[erlang-questions] UDP receive performance

[Valentin Micic]

On 20 Jun 2018, at 7:06 PM, Lukas Larsson wrote:

> On Wed, Jun 20, 2018 at 1:56 PM Jonas Falkevik <jonas.falkevik@REDACTED> wrote:
>  
> Are you adding the multicast network to the loop back interface? Using some other interface that does not allow multicast traffic?
> 
> 
> I though I had, but apparently not.... I managed to reproduce the test case now.
> 
> I also spent some time staring at the code in the inet_driver and realized what the problem was in there. It would seem that a performance/feature fix for SCTP in R15B inadvertently introduced this behaviour for UDP :( I've pushed a bug fix to that problem in the inet_driver to the same branch. In my tests the scheduler utilization goes form 71% to about 4% with both patches. With the fix in the inet driver, my allocator patch does not really do any difference for this specific testcase, but I'll keep that fix anyways as it is a "good thing". I should probably also add a benchmark for this so that it does not happen again....
> 
> I'd also like to add that changing the test-case to use a smaller user-space buffer also has the same effect. So if anyone is running a performance critical UDP server that has set the "recbuf" to a high value, I really recommend that you lower the "buffer" size to something close to the max expected packet size. In the example that Danil provided I applied the following patch:
> 
> diff --git a/udptest.erl b/udptest.erl
> index 16a1798..4edeef0 100644
> --- a/udptest.erl
> +++ b/udptest.erl
> @@ -33,7 +33,9 @@ send_packet(ID, S, Addr, Port) ->
>  
>  start_reader(Addr, Port) ->
>    GwIP = {0,0,0,0}, % {127,0,0,1},
> -  Common = [binary,{reuseaddr,true},{recbuf,2*1024*1024},inet,{read_packets,100},{active,500}],
> +  Common = [binary,{reuseaddr,true},
> +            {buffer,1500}, %% 1500 is just an example value, don't just copy this. You need to know what you max UDP packet size will be.
> +            {recbuf,2*1024*1024},inet,{read_packets,100},{active,500}],
>    Multicast = [{multicast_ttl,4},{multicast_loop,true},{ip,Addr},{add_membership,{Addr,GwIP}}],
>    Options = Common ++ Multicast,
>    spawn(fun() -> run_reader(Port, Options) end).
> 
> and the scheduler utilization dropped to about 4% there as well.
> 
> Lukas
> 
> _______________________________________________
> erlang-questions mailing list
> erlang-questions@REDACTED
> http://erlang.org/mailman/listinfo/erlang-questions

Juts one observation about inet:setopts/2  which may not be present in R21, but certainly was in earlier releases.

When setting the "user buffer", the order of attributes specification seems to matter, e.g.

inet:setopts( S, [{buffer, 512}, {recbuf, 16#1ffff}, {sndbuf, 16#1ffff}]).  

will set the buffer to 512, whilst:

inet:setopts( S, [{recbuf, 16#1ffff}, {sndbuf, 16#1ffff}, {buffer, 512}]). 

will leave it as 16#1ffff! (e.g. set to recbuf size).

Kind regards

V/

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20180625/1cb497a4/attachment.htm>