April 09, 2013

Erlang, tcp sockets, and active true

If you don’t know erlang then you’re missing out. If you do know erlang, you’ve probably at some point done something with tcp sockets. Erlang’s highly concurrent model of execution lends itself well to server programs where a high number of active connections is desired. Each thread can autonomously handle its single client, greatly simplifying the logic of the whole application while still retaining great performance characteristics.

Background

For an erlang thread which owns a single socket there are three different ways to receive data off of that socket. These all revolve around the active setopts flag. A socket can be set to one of:

  • {active,false} - All data must be obtained through recv/2 calls. This amounts to syncronous socket reading.

  • {active,true} - All data on the socket gets sent to the controlling thread as a normal erlang message. It is the thread’s responsibility to keep up with the buffered data in the message queue. This amounts to asyncronous socket reading.

  • {active,once} - When set the socket is placed in {active,true} for a single packet. That is, once set the thread can expect a single message to be sent to when data comes in. To receive any more data off of the socket the socket must either be read from using recv/2 or be put in {active,once} or {active,true}.

Which to use?

Many (most?) tutorials advocate using {active,once} in your application [0][1][2]. This has to do with usability and security. When in {active,true} it’s possible for a client to flood the connection faster than the receiving process will process those messages, potentially eating up a lot of memory in the VM. However, if you want to be able to receive both tcp data messages as well as other messages from other erlang processes at the same time you can’t use {active,false}. So {active,once} is generally preferred because it deals with both of these problems quite well.

Why not to use {active,once}

Here’s what your classic {active,once} enabled tcp socket implementation will probably look like:

-module(tcp_test).
-compile(export_all).

-define(TCP_OPTS, [
    binary,
    {packet, raw},
    {nodelay,true},
    {active, false},
    {reuseaddr, true},
    {keepalive,true},
    {backlog,500}
]).

%Start listening
listen(Port) ->
    {ok, L} = gen_tcp:listen(Port, ?TCP_OPTS),
    ?MODULE:accept(L).

%Accept a connection
accept(L) ->
    {ok, Socket} = gen_tcp:accept(L),
    ?MODULE:read_loop(Socket),
    io:fwrite("Done reading, connection was closed\n"),
    ?MODULE:accept(L).

%Read everything it sends us
read_loop(Socket) ->
    inet:setopts(Socket, [{active, once}]),
    receive
    {tcp, _, _} ->
        do_stuff_here,
        ?MODULE:read_loop(Socket);
    {tcp_closed, _}-> donezo;
    {tcp_error, _, _} -> donezo
    end.

This code isn’t actually usable for a production system; it doesn’t even spawn a new process for the new socket. But that’s not the point I’m making. If I run it with tcp_test:listen(8000), and in other window do:

while [ 1 ]; do echo "aloha"; done | nc localhost 8000

We’ll be flooding the the server with data pretty well. Using eprof we can get an idea of how our code performs, and where the hang-ups are:

1> eprof:start().
{ok,<0.34.0>}

2> P = spawn(tcp_test,listen,[8000]).
<0.36.0>

3> eprof:start_profiling([P]).
profiling

4> running_the_while_loop.
running_the_while_loop

5> eprof:stop_profiling().
profiling_stopped

6> eprof:analyze(procs,[{sort,time}]).

****** Process <0.36.0>    -- 100.00 % of profiled time *** 
FUNCTION                           CALLS      %      TIME  [uS / CALLS]
--------                           -----    ---      ----  [----------]
prim_inet:type_value_2/2               6   0.00         0  [      0.00]

....snip....

prim_inet:enc_opts/2                   6   0.00         8  [      1.33]
prim_inet:setopts/2             12303599   1.85   1466319  [      0.12]
tcp_test:read_loop/1            12303598   2.22   1761775  [      0.14]
prim_inet:encode_opt_val/1      12303599   3.50   2769285  [      0.23]
prim_inet:ctl_cmd/3             12303600   4.29   3399333  [      0.28]
prim_inet:enc_opt_val/2         24607203   5.28   4184818  [      0.17]
inet:setopts/2                  12303598   5.72   4533863  [      0.37]
erlang:port_control/3           12303600  77.13  61085040  [      4.96]

eprof shows us where our process is spending the majority of its time. The % column indicates percentage of time the process spent during profiling inside any function. We can pretty clearly see that the vast majority of time was spent inside erlang:port_control/3, the BIF that inet:setopts/2 uses to switch the socket to {active,once} mode. Amongst the calls which were called on every loop, it takes up by far the most amount of time. In addition all of those other calls are also related to inet:setopts/2.

I’m gonna rewrite our little listen server to use {active,true}, and we’ll do it all again:

-module(tcp_test).
-compile(export_all).

-define(TCP_OPTS, [
    binary,
    {packet, raw},
    {nodelay,true},
    {active, false},
    {reuseaddr, true},
    {keepalive,true},
    {backlog,500}
]).

%Start listening
listen(Port) ->
    {ok, L} = gen_tcp:listen(Port, ?TCP_OPTS),
    ?MODULE:accept(L).

%Accept a connection
accept(L) ->
    {ok, Socket} = gen_tcp:accept(L),
    inet:setopts(Socket, [{active, true}]), %Well this is new
    ?MODULE:read_loop(Socket),
    io:fwrite("Done reading, connection was closed\n"),
    ?MODULE:accept(L).

%Read everything it sends us
read_loop(Socket) ->
    %inet:setopts(Socket, [{active, once}]),
    receive
    {tcp, _, _} ->
        do_stuff_here,
        ?MODULE:read_loop(Socket);
    {tcp_closed, _}-> donezo;
    {tcp_error, _, _} -> donezo
    end.

And the profiling results:

1> eprof:start().
{ok,<0.34.0>}

2>  P = spawn(tcp_test,listen,[8000]).
<0.36.0>

3> eprof:start_profiling([P]).
profiling

4>  running_the_while_loop.
running_the_while_loop

5> eprof:stop_profiling().
profiling_stopped

6> eprof:analyze(procs,[{sort,time}]).

****** Process <0.36.0>    -- 100.00 % of profiled time *** 
FUNCTION                           CALLS       %      TIME  [uS / CALLS]
--------                           -----     ---      ----  [----------]
prim_inet:enc_value_1/3                7    0.00         1  [      0.14]
prim_inet:decode_opt_val/1             1    0.00         1  [      1.00]
inet:setopts/2                         1    0.00         2  [      2.00]
prim_inet:setopts/2                    2    0.00         2  [      1.00]
prim_inet:enum_name/2                  1    0.00         2  [      2.00]
erlang:port_set_data/2                 1    0.00         2  [      2.00]
inet_db:register_socket/2              1    0.00         3  [      3.00]
prim_inet:type_value_1/3               7    0.00         3  [      0.43]

.... snip ....

prim_inet:type_opt_1/1                19    0.00         7  [      0.37]
prim_inet:enc_value/3                  7    0.00         7  [      1.00]
prim_inet:enum_val/2                   6    0.00         7  [      1.17]
prim_inet:dec_opt_val/1                7    0.00         7  [      1.00]
prim_inet:dec_value/2                  6    0.00        10  [      1.67]
prim_inet:enc_opt/1                   13    0.00        12  [      0.92]
prim_inet:type_opt/2                  19    0.00        33  [      1.74]
erlang:port_control/3                  3    0.00        59  [     19.67]
tcp_test:read_loop/1            20716370  100.00  12187488  [      0.59]

This time our process spent almost no time at all (according to eprof, 0%) fiddling with the socket opts. Instead it spent all of its time in the read_loop doing the work we actually want to be doing.

So what does this mean?

I’m by no means advocating never using {active,once}. The security concern is still a completely valid concern and one that {active,once} mitigates quite well. I’m simply pointing out that this mitigation has some fairly serious performance implications which have the potential to bite you if you’re not careful, especially in cases where a socket is going to be receiving a large amount of traffic.

Meta

These tests were done using R15B03, but I’ve done similar ones in R14 and found similar results. I have not tested R16.

  • [0] http://learnyousomeerlang.com/buckets-of-sockets
  • [1] http://www.erlang.org/doc/man/gen_tcp.html#examples
  • [2] http://erlycoder.com/25/erlang-tcp-server-tcp-client-sockets-with-gen_tcp