A sample HTTP Server implemented using libevent.

This repository also contains several versions of a HTTP Server implemented in cpp.

To run the server, use the following command, default host is 14396:

bazel run -c opt src/main/cpp:server

The base class for server is ServerSocket at src/main/cpp/server_socket.h. I use an evconnlistener from libevent to bind and listen on the server address. For each new connection , accept_conn_cb is called. The server passes the connection to one of the workers to handle (defined in src/main/cpp/server).

The server worker runs an event loop, connections received from the server will be added to the loop for monitoring. When the connection becomes active, conn_read_callback will be triggered. It reads from the socket, parses the HTTP request, call HandleHttpRequest to produce an HttpResponse and write the response to the connection socket.

Some notes:

• Connections received from evconnlistener are non-blocking by default. When reading from non-blocking fd, if the buffer is not ready, error EAGAIN or EWOULDBLOCK would be returned. We should not close the connection there, but simply skip it, continue to proceed other connections.
• When the event loop triggers conn_read_callback, the event associated with the connection becomes non-pending and will not trigger callbacks anymore, so in the callback function, we need to add the connection again after we're done processing it.
• We run the event loop in one thread, when there's no event, and then adding connection to it from another thread, so we need to set flag EVLOOP_NO_EXIT_ON_EMPTY.

HTTP Server with N workers. Connections are monitored by libevent, active connections are passed to workers for handling.

user~/wrk (master) ./wrk -t8 -c10000 -d20s http://host:14396
Running 20s test @ http://host:14396
  8 threads and 10000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     3.39ms   16.09ms   1.68s    98.65%
    Req/Sec    37.57k    11.24k  107.14k    71.88%
  5897359 requests in 20.08s, 224.97MB read
  Socket errors: connect 8987, read 0, write 0, timeout 0
Requests/sec: 293685.43
Transfer/sec:     11.20MB

HTTP Server with N workers. Connections are distributed to the workers, each work handles IO using EPOLL with non-blocking IOs.

user~/wrk (master) ./wrk -t8 -c 10000 -d20s http://host:14396
Running 20s test @ http://host:14396
  8 threads and 10000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     2.17ms    3.95ms 240.01ms   89.26%
    Req/Sec    48.79k     9.00k  108.25k    83.06%
  7711554 requests in 20.09s, 294.17MB read
  Socket errors: connect 8987, read 0, write 0, timeout 0
Requests/sec: 383928.67
Transfer/sec:     14.65MB

HTTP Server with N workers. Connections are distributed to the workers, each work handles IO using EPOLL.

user~/wrk (master) ./wrk -t8 -c 10000 -d20s http://host:14396
Running 20s test @ http://host:14396
  8 threads and 10000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    31.02us   37.16us  15.38ms   99.95%
    Req/Sec    74.05k    26.72k  103.45k    59.33%
  4420272 requests in 20.03s, 168.62MB read
  Socket errors: connect 8987, read 0, write 0, timeout 0
Requests/sec: 220632.74
Transfer/sec:      8.42MB

HTTP Server with a multi-threaded backend. The backend employs a ThreadPool, which manages N worker threads. Each request is encapsulated in a Task object and passed to one of the worker threads for handling.

Benchmark with WRK:

user@another_host:~/wrk$ ./wrk -t16 -c10000 -d20s http://host:14396
Running 20s test @ http://host:14396
  16 threads and 10000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   142.03us  462.59us 205.24ms   99.97%
    Req/Sec    20.89k     2.72k   24.44k    70.00%
  415862 requests in 20.09s, 15.86MB read
  Socket errors: connect 19, read 15, write 0, timeout 0
Requests/sec:  20699.89
Transfer/sec:    808.59KB