* req, int status) {
+  struct client_ctx* ctx = container_of(req, struct client_ctx, connect_req);
+  uv_close((uv_handle_t*) &ctx->client_handle, cl_close_cb);
+  ASSERT(status == 0);
+}
+
+
+static void cl_close_cb(uv_handle_t* handle) {
+  struct client_ctx* ctx;
+
+  ctx = container_of(handle, struct client_ctx, client_handle);
+
+  if (--ctx->num_connects == 0)
+    return;
+
+  ASSERT(0 == uv_tcp_init(handle->loop, (uv_tcp_t*) &ctx->client_handle));
+  ASSERT(0 == uv_tcp_connect(&ctx->connect_req,
+                             (uv_tcp_t*) &ctx->client_handle,
+                             listen_addr,
+                             cl_connect_cb));
+}
+
+
+static int test_tcp(unsigned int num_servers, unsigned int num_clients) {
+  struct server_ctx* servers;
+  struct client_ctx* clients;
+  uv_loop_t* loop;
+  uv_tcp_t* handle;
+  unsigned int i;
+  double time;
+
+  listen_addr = uv_ip4_addr("127.0.0.1", TEST_PORT);
+  loop = uv_default_loop();
+
+  servers = calloc(num_servers, sizeof(servers[0]));
+  clients = calloc(num_clients, sizeof(clients[0]));
+  ASSERT(servers != NULL);
+  ASSERT(clients != NULL);
+
+  /* We're making the assumption here that from the perspective of the
+   * OS scheduler, threads are functionally equivalent to and interchangeable
+   * with full-blown processes.
+   */
+  for (i = 0; i < num_servers; i++) {
+    struct server_ctx* ctx = servers + i;
+    ASSERT(0 == uv_sem_init(&ctx->semaphore, 0));
+    ASSERT(0 == uv_thread_create(&ctx->thread_id, server_cb, ctx));
+  }
+
+  send_listen_handles(UV_TCP, num_servers, servers);
+
+  for (i = 0; i < num_clients; i++) {
+    struct client_ctx* ctx = clients + i;
+    ctx->num_connects = NUM_CONNECTS / num_clients;
+    handle = (uv_tcp_t*) &ctx->client_handle;
+    handle->data = "client handle";
+    ASSERT(0 == uv_tcp_init(loop, handle));
+    ASSERT(0 == uv_tcp_connect(&ctx->connect_req,
+                               handle,
+                               listen_addr,
+                               cl_connect_cb));
+  }
+
+  {
+    uint64_t t = uv_hrtime();
+    ASSERT(0 == uv_run(loop));
+    t = uv_hrtime() - t;
+    time = t / 1e9;
+  }
+
+  for (i = 0; i < num_servers; i++) {
+    struct server_ctx* ctx = servers + i;
+    uv_async_send(&ctx->async_handle);
+    ASSERT(0 == uv_thread_join(&ctx->thread_id));
+    uv_sem_destroy(&ctx->semaphore);
+  }
+
+  printf("accept%u: %.0f accepts/sec (%u total)\n",
+         num_servers,
+         NUM_CONNECTS / time,
+         NUM_CONNECTS);
+
+  for (i = 0; i < num_servers; i++) {
+    struct server_ctx* ctx = servers + i;
+    printf("  thread #%u: %.0f accepts/sec (%u total, %.1f%%)\n",
+           i,
+           ctx->num_connects / time,
+           ctx->num_connects,
+           ctx->num_connects * 100.0 / NUM_CONNECTS);
+  }
+
+  free(clients);
+  free(servers);
+  uv_loop_delete(uv_default_loop()); /* Silence valgrind. */
+
+  return 0;
+}
+
+
+BENCHMARK_IMPL(tcp_multi_accept2) {
+  return test_tcp(2, 40);
+}
+
+
+BENCHMARK_IMPL(tcp_multi_accept4) {
+  return test_tcp(4, 40);
+}
+
+
+BENCHMARK_IMPL(tcp_multi_accept8) {
+  return test_tcp(8, 40);
+}

uv.gyp

@@ -372,6 +372,7 @@
         'test/benchmark-list.h',
         'test/benchmark-loop-count.c',
         'test/benchmark-million-timers.c',
+        'test/benchmark-multi-accept.c',
         'test/benchmark-ping-pongs.c',
         'test/benchmark-pound.c',
         'test/benchmark-pump.c',