c++ - 高效解析FIX消息c++

Question

我需要解析一个包含金融 FIX 协议(protocol)的文件。示例如下:

1128=99=24535=X49=CME75=2017040934=82452=2017040920070508394791460=201704092007050800000005799=10000000268=2279=0269=B48=900655=ESM783=23271=1473460731=100000005796=17263279=0269=C48=900655=ESM783=24271=2861528731=100000005796=1726310=219

我的应用程序将加载许多文件，每个文件都有数百万行历史数据，因此需要考虑性能。

我已经在线查看了有关 FIX 解析的类似问题，并探索了 QuickFix 库(特别是使用 FIX::Message(string) 来破解消息)，但我的目标是获得比我能够实现的更好的吞吐量使用 quickfix。

我为最常见的消息类型(市场数据增量刷新)编写了一个模拟，以查看我所达到的速度，并且对每秒约 60,000 条消息的结果最不满意，包括文件解析一个 3m 的线文件。

这是我的第一个 C++ 应用程序，所以我预计我的方法会有很多缺陷，如果您能提出任何关于如何提高其性能的建议，我们将不胜感激。

目前流程是file->string->MDIncrementalRefresh。 MDIncrementalRefresh 有两个可选的重复组，我使用 vector 来存储它们，因为它们在消息之间的大小未知。

我猜我在每次更新时重建 MDIncrementalRefresh 与我通过更新之前的 MDIncrementalRefresh 的内容重新使用该对象相比会导致不必要的开销吗？

提前致谢

#include <string>
#include <vector>
#include <iostream>
#include <fstream>

using namespace std;

std::vector<std::string> string_split(std::string s, const char delimiter)
{
    size_t start=0;
    size_t end=s.find_first_of(delimiter);

    std::vector<std::string> output;

    while (end <= std::string::npos)
    {
        output.emplace_back(s.substr(start, end-start));

        if (end == std::string::npos)
            break;

        start=end+1;
        end = s.find_first_of(delimiter, start);
    }

    return output;
}

const char FIX_FIELD_DELIMITER = '\x01';
const char FIX_KEY_DELIMITER = '=';

const int STR_TO_CHAR = 0;
const int KEY = 0;
const int VALUE = 1;

const string Field_TransactTime = "60";
const string Field_MatchEventIndicator = "5799";
const string Field_NoMDEntries = "268";
const string Field_MDUpdateAction = "279";
const string Field_MDEntryType = "269";
const string Field_SecurityID = "48";
const string Field_RptSeq = "83";
const string Field_MDEntryPx = "270";
const string Field_MDEntrySize = "271";
const string Field_NumberOfOrders = "346";
const string Field_MDPriceLevel = "1023";
const string Field_OpenCloseSettlFlag = "286";
const string Field_AggressorSide = "5797";
const string Field_TradingReferenceDate = "5796";
const string Field_HighLimitPrice = "1149";
const string Field_LowLimitPrice = "1148";
const string Field_MaxPriceVariation = "1143";
const string Field_ApplID = "1180";
const string Field_NoOrderIDEntries = "37705";
const string Field_OrderID = "37";
const string Field_LastQty = "32";
const string Field_SettlPriceType= "731";

class OrderIdEntry {
public:
    string OrderID;
    int LastQty;
};

struct MDEntry {
public:
    // necessary for defaults?
    char MDUpdateAction;
    char MDEntryType;
    int SecurityID;
    int RptSeq;
    double MDEntryPx;
    int MDEntrySize;
    int NumberOfOrders = 0;
    int MDPriceLevel = 0;
    int OpenCloseSettlFlag = 0;
    string SettlPriceType = "";
    int AggressorSide = 0;
    string TradingReferenceDate = "";
    double HighLimitPrice = 0.0;
    double LowLimitPrice = 0.0;
    double MaxPriceVariation = 0.0;
    int ApplID = 0;

};

class MDIncrementalRefresh {

public:
    string TransactTime;
    string MatchEventIndicator;
    int NoMDEntries;
    int NoOrderIDEntries = 0;
    vector<MDEntry> MDEntries;
    vector<OrderIdEntry> OrderIdEntries;

    MDIncrementalRefresh(const string& message)
    {

        MDEntry* currentMDEntry = nullptr;
        OrderIdEntry* currentOrderIDEntry = nullptr;

        for (auto fields : string_split(message, FIX_FIELD_DELIMITER))
        {
            vector<string> kv = string_split(fields, FIX_KEY_DELIMITER);

            // Header :: MDIncrementalRefresh

            if (kv[KEY] == Field_TransactTime) this->TransactTime = kv[VALUE];

            else if (kv[KEY] == Field_MatchEventIndicator) this->MatchEventIndicator = kv[VALUE];
            else if (kv[KEY] == Field_NoMDEntries) this->NoMDEntries = stoi(kv[VALUE]);
            else if (kv[KEY] == Field_NoOrderIDEntries) this->NoOrderIDEntries = stoi(kv[VALUE]);

            // Repeating Group :: MDEntry

            else if (kv[KEY] == Field_MDUpdateAction)
            {
                MDEntries.push_back(MDEntry());
                currentMDEntry = &MDEntries.back(); // use pointer for fast lookup on subsequent repeating group fields
                currentMDEntry->MDUpdateAction = kv[VALUE][STR_TO_CHAR];
            }
            else if (kv[KEY] == Field_MDEntryType) currentMDEntry->MDEntryType = kv[VALUE][STR_TO_CHAR];
            else if (kv[KEY] == Field_SecurityID) currentMDEntry->SecurityID = stoi(kv[VALUE]);
            else if (kv[KEY] == Field_RptSeq) currentMDEntry->RptSeq = stoi(kv[VALUE]);
            else if (kv[KEY] == Field_MDEntryPx) currentMDEntry->MDEntryPx = stod(kv[VALUE]);
            else if (kv[KEY] == Field_MDEntrySize) currentMDEntry->MDEntrySize = stoi(kv[VALUE]);
            else if (kv[KEY] == Field_NumberOfOrders) currentMDEntry->NumberOfOrders = stoi(kv[VALUE]);
            else if (kv[KEY] == Field_MDPriceLevel) currentMDEntry->MDPriceLevel = stoi(kv[VALUE]);
            else if (kv[KEY] == Field_OpenCloseSettlFlag) currentMDEntry->OpenCloseSettlFlag = stoi(kv[VALUE]);
            else if (kv[KEY] == Field_SettlPriceType) currentMDEntry->SettlPriceType= kv[VALUE];
            else if (kv[KEY] == Field_AggressorSide) currentMDEntry->AggressorSide = stoi(kv[VALUE]);
            else if (kv[KEY] == Field_TradingReferenceDate) currentMDEntry->TradingReferenceDate = kv[VALUE];
            else if (kv[KEY] == Field_HighLimitPrice) currentMDEntry->HighLimitPrice = stod(kv[VALUE]);
            else if (kv[KEY] == Field_LowLimitPrice) currentMDEntry->LowLimitPrice = stod(kv[VALUE]);
            else if (kv[KEY] == Field_MaxPriceVariation) currentMDEntry->MaxPriceVariation = stod(kv[VALUE]);
            else if (kv[KEY] == Field_ApplID) currentMDEntry->ApplID = stoi(kv[VALUE]);

            // Repeating Group :: OrderIDEntry
            else if (kv[KEY] == Field_OrderID) {
                OrderIdEntries.push_back(OrderIdEntry());
                currentOrderIDEntry = &OrderIdEntries.back();
                currentOrderIDEntry->OrderID = kv[VALUE];
            }

            else if (kv[KEY] == Field_LastQty) currentOrderIDEntry->LastQty = stol(kv[VALUE]);
        }
    }


};

int main() {

    //std::string filename = "test/sample";

    std::string line;
    std::ifstream file (filename);

    int count = 0;
    if (file.is_open())
    {
        while ( std::getline( file, line ) )
        {
            MDIncrementalRefresh md(line);
            if (md.TransactTime != "") {
                count++;
            }
        }
        file.close();
    }
    cout << count << endl;
    return 0;
}

Answer 1

对于那些感兴趣的人来说，处理上面代码的大部分时间都花在了 split_string 函数上。对 split_string 的大量调用导致在堆上进行许多(昂贵的)分配。

另一种实现 split_string_optim 重新使用预分配的 vector 。这可以防止在每次 split_string 函数调用时不必要的堆分配/扩展。以下运行 1.5m 迭代的示例表明速度提高了 3.4 倍。通过利用本身不会将分配的内存释放回堆的 vector.clear()，它可以确保后续对 split_string_optim 的 split_string 调用，其中生成的 vector 大小 <= 先前没有额外分配。

#include <string>
#include <vector>

void string_split_optim(std::vector<std::string>& output, const std::string &s, const char delimiter)
{
    output.clear();

    size_t start = 0;
    size_t end = s.find_first_of(delimiter);


    while (end <= std::string::npos)
    {
        output.emplace_back(s.substr(start, end - start));

        if (end == std::string::npos)
            break;

        start = end + 1;
        end = s.find_first_of(delimiter, start);
    }

}


int main()
{
    const int NUM_RUNS = 1500000;
    const std::string s = "1128=9\u00019=174\u000135=X\u000149=CME\u000175=20170403\u000134=1061\u000152=20170402211926965794928\u000160=20170402211926965423233\u00015799=10000100\u0001268=1\u0001279=1\u0001269=1\u000148=9006\u000155=ESM7\u000183=118\u0001270=236025.0\u0001271=95\u0001346=6\u00011023=9\u000110=088\u0001";

    std::vector<std::string> vec;

    // standard
    clock_t tStart = clock();
    for (int i = 0; i < NUM_RUNS; ++i)
    {
        vec = string_split(s, '=');
    }

    printf("Time taken: %.2fs\n", (double) (clock() - tStart) / CLOCKS_PER_SEC);

    // reused vector
    tStart = clock();
    for (int i = 0; i < NUM_RUNS; ++i)
    {
        string_split_optim(vec, s, '=');
        vec.clear();
    }

    printf("Time taken: %.2fs\n", (double) (clock() - tStart) / CLOCKS_PER_SEC);
}

我的 macbook 上的结果是提高了 3.4 倍。

Time taken: 6.60s
Time taken: 1.94s

此外，MDIncrementalRefresh 对象被重复构造(在堆栈上，但它的 vector 成员也在堆上扩展)。根据上述关于 split_string 的发现，我决定重新使用临时对象并简单地清除其先前的状态，从而导致另一个显着的性能提升。