多线程测试下，性能波动大

[Cai-Yao]

client

#include <rest_rpc.hpp>
#include <sys/time.h>
#include <vector>
#include <thread>
#include <chrono>
static long long getCurrentPreciseTime() {
    struct timeval tv;  
    gettimeofday(&tv,NULL); 
    return (long long)tv.tv_sec * 1000 * 1000 + tv.tv_usec;
}

using namespace rest_rpc;
using namespace rest_rpc::rpc_service;
void test_add() {
	rpc_client client("127.0.0.1", 9123);
	{
		std::vector<char> req;
		req.resize(128 * 100);
		for (int i = 0; i < 10; i++) {
			bool r = client.connect();
			if (!r) {
				std::cout << "connect timeout" << std::endl;
				return;
			}
			
			auto result = client.call<std::tuple<int32_t, std::vector<char>, int32_t>>("tuple", i, i, req);
		}
		
	}
	// client.close();
	client.stop();
}

int main()
{
	std::vector<std::thread> threads;
	long long start_time = getCurrentPreciseTime();
	for(int i = 0; i < 150; i++) {
		threads.push_back(std::thread(test_add));
	}
	for(int i = 0; i < 150; i++) {
		threads[i].join();
	}
	long long end_time = getCurrentPreciseTime();
	printf("read cost %lld\n", end_time - start_time);
	return 0;
}

server

#include "rest_rpc.hpp"
#include <string>
using namespace rest_rpc;
using namespace rpc_service;

struct dummy{
	std::tuple<int32_t, std::vector<char>, int32_t> rpc_test_tuple(rpc_conn conn, int32_t x, int32_t y, std::vector<char> req) {
		std::vector<char> res(8);
		auto tuple = std::make_tuple(x, res, y);

		return tuple;
	}
};

int main() {
    rpc_server server(9123, std::thread::hardware_concurrency() * 3);

	dummy d;
	//server.register_handler("add", &dummy::add, &d);
	server.register_handler("tuple", &dummy::rpc_test_tuple, &d);
	server.run();
	
	return 0;
}

result

$ ./client 
read cost 48257
$ ./client 
read cost 50969
$ ./client 
read cost 3021651
$ ./client 
read cost 3020548
$ ./client 
read cost 3030853
$ ./client 
read cost 3034473
$ ./client 
read cost 3031834
$ ./client 
read cost 3039597
$ ./client 
read cost 48282
$ ./client 
read cost 51913
$ ./client 
read cost 3038084

请问这是什么原因，是因为哪里的细节没有注意到吗

[qicosmos]

你这测试方法不对呀，你用100多个线程去测试，线程切换的开销会很大，不是已经有一个benchamrk client的例子了吗？
https://github.com/qicosmos/rest_rpc/blob/master/examples/client/main.cpp#L616
用client的异步call，启动这样的几个client就够了。

[qicosmos]

同步client call用起来方便，但是效率没有client call效率高，不过平时开发的时候同步call一般情况下性能也够用，性能测试下用异步call可以让client请求的效率很高，性能测试数据才会比较准确。

[Cai-Yao]

嗯好，我使用benchamrk client的例子试试