更新
\n添加 pthreaded C 客户端后，问题又出现了，这表明长连接时间是 TCP 协议的一部分，而不是具体实现的一部分。改变这些协议似乎并不容易。
\n初始问题
\nnet我相信我的问题主要是：当尝试通过 TCP 连接到服务器时，Golang 包会做什么：
在该连接中似乎存在大量开销，服务器响应时间从5 ms几秒增加到几秒。这在生产环境和下面的最小示例中都可以看到。正确的解决方案是使用服务器的连接池，这将被实现。这很大程度上是我的好奇心。
重现：
\nbacklog = 1，运行 client.go。\nbacklog = 100，运行 client.go。\n~260 ms.50 us能够完成连接12 ms，因此没有看到此问题。示例输出backlog = 1（第一次是拨号时间，第二次是完成时间）：
user@computer ~/tcp-tests $ go run client.go 127.0.0.1:46999\nLong Elapsed Time: 216.579\xc2\xb5s, 315.196\xc2\xb5s\nLong Elapsed Time: 274.169\xc2\xb5s, 5.970873ms\nLong Elapsed Time: 74.4\xc2\xb5s, 10.753871ms\nLong Elapsed Time: 590.965\xc2\xb5s, 205.851066ms\nLong Elapsed Time: 1.029287689s, 1.029574065s\nLong Elapsed Time: 1.02945649s, 1.035098229s\n...\nLong Elapsed Time: 3.045881865s, 6.378597166s\nLong Elapsed Time: 3.045314838s, 6.383783688s\nTime taken stats: 2.85 +/- 1.59 s // average +/- STDEV\nMain Taken: 6.384677948s\n示例输出backlog = 100：
...\nLong Elapsed Time: 330.098\xc2\xb5s, 251.004077ms\nLong Elapsed Time: 298.146\xc2\xb5s, 256.187795ms\nLong Elapsed Time: 315.832\xc2\xb5s, 261.523685ms\nTime taken stats: 0.13 +/- 0.08 s\nMain Taken: 263.186955ms\n那么到底发生了什么net.DialTCP（我们也使用了其他风格的表盘，没有明显的差异）导致表盘时间增长？
我倾向于前两个，因为这些1s, 3s, 5s值似乎是神奇的数字。它们既出现在我不起眼的本地机器上，也出现在大规模生产环境中。
这是用 C 编写的最小服务器。感兴趣的配置值是Listenbacklog的参数。
/*\n Adapted from\n https://www.geeksforgeeks.org/tcp-server-client-implementation-in-c/\n \n Compile and run with:\n gcc server.c -o server; ./server\n*/\n#include <stdio.h>\n#include <string.h>\n#include <sys/socket.h>\n#include <arpa/inet.h>\n#include <unistd.h>\n#include <sys/time.h>\n\nint main(void)\n{\n int socket_desc, client_sock, client_size;\n struct sockaddr_in server_addr, client_addr;\n char server_message[2000], client_message[2000];\n\n // Clean buffers:\n memset(server_message, \'\\0\', sizeof(server_message));\n memset(client_message, \'\\0\', sizeof(client_message));\n\n // Create socket:\n socket_desc = socket(AF_INET, SOCK_STREAM, 0);\n\n if(socket_desc < 0){\n printf("Error while creating socket\\n");\n return -1;\n }\n printf("Socket created successfully\\n");\n\n // Set port and IP:\n server_addr.sin_family = AF_INET;\n server_addr.sin_port = htons(46999);\n server_addr.sin_addr.s_addr = inet_addr("127.0.0.1");\n\n // Bind to the set port and IP:\n if(bind(socket_desc, (struct sockaddr*)&server_addr, sizeof(server_addr))<0){\n printf("Couldn\'t bind to the port\\n");\n return -1;\n }\n printf("Done with binding\\n");\n\n // Listen for clients:\n // Increasing the backlog allows the Go client to connect and wait\n // rather than poll/retry.\n if(listen(socket_desc, 100) < 0){\n printf("Error while listening\\n");\n return -1;\n }\n printf("\\nListening for incoming connections.....\\n");\n\n // Accept an incoming connection:\n client_size = sizeof(client_addr);\n int server_run = 1;\n do\n {\n struct timeval start, end;\n double cpu_time_used;\n gettimeofday(&start, NULL);\n client_sock = accept(socket_desc, (struct sockaddr*)&client_addr, &client_size);\n\n if (client_sock < 0){\n printf("Can\'t accept\\n");\n return -1;\n }\n\n // Receive client\'s message:\n if (recv(client_sock, client_message, sizeof(client_message), 0) < 0){\n printf("Couldn\'t receive\\n");\n return -1;\n }\n if (strcmp(client_message, "stop") == 0)\n {\n server_run = 0;\n printf("Received stop message.\\n");\n }\n\n // Respond to client:\n strcpy(server_message, "This is the server\'s message.");\n\n if (send(client_sock, server_message, strlen(server_message), 0) < 0){\n printf("Can\'t send\\n");\n return -1;\n }\n\n // sleep for 5 ms\n usleep(5000);\n\n // Closing the socket:\n close(client_sock);\n gettimeofday(&end, NULL);\n cpu_time_used = (end.tv_usec - start.tv_usec) / 1000.0;\n if (cpu_time_used > 0.0) // overflow in tv_usec if negative\n printf("Server Time: %.4f ms\\n", cpu_time_used);\n } while(server_run);\n\n close(socket_desc);\n\n return 0;\n}\n这是测试 Go 客户端
\n/*\n Adapted from\n https://www.linode.com/docs/guides/developing-udp-and-tcp-clients-and-servers-in-go/\n\n Run once the server.c is compiled and running with:\n go run client.go 127.0.0.1:46999\n*/\npackage main\n\nimport (\n "fmt"\n "net"\n "os"\n "time"\n "github.com/montanaflynn/stats"\n "sync"\n)\nfunc do_message(wg *sync.WaitGroup, connect string, time_taken *float64) {\n defer wg.Done()\n message := make([]byte, 128)\n start_time := time.Now()\n pAddr, err := net.ResolveTCPAddr("tcp", connect)\n if err != nil {\n return\n }\n\n c, err := net.DialTCP("tcp", nil, pAddr)\n if err != nil {\n fmt.Println(err)\n return\n }\n c.SetLinger(0)\n dialed_time := time.Since(start_time)\n\n defer func() {\n c.Close()\n elapsed_time := time.Since(start_time)\n if elapsed_time.Microseconds() > 60 { // microseconds\n fmt.Println("Long Elapsed Time: " + dialed_time.String() + ", " + elapsed_time.String())\n }\n *time_taken = float64(elapsed_time.Microseconds())\n }()\n\n text := "{\\"service\\": \\"magic_service_str\\"}"\n c.Write([]byte(text))\n code, _ := c.Read(message) // Does not actually wait for response.\n code = code\n}\nfunc main() {\n main_start := time.Now()\n arguments := os.Args\n if len(arguments) == 1 {\n fmt.Println("Please provide host:port.")\n return\n }\n n_messages := 50\n wg := new(sync.WaitGroup)\n wg.Add(n_messages)\n times := make([]float64, n_messages)\n for i := 0; i < n_messages; i++ {\n // Used to turn the goroutines into serial implementation\n // time.Sleep(5500 * time.Microsecond)\n go do_message(wg, arguments[1], ×[i])\n }\n wg.Wait()\n avg, _ := stats.Mean(times)\n std, _ := stats.StandardDeviation(times)\n fmt.Println("Time taken stats: " + fmt.Sprintf("%.2f", avg / 1000000.0) + " +/- " + fmt.Sprintf("%.2f", std / 1000000.0) + " s")\n main_taken := time.Since(main_start)\n fmt.Println("Main Taken: " + main_taken.String())\n}\n更新了 C 语言的 pthreaded 客户端并确认问题不是 Golang 实现：
\n// gcc client_p.c -o pclient -lpthread\n#include <stdio.h>\n#include <string.h>\n#include <sys/socket.h>\n#include <arpa/inet.h>\n#include <unistd.h>\n#include <stdlib.h>\n#include<sys/time.h>\n\n\n#include <pthread.h>\n#include <errno.h>\n\n#ifndef THREAD_LOOP_COUNT\n#define THREAD_LOOP_COUNT 1\n#endif\n\n/* Subtract the \xe2\x80\x98struct timeval\xe2\x80\x99 values X and Y,\n storing the result in RESULT.\n Return 1 if the difference is negative, otherwise 0.\n https://www.gnu.org/software/libc/manual/html_node/Calculating-Elapsed-Time.html\n*/\n\nint\ntimeval_subtract (struct timeval *result, struct timeval *x, struct timeval *y)\n{\n /* Perform the carry for the later subtraction by updating y. */\n if (x->tv_usec < y->tv_usec) {\n int nsec = (y->tv_usec - x->tv_usec) / 1000000 + 1;\n y->tv_usec -= 1000000 * nsec;\n y->tv_sec += nsec;\n }\n if (x->tv_usec - y->tv_usec > 1000000) {\n int nsec = (x->tv_usec - y->tv_usec) / 1000000;\n y->tv_usec += 1000000 * nsec;\n y->tv_sec -= nsec;\n }\n\n /* Compute the time remaining to wait.\n tv_usec is certainly positive. */\n result->tv_sec = x->tv_sec - y->tv_sec;\n result->tv_usec = x->tv_usec - y->tv_usec;\n\n /* Return 1 if result is negative. */\n return x->tv_sec < y->tv_sec;\n}\n\n\nstatic void* workerThreadFunc(void* arg)\n{\n int socket_desc;\n struct sockaddr_in server_addr;\n char server_message[2000], client_message[2000];\n // Clean buffers:\n memset(server_message,\'\\0\',sizeof(server_message));\n memset(client_message,\'\\0\',sizeof(client_message));\n // Set port and IP the same as server-side:\n server_addr.sin_family = AF_INET;\n server_addr.sin_port = htons(46999);\n server_addr.sin_addr.s_addr = inet_addr("127.0.0.1");\n int retries = 0;\n struct timeval start, end, difference;\n double cpu_time_used;\n for(int i = 0; i < THREAD_LOOP_COUNT; i++)\n {\n gettimeofday(&start, NULL);\n // Create socket:\n socket_desc = socket(AF_INET, SOCK_STREAM, 0);\n if(socket_desc < 0){\n printf("Unable to create socket\\n");\n return;\n }\n // Send connection request to server:\n while(connect(socket_desc, (struct sockaddr*)&server_addr, sizeof(server_addr)) < 0){\n retries++;\n if (retries > 10)\n {\n printf("Unable to connect\\n");\n retries = 0;\n }\n usleep(5);\n }\n int retries = 0;\n\n // Send the message to server:\n if(send(socket_desc, client_message, strlen("client message."), 0) < 0){\n printf("Unable to send message\\n");\n close(socket_desc);\n return;\n }\n\n // Receive the server\'s response:\n if(recv(socket_desc, server_message, sizeof(server_message), 0) < 0){\n printf("Error while receiving server\'s msg\\n");\n close(socket_desc);\n return;\n }\n\n // Close the socket:\n close(socket_desc);\n gettimeofday(&end, NULL);\n timeval_subtract (&difference, &end, &start);\n double cpu_time_used = (double)difference.tv_sec + (double)difference.tv_usec / 1000000.0;\n printf("Client Time: %.4e s\\n", cpu_time_used);\n }\n}\n\nint main(int argc, char **argv)\n{\n int n_threads = 50; // default value\n if (argc > 1)\n n_threads = atoi(argv[1]);\n\n pthread_t *threads = (pthread_t*)malloc(n_threads * sizeof(pthread_t));\n\n struct timeval start, end, difference;\n gettimeofday(&start, NULL);\n for(int i = 0; i < n_threads; i++)\n {\n int createRet = pthread_create(&threads[i], NULL, workerThreadFunc, NULL);\n if (createRet != 0)\n {\n printf("failed to create thread\\n");\n }\n }\n for(int i = 0; i < n_threads; i++)\n pthread_join(threads[i], NULL);\n gettimeofday(&end, NULL);\n timeval_subtract (&difference, &end, &start);\n double cpu_time_used = (double)difference.tv_sec + (double)difference.tv_usec / 1000000.0;\n printf("Total Client Time: %.4e s\\n", cpu_time_used);\n free(threads);\n return 0;\n}\n正如 @user207421 所指出的，问题在于 TCP 实现，其中包括重试时的指数退避。Golang 和 C 似乎都没有简单的方法来改变这种行为。
答案是：如果吞吐量较高，请不要打开和关闭 TCP 连接——使用连接池。
有一些工作正在考虑消除指数退避，如下链接，但对于特定情况可能有更好的解决方案。有给我的。
ACM SIGCOMM 计算机通信评论，“从 TCP 中删除指数退避”，第 38 卷，第 5 期，2008 年 10 月。