Re: upstream keepalive - call for testing - Matthieu Tourne - ru.sysoev.nginx

75 messages in ru.sysoev.nginxRe: upstream keepalive - call for tes...

From	Sent On	Attachments
Maxim Dounin	Aug 1, 2011 9:07 am
liseen	Aug 2, 2011 6:36 am
António P. P. Almeida	Aug 2, 2011 8:24 am
Maxim Dounin	Aug 2, 2011 10:32 am
Maxim Dounin	Aug 2, 2011 10:35 am
David Yu	Aug 2, 2011 10:41 am
Maxim Dounin	Aug 2, 2011 10:49 am
David Yu	Aug 2, 2011 10:52 am
Maxim Dounin	Aug 2, 2011 11:46 am
David Yu	Aug 2, 2011 12:09 pm
卫越	Aug 2, 2011 7:48 pm
liseen	Aug 2, 2011 8:56 pm
SplitIce	Aug 2, 2011 10:20 pm
Maxim Dounin	Aug 3, 2011 12:37 am
Charles Chen	Aug 3, 2011 2:18 am
Matthieu Tourne	Aug 3, 2011 5:06 pm
Maxim Dounin	Aug 3, 2011 11:51 pm
SplitIce	Aug 7, 2011 9:43 pm
Maxim Dounin	Aug 8, 2011 2:21 am
SplitIce	Aug 8, 2011 2:34 am
SplitIce	Aug 8, 2011 2:35 am
Matthieu Tourne	Aug 12, 2011 12:32 pm	.patch, .patch
Maxim Dounin	Aug 12, 2011 12:59 pm
Matthieu Tourne	Aug 12, 2011 2:11 pm
Maxim Dounin	Aug 12, 2011 3:26 pm
Matthieu Tourne	Aug 12, 2011 3:41 pm	.patch
Matthieu Tourne	Aug 16, 2011 4:29 pm
Maxim Dounin	Aug 16, 2011 5:21 pm
magicbear	Aug 24, 2011 10:11 am
Maxim Dounin	Aug 24, 2011 5:04 pm	.txt, .txt
Shaun savage	Aug 24, 2011 6:16 pm
magicbear	Aug 24, 2011 10:30 pm
magicbear	Aug 26, 2011 12:07 am
Maxim Dounin	Aug 26, 2011 2:38 am
magicbear	Aug 26, 2011 4:00 am
magicbear	Aug 26, 2011 4:04 am
magicbear	Aug 26, 2011 4:27 am
Maxim Dounin	Aug 26, 2011 4:36 am
magicbear	Aug 26, 2011 4:53 am
Maxim Dounin	Aug 26, 2011 8:54 am
magicbear	Aug 26, 2011 9:16 am
magicbear	Aug 26, 2011 9:27 am
magicbear	Aug 26, 2011 10:00 am
magicbear	Aug 26, 2011 10:51 am
Maxim Dounin	Aug 26, 2011 11:05 am
magicbear	Aug 26, 2011 12:00 pm
magicbear	Aug 28, 2011 10:06 am
magicbear	Aug 28, 2011 10:10 am
Maxim Dounin	Aug 28, 2011 6:46 pm	.txt
magicbear	Aug 31, 2011 1:04 pm
SplitIce	Aug 31, 2011 6:56 pm
magicbear	Sep 1, 2011 6:37 am
magicbear	Sep 4, 2011 10:33 am
Maxim Dounin	Sep 4, 2011 11:20 am
MagicBear	Sep 4, 2011 11:31 am
Maxim Dounin	Sep 5, 2011 12:07 am
ビリビリⅤ	Sep 5, 2011 8:41 am
Maxim Dounin	Sep 5, 2011 11:01 am
magicbear	Sep 5, 2011 11:39 pm
Matthieu Tourne	Sep 7, 2011 4:33 pm
Maxim Dounin	Sep 8, 2011 2:26 am
Maxim Dounin	Sep 8, 2011 8:41 am
Matthieu Tourne	Sep 8, 2011 3:04 pm
magicbear	Sep 14, 2011 3:53 pm
MagicBear	Sep 15, 2011 10:50 am	.txt
SplitIce	Sep 15, 2011 6:41 pm
philipp	Dec 29, 2011 4:46 am
Maxim Dounin	Dec 29, 2011 7:03 am
alexscott	Mar 8, 2012 6:29 am
Andrew Alexeev	Mar 8, 2012 10:17 pm
alexscott	Mar 12, 2012 7:34 am
Maxim Dounin	Mar 12, 2012 7:53 am
alexscott	Mar 12, 2012 10:39 am
Maxim Dounin	Mar 12, 2012 10:58 am
alexscott	Mar 12, 2012 12:55 pm

Subject:	Re: upstream keepalive - call for testing
From:	Matthieu Tourne (matt...@gmail.com)
Date:	Aug 16, 2011 4:29:29 pm
List:	ru.sysoev.nginx

I'm still a little confused,

Is the peer selection algorithm guaranteed to never run at the same time for different workers? (i-e: creating race conditions in the keepalive queue) I see that the round robin code has a bunch of mutex locks all commented out.. On the other hand nginx_http_upstream_check_module (healthchecks) uses mutexes.

For faster lookups, I was thinking about a hashmap of queues, hashed on sockaddr. This is probably overkill for a small amount of keepalives connections though. I'll send a patch if I get around to implement it.

Matthieu.

Thanks for the help Maxim, I'll submit this code if I get around implementing it.

Also, I think I used the wrong string comparison function in the patch I sent earlier. This one should work as intended in the description ..

On Fri, Aug 12, 2011 at 3:27 PM, Maxim Dounin <mdou...@mdounin.ru> wrote:

Hello!

On Fri, Aug 12, 2011 at 02:11:51PM -0700, Matthieu Tourne wrote:

Also, if I was planning on having a lot of different connections using the upstream keepalive module. Would it make sense to convert the queues into rbtrees for faster lookup

Yes, it may make sense if you are planning to keep lots of connections to lots of different backends (you'll still need queues though, but that's details).

Maxim Dounin

Thank you!

Matthieu.

On Fri, Aug 12, 2011 at 12:59 PM, Maxim Dounin <mdou...@mdounin.ru> wrote:

Hello!

On Fri, Aug 12, 2011 at 12:32:26PM -0700, Matthieu Tourne wrote:

Hi all,

I think I have found a small issue, if we're using proxy_pass to

talk to

origin that doesn't support keep alives. The origin will return a HTTP header "Connection: close", and

terminate

the

connection (TCP FIN). We don't take this into account, and assume there is a keep-alive connection available. The next time the connection is used, it won't be part of a valid TCP stream, and the origin server will send a TCP RST.

Yes, I'm aware of this, thank you. Actually, this is harmless: upstream keepalive module should detect connection was closed while keeping it, and even if it wasn't able to do so - nginx will re-try sending request if sending to cached connection fails.

This can be simulated with 2 nginx instances, one acting as a

reverse

proxy

with keep alive connection. And the other using the directive keepalive_timeout 0; (which will always terminate connections right away).

The patches attached take into account the response of the origin, and should fix this issue.

I'm planing to add similar patch, thanks.

Maxim Dounin

Matthieu.

On Mon, Aug 8, 2011 at 2:36 AM, SplitIce <mat...@gmail.com> wrote:

Oh and I havent been able to reproduce the crash, I tried for a

while

but

gave up. if it happens again ill build with debugging and restart howeaver so far its been 36hours without issues (under a significant amount of traffic)

On Mon, Aug 8, 2011 at 7:35 PM, SplitIce <mat...@gmail.com> wrote:

50ms per HTTP request (taken from firebug and chrome resource

panel)

the time it takes the html to load from request to arrival. 200ms is the time saved by when the http starts transfering to me (allowing other resources to begin downloading before the HTML completes), previously the html only started transfering after the full

request

was

downloaded to the proxy server (due to buffering)

HTTP to talk to the backends (between countries)

The node has a 30-80ms ping time between the backend and frontend. (Russia->Germany, Sweden->NL, Ukraine->Germany/NL etc)

On Mon, Aug 8, 2011 at 7:22 PM, Maxim Dounin <mdou...@mdounin.ru

wrote:

Hello!

On Mon, Aug 08, 2011 at 02:44:12PM +1000, SplitIce wrote:

Been testing this on my servers now for 2 days now, handling approximately 100mbit of constant traffic (3x20mbit, 1x40mbit).

Havent noticed any large bugs, had an initial crash on one of the servers however havent been able to replicate. The servers are

mixture of

openvz, XEN and one vmware virtualised containers running

debian

lenny

squeeze,

By "crash" you mean nginx segfault? If yes, it would be great to track it down (either to fix problem in keepalive patch or to prove it's unrelated problem).

Speed increases from this module are decent, approximately

50ms

from

the

request time and the HTTP download starts 200ms earler

resulting in

150ms

quicker load time on average.

Sounds cool, but I don't really understand what "50ms from the request time" and "download starts 200ms earler" actually means. Could you please elaborate?

And, BTW, do you use proxy or fastcgi to talk to backends?

Maxim Dounin

all in all, seems good.

Thanks for all your hard work Maxim.

On Thu, Aug 4, 2011 at 4:51 PM, Maxim Dounin <

mdou...@mdounin.ru>

wrote:

Hello!

On Wed, Aug 03, 2011 at 05:06:56PM -0700, Matthieu Tourne wrote:

Hi,

I'm trying to use keepalive http connections for

proxy_pass

directives

containing variables. Currently it only works for named upstream blocks.

I'm wondering what would be the easiest way, maybe setting peer->get to

ngx_http_upstream_get_keepalive_peer

and

kp->original_get_peer to

ngx_http_upstream_get_round_robin_peer()

towards

the end of ngx_http_create_round_robin_peer(). If I can figure how to set kp->conf to something sane this

might

work :)

Thoughts ?

You may try to pick one from upstream's main conf upstreams array (e.g. one from first found upstream with init set to ngx_http_upstream_init_keepalive). Dirty, but should work.

Maxim Dounin

Thank you, Matthieu.

On Tue, Aug 2, 2011 at 10:21 PM, SplitIce <

mat...@gmail.com>

wrote:

Ive been testing this on my localhost and one of my live

servers

(http

backend) for a good week now, I haven't had any issues

that I

have

noticed

as of yet.

Servers are Debian Lenny and Debian Squeeze (oldstable, stable)

Hoping it will make it into the developer (1.1.x) branch

soon

On Wed, Aug 3, 2011 at 1:57 PM, liseen <

lise...@gmail.com

wrote:

Could nginx keepalive work with HealthCheck? Does

Maxim

Dounin

have

support plan?

On Wed, Aug 3, 2011 at 3:09 AM, David Yu <

davi...@gmail.com>

wrote:

On Wed, Aug 3, 2011 at 2:47 AM, Maxim Dounin <

mdou...@mdounin.ru>

wrote:

Hello!

On Wed, Aug 03, 2011 at 01:53:30AM +0800, David Yu wrote:

On Wed, Aug 3, 2011 at 1:50 AM, Maxim Dounin <

mdou...@mdounin.ru>

wrote:

Hello!

On Wed, Aug 03, 2011 at 01:42:13AM +0800, David

wrote:

On Wed, Aug 3, 2011 at 1:36 AM, Maxim Dounin <

mdou...@mdounin.ru>

wrote:

Hello!

On Tue, Aug 02, 2011 at 04:24:45PM +0100,

António

Almeida

wrote:

On 1 Ago 2011 17h07 WEST, mdounin@mdounin.ruwrote:

Hello!

JFYI:

Last week I posted patch to

nginx-devel@which

adds

keepalive

support to various backends (as with

upstream

keepalive

module),

including fastcgi and http backends (this

turn

means

nginx now

able to talk HTTP/1.1 to backends, in

particular

it now

understands chunked responses). Patch

applies

1.0.5

and

1.1.0.

Testing is appreciated.

You may find patch and description here:

http://mailman.nginx.org/pipermail/nginx-devel/2011-July/001057.html

Patch itself may be downloaded here:

http://nginx.org/patches/patch-nginx-keepalive-full.txt

Upstream keepalive module may be

downloaded

here:

http://mdounin.ru/hg/ngx_http_upstream_keepalive/

http://mdounin.ru/files/ngx_http_upstream_keepalive-0.4.tar.gz

So *either* we use the patch or use the

module.

Correct?

No, to keep backend connections alive you

need

module

*and*

patch.

Patch provides foundation in nginx core for

module

work

with

fastcgi and http.

With a custom nginx upstream binary protocol, I

believe

multiplexing will

now be possible?

ENOPARSE, sorry.

After some googling ... ENOPARSE is a nerdy term. It is one of the standard

library

error

codes

that can be set in the global variable "errno" and

stands

for

Error No

Parse. Since you didn't get it, I can thus conclude

that

unlike me

you

are probably

a normal, well adjusted human being ;-)

Actually, this definition isn't true: there is no

such

error

code,

it's rather imitation. The fact that author of

definition

claims

it's real error indicate that unlike me, he is

normal,

well

adjusted human being. ;)

Now I get it. Well adjusted I am.

Now you may try to finally explain what you mean to

ask in

your

original message. Please keep in mind that your are

talking

somebody far from being normal and well adjusted. ;)

Maxim Dounin

p.s. Actually, I assume you are talking about fastcgi multiplexing.

Nope not fastcgi multiplexing. Multiplexing over a custom/efficient nginx binary protocol. Where requests sent to upstream include a unique id

w/c the

upstream

will

also send on response. This allows for asychronous, out-of-bands, messaging. I believe this is what mongrel2 is trying to do now

...

though

as an

http

server, it is nowhere near as robust/stable as nginx. If nginx implements this (considering nginx already

has a

lot

market

share), it certainly would bring more developers/users

(especially

the

ones needing async, out-of-bands request handling)

Short answer is: no, it's still not possible.

-- Warez Scene <http://thewarezscene.org> Free Rapidshare

Downloads<

http://www.nexusddl.com>

-- Warez Scene <http://thewarezscene.org> Free Rapidshare Downloads<http://www.nexusddl.com>

-- Warez Scene <http://thewarezscene.org> Free Rapidshare

Downloads<

http://www.nexusddl.com>

	Start a set with this search
	Include this search in one of my sets
	Exclude this search from one of my sets