[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [sc-users] TCP default?



that would be great. High density of small messages seems an important parameter (think granular synthesis, > 1000 messages/sec). Maybe you can add a test for broadcasting, too.


> On 20.12.2016, at 16:37, amindfv@xxxxxxxxx wrote:
> 
> Thanks for the first batch of numbers, Glen!
> 
> I see very similar results testing TCP vs UDP in general (ie a ~0-3% time increase for TCP). I'm working on some simple SC tests too.
> 
> tom
> 
> 
> El 20 dic 2016, a las 08:11, Glen Fraser <holaglen@xxxxxxxxx> escribió:
> 
>> Hmm, not sure why the test results didn’t show up in my mail (they do appear in the Nabble forum)… For the record, they should have read:
>> 
>> ./tcp_lat
>>  100 10000
>> message size: 100 octets
>> roundtrip count: 10000
>> average latency: 17770 ns
>> ./
>> udp_lat
>>  100 10000
>> message size: 100 octets
>> roundtrip count: 10000
>> average latency: 17528 ns
>> 
>> 
>> 
>>> On 20 Dec 2016, at 15:07, Glen Fraser <holaglen@xxxxxxxxx> wrote:
>>> 
>>> Julian Rohrhuber-3 wrote
>>>> good point. Careful benchmarking really might be necessary to decide this,
>>>> see e.g.
>>>> http://stackoverflow.com/questions/47903/udp-vs-tcp-how-much-faster-is-it
>>> 
>>> For fun, I looked for some latency benchmarking tests.  I came across  this
>>> <https://github.com/rigtorp/ipc-bench>  .  He includes no UDP test, so I
>>> added one (submitted as a pull request to his repo; in the meantime you can
>>> see my forked version  here <https://github.com/totalgee/ipc-bench>  ).  A
>>> quick test of the TCP and (now) UDP latency tests, give results (on my
>>> MacBook Air) such as:
>>> 
>>> 
>>> 
>>> On my machine, the UDP/TCP latencies tend to be very similar, for a given
>>> message size and iteration count.
>>> 
>>> Note that for the above test, the UNIX pipe latency was a mere 5400 ns
>>> whereas UNIX domain sockets achieved about 7300 ns, about 2-3 times faster
>>> than using localhost (TCP or UDP) network sockets.  Still, 18 microseconds
>>> isn't too bad either...
>>> 
>>> Also note that it would be useful to look at the min and max latencies, not
>>> just average.  If you only send one small packet, you'll see that the
>>> latency can be quite large, around a couple of hundred microseconds.  The
>>> average latencies go up when you do fewer roundtrips in a burst (i.e. this
>>> test is not exactly representative of how SC would be sending data).
>>> 
>>> Still, it might be useful, especially to see that (on my machine at least)
>>> the results look quite similar for TCP and UDP.
>>> 
>>> Glen.
>>> 
>>> 
>>> 
>>> --
>>> View this message in context: http://new-supercollider-mailing-lists-forums-use-these.2681727.n2.nabble.com/TCP-default-tp7629613p7629627.html
>>> Sent from the SuperCollider Users New (Use this!!!!) mailing list archive at Nabble.com.
>>> 
>>> _______________________________________________
>>> sc-users mailing list
>>> 
>>> info (subscription, etc.): http://www.birmingham.ac.uk/facilities/ea-studios/research/supercollider/mailinglist.aspx
>>> archive: https://listarc.bham.ac.uk/marchives/sc-users/
>>> search: https://listarc.bham.ac.uk/lists/sc-users/search/
>> 

Attachment: signature.asc
Description: Message signed with OpenPGP using GPGMail