[erlang-questions] Improved IO polling in erts

[Ameretat Reith]

This is amazing!

Here is result of simple UDP echo server. It spans 1000 processes for
client and servers [1]. I ran it on Xeon E5-2620 with eight physical cores.

On baseline 20:

16:22:34.977 [notice] 1505303554: recv_pkts: 1793113 recv_size: 8965565
sent_pkts: 1793113 sent_size: 2639462336 recv: 7.172452 Mbit/s send:
2111.569869 Mbit/s
16:22:44.978 [notice] 1505303564: recv_pkts: 1801476 recv_size: 9007380
sent_pkts: 1801476 sent_size: 2651772672 recv: 7.205904 Mbit/s send:
2121.418138 Mbit/s
16:22:54.979 [notice] 1505303574: recv_pkts: 1799551 recv_size: 8997755
sent_pkts: 1799551 sent_size: 2648939072 recv: 7.198204 Mbit/s send:
2119.151258 Mbit/s

On poll-threads branch I get best results by two polling threads:

16:58:14.696 [notice] 1505305694: recv_pkts: 2835781 recv_size: 14178905
sent_pkts: 2835781 sent_size: 4174269632 <%28417%29%20426-9632> recv:
11.343124 Mbit/s send: 3339.415706 Mbit/s
16:58:24.698 [notice] 1505305704: recv_pkts: 2836201 recv_size: 14181005
sent_pkts: 2836201 sent_size: 4174887872 <%28417%29%20488-7872> recv:
11.344804 Mbit/s send: 3339.910298 Mbit/s
16:58:34.699 [notice] 1505305714: recv_pkts: 2833905 recv_size: 14169525
sent_pkts: 2833905 sent_size: 4171508160 recv: 11.335620 Mbit/s send:
3337.206528 Mbit/s

A %50 throughput increase for this small system! I expect more enhancements
in bigger systems; those involving more message passings or using
encryption.

1: https://github.com/reith/udpstress
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20170913/cae0b7f8/attachment.htm>