On Fri, Jul 1, 2016 at 9:56 AM, Ray Satiro via curl-library
<curl-library_at_cool.haxx.se> wrote:
> On 6/30/2016 9:11 AM, Daniel Stenberg wrote:
>>
>> Since we once again got hit by the TCP_NODELAY causing a weirdo
>> performance behavior, I move to making it enabled by default. As the
>> attached patch shows.
>>
>> Mentioned here to allow you to protest, agree or discuss bikeshed colors.
>>
>> If no objections show up, I'll merge tomorrow.
>
>
> It seems possible you could actually hurt performance in some cases and end
> up doing the opposite of what you intended. My understanding of this is
> we're essentially overriding the operating system's ability to combine
> packets for efficiency. Maybe there are some specific scenarios where all
> users could get provably better performance using NODELAY but using it for
> everything seems too far. Can you make a better case for this? I haven't
> seen enough to back this up so for now I vote -1.

FWIW, I often seen Nagle's alg disabled (nodelay) on network applications.

Also, as the trade-off to my understanding is bandwidth (of which we
have more abundance nowadays) vs speed, and since curl is unlikely to
make real small writes anyway, I'd vote for disabling it by default as
suggested.
-------------------------------------------------------------------
List admin: https://cool.haxx.se/list/listinfo/curl-library
Etiquette: https://curl.haxx.se/mail/etiquette.html
Received on 2016-07-01