从 Beast.1.0.0-b66 到 Boost.1.67.0.Beast 的端口性能下降

Question

我观察到从 Beast.1.0.0-b66（使用 Boost.1.64.0）迁移到 Boost.1.67.0.Beast（即集成了 Beast）后性能急剧下降（CPU 消耗急剧增加）进入升压）。毫无疑问我做错了什么，但我无法想象是什么。

什么是：

typedef beast::http::request<beast::http::string_body> BeastHttpRequest;

现在：

namespace http = boost::beast::http;    
typedef http::request<http::string_body> BeastHttpRequest;

什么是：

beast::http::prepare(req);
beast::http::write(stream, req);

现在：

req.prepare_payload();
http::write(stream, req);

当然，我还必须进行大量 API 更改。例如：

req.fields.replace(hdrName, hdrValue);

现在：

req.set(hdrName, hdrValue);

应用程序工作正常——包括 SSL 握手和代理协商——但我必须解决 CPU 消耗的峰值和相应的性能下降问题。我想知道是否有人知道我忽略了一些明显的事情。

编辑： 我应该提到我正在为 SSL 流使用 flat_buffer。

我有机会分析“移植前”和“移植后”的性能数据。这是“端口前”（即良好的性能）调用链：

- 34.61% HttpRequest::send
   - 32.02% beast::http::write<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >&>, true, beast::http::string_body,
      - 31.90% beast::http::write<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >&>, true, beast::http::string_bo
         - 20.77% beast::http::detail::write_preparation<true, beast::http::string_body, beast::http::basic_fields<std::allocator<char> > >::init
            - 18.25% beast::http::detail::write_fields<beast::basic_streambuf<std::allocator<char> >, beast::http::basic_fields<std::allocator<char> > >
               - 10.38% beast::write<beast::basic_streambuf<std::allocator<char> >, boost::basic_string_ref<char, std::char_traits<char> > >
                  - 10.20% beast::detail::write_dynabuf<beast::basic_streambuf<std::allocator<char> >, boost::basic_string_ref<char, std::char_traits<char> > >
                     + 3.66% beast::basic_streambuf<std::allocator<char> >::prepare
                     + 2.90% beast::basic_streambuf<std::allocator<char> >::commit
                     + 1.81% boost::lexical_cast<std::string, boost::basic_string_ref<char, std::char_traits<char> > >
                     + 1.47% boost::asio::buffer_copy<beast::basic_streambuf<std::allocator<char> >::mutable_buffers_type>
               - 7.51% beast::write<beast::basic_streambuf<std::allocator<char> >, char [3]>
                  - 7.47% beast::detail::write_dynabuf<beast::basic_streambuf<std::allocator<char> >, 3ul>
                     + 3.14% beast::basic_streambuf<std::allocator<char> >::prepare
                     + 2.84% beast::basic_streambuf<std::allocator<char> >::commit
                     + 1.32% boost::asio::buffer_copy<beast::basic_streambuf<std::allocator<char> >::mutable_buffers_type>
            + 2.04% beast::http::detail::write_start_line<beast::basic_streambuf<std::allocator<char> >, beast::http::basic_fields<std::allocator<char> > >
         - 9.60% beast::http::string_body::writer::write<beast::http::detail::writef0_lambda<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_serv
            - 9.57% beast::http::detail::writef0_lambda<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >&>, beast:
               - 9.54% boost::asio::write<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >&>, beast::detail::buffe
                  - 9.48% boost::asio::write<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >&>, beast::detail::bu
                     - 8.55% boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >&>::write_some<boost::asio::detail::c
                        - 7.56% boost::asio::ssl::detail::io<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >, boost::asio::ssl::detail::wr
                           - 6.98% boost::asio::write<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >, boost::asio::mutable_buffers_1>
                              - 6.93% boost::asio::write<boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >, boost::asio::mutable_buffers_1,
                                 - 6.69% boost::asio::basic_stream_socket<boost::asio::ip::tcp, boost::asio::stream_socket_service<boost::asio::ip::tcp> >::write_some<boost::asio::detail::consuming_buffer
                                    - 6.69% boost::asio::stream_socket_service<boost::asio::ip::tcp>::send<boost::asio::detail::consuming_buffers<boost::asio::const_buffer, boost::asio::mutable_buffers_1>
                                       - 6.67% boost::asio::detail::reactive_socket_service_base::send<boost::asio::detail::consuming_buffers<boost::asio::const_buffer, boost::asio::mutable_buffers_1> >
                                          - 6.51% boost::asio::detail::socket_ops::sync_send
                                             - 6.32% 0xec6d
                                                - 6.08% system_call_fastpath
                                                   - 6.07% sys_sendmsg
                                                      - 6.06% __sys_sendmsg
                                                         - 5.94% ___sys_sendmsg
                                                            + 5.84% sock_send

这是“后端口”（即性能不佳）调用链：

- 53.77% HttpRequest::send
   - 53.32% boost::beast::http::write<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp>&>, true, boost::beast::http::basic_string_body<char, std::char_traits<char>, std::allo
      - 53.30% boost::beast::http::write<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp>&>, true, boost::beast::http::basic_string_body<char, std::char_traits<char>, std::a
         - 53.14% boost::beast::http::write<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp>&>, true, boost::beast::http::basic_string_body<char, std::char_traits<char>, std
            - 53.03% boost::beast::http::write_some<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp>&>, true, boost::beast::http::basic_string_body<char, std::char_traits<ch
               - 52.95% boost::beast::http::detail::write_some_impl<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp>&>, true, boost::beast::http::basic_string_body<char, std
                  - 36.58% boost::beast::http::serializer<true, boost::beast::http::basic_string_body<char, std::char_traits<char>, std::allocator<char> >, boost::beast::http::basic_fields<std::allocator<
                     - 35.28% boost::beast::http::serializer<true, boost::beast::http::basic_string_body<char, std::char_traits<char>, std::allocator<char> >, boost::beast::http::basic_fields<std::allocat
                        - 25.93% boost::beast::http::detail::write_some_lambda<boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp>&> >::operator()<boost::beast::detail::buffers
                           - 25.82% boost::asio::ssl::stream<boost::asio::basic_stream_socket<boost::asio::ip::tcp>&>::write_some<boost::beast::detail::buffers_ref<boost::beast::buffers_prefix_view<boost:
                              - 25.69% boost::asio::ssl::detail::io<boost::asio::basic_stream_socket<boost::asio::ip::tcp>, boost::asio::ssl::detail::write_op<boost::beast::detail::buffers_ref<boost::beas
                                 - 22.20% boost::asio::write<boost::asio::basic_stream_socket<boost::asio::ip::tcp>, boost::asio::mutable_buffer>
                                    - 22.13% boost::asio::write<boost::asio::basic_stream_socket<boost::asio::ip::tcp>, boost::asio::mutable_buffer, boost::asio::detail::transfer_all_t>
                                       - 21.96% boost::asio::detail::write_buffer_sequence<boost::asio::basic_stream_socket<boost::asio::ip::tcp>, boost::asio::mutable_buffer, boost::asio::mutable_buffer
                                          - 21.47% boost::asio::basic_stream_socket<boost::asio::ip::tcp>::write_some<boost::asio::const_buffers_1>
                                             - 21.34% boost::asio::detail::reactive_socket_service_base::send<boost::asio::const_buffers_1>
                                                - 21.17% boost::asio::detail::socket_ops::sync_send
                                                   - 20.45% 0xec6d
                                                      - 19.63% system_call_fastpath
                                                         - 19.57% sys_sendmsg
                                                            - 19.52% __sys_sendmsg
                                                               - 19.16% ___sys_sendmsg
                                                                  + 18.90% sock_sendmsg

Answer 1

它应该更快，而不是更慢，因为 HTTP 算法已经过优化。我想我知道发生了什么。如果您可以在使用常规套接字而不是 SSL 时比较两个版本的性能，那将会很有帮助。 boost::asio::ssl::stream 有一个缺点，即当写入长度大于 1 的缓冲区序列时，它会为序列中的每个缓冲区写入套接字，而不是将这些加密缓冲区合并为单个写入。这可能会对性能产生重大影响。

这确实需要在 Boost.Asio 中修复，但一种解决方法是您可以编写自己的流包装器，其写入算法在出现长度大于 1 的缓冲区序列时，通过使用创建一个长度为 1 的新序列memcpy 和动态分配。我也会在这里打开一个问题：https://github.com/boostorg/asio/issues

为什么两个版本的 Beast 有区别？旧版本的 Beast 在序列化期间分配内存以保存消息的线性版本。当前版本使用一种根本不需要内存分配的算法。在正常情况下它会更快，但正如您在使用 ssl::stream 类型时发现的那样，它可能会更慢。

见https://github.com/boostorg/beast/issues/1108