incubator-couchdb-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From kowsik <kow...@gmail.com>
Subject Re: CouchDB performance
Date Wed, 17 Aug 2011 05:12:18 GMT
Well, for nodelay (we set this to true on the blitz.io scale engines,
BTW) to be appreciable you need some kinda geo latency where the
buffering and the Nagle algorithm is recognizable. For local or really
fast networks, unless you are using an interactive client like SSH or
telnet (stop using that!), you are not going to see much of a
difference.

Thanks much and the blitz.io couch cluster is now running with this setting. :)

K.
---
http://blitz.io
@pcapr

On Tue, Aug 16, 2011 at 10:05 PM, Filipe David Manana
<fdmanana@apache.org> wrote:
> On Tue, Aug 16, 2011 at 9:48 PM, kowsik <kowsik@gmail.com> wrote:
>> This is awesome. Didn't know about this setting. Is there a reason why
>> we don't set nodelay to true by default?
>
> On some TCP/IP stack implementations, or networks, the gains can be
> much smaller or none. I remember trying several times in a Mac OS X 32
> bits and not seeing much different, while on Linux I saw a very
> significant difference.
>
>  It seems that even the
>> [continuous] _changes feed will do *much* better with this setting,
>> notifying the listeners immediately without any buffering. True?
>>
>> One more Q: After changing this and if I invoke _restart, will the
>> settings take into effect?
>
> As soon as you change this setting via the REST API, subsequent
> requests should benefit from nodelay=true.
>
>>
>> Thanks,
>
> yw
>
>>
>> K.
>> ---
>> http://blitz.io
>> @pcapr
>>
>> On Tue, Aug 16, 2011 at 9:36 PM, Filipe David Manana
>> <fdmanana@apache.org> wrote:
>>> On Tue, Aug 16, 2011 at 9:18 PM, John Cheng
>>>>> Have you tried setting server's socket options, in the .ini config, as
>>>>> I suggested before? (setting nodelay to true)
>>>>>
>>>>>
>>>> You know I really thought I had tried that last night without seeing a
>>>> difference. I just tried it again right now and saw that was indeed the
>>>> source of my problem. Boy do I feel embarrassed right now.
>>>>
>>>> Thank you for your patience!
>>>
>>> Np.
>>> Yes, default is false. For me, on Linux 32 bits at least, nodelay set
>>> true makes a huge difference as well.
>>>
>>>
>>>>
>>>> Just for the record, this is the setting
>>>>
>>>> socket_options = [{recbuf, 262144}, {sndbuf, 262144}, {nodelay, true}]
>>>>
>>>> In default.ini it is commented out, and I assume by default nodelay is set
>>>> to false.
>>>>
>>>
>>>
>>>
>>> --
>>> Filipe David Manana,
>>> fdmanana@gmail.com, fdmanana@apache.org
>>>
>>> "Reasonable men adapt themselves to the world.
>>>  Unreasonable men adapt the world to themselves.
>>>  That's why all progress depends on unreasonable men."
>>>
>>
>
>
>
> --
> Filipe David Manana,
> fdmanana@gmail.com, fdmanana@apache.org
>
> "Reasonable men adapt themselves to the world.
>  Unreasonable men adapt the world to themselves.
>  That's why all progress depends on unreasonable men."
>

Mime
View raw message