Skip to contentSkip to conversation

urgent question about tcng!

LARTC Thomas Mandl 4 participants 9 messages Apr 27, 2005 – May 4, 2005

lore
Share & tools
lore
#0 lore raw

Hello List,
I'm new to QoS/tcng/HTB and friends, so please forgive me if my question might be silly...
After having read lots of HowTo documents I'm totally confused...

The Challenge:
=======
I'll have to deploy several "mirror" download servers (Linux) which must be able to handle a huge number of HTTP download requests (about 10k to 20k unicast requests per server within 120 min.) which will hit my servers . The mirror servers will be connected with 100MBit to the Internet backbone. Each "client" will download between 7MB up to 30MB per request. The download mirror servers will be directly attached to the Internet (no FW or router between, which could handle QoS for me). Of course will these download mirrors be specially hardened :-)

The Idea:
====I can not handle all the traffic at once without QoS, so I thought it should be possible to establish a traffic control mechanism to guarantee bandwith for each client download (e.g. 256kbps CIR up to 512kbps ceil). This should allow me, to "serve" all clients within 120 min, before a new download wave will hit my servers. If the available bandwith for HTTP traffic is used, clients shall not be able to start a download, until some bandwith is freed by another client who finished a download.

Note: I would rather go with "tcng" than with the old "tc".

The Problem:
======
While playing with tcng and HTB I managed to shape traffic (to limit the total bandwith for HTTP traffic to e.g. 50Mbps) but I did not manage to write a tcng configuration which guarantees a certain bandwith per client request. I must be able to assign (and guarantee) a defined bandwith to a client request, to make the downloads more predictable. My experiments so far only allowed me to limit e.g. HTTP traffic, which then was equally distributed to the number of clients, but this led to the usual "traffic jam" when too many clients initiated a simultanes download.

Can anybody please assist me, or does anybody have a working "tcng" config file to do the job. Additional readings/comments/links (beside those listed in the LDB/Linux Traffic Control Howtos are also welcome.)

Many thanks in advance and
best regards from Austria

Thomas

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#1 lore raw
Quoted 2 lines · Thomas Mandl

Hello List,
I'm new to QoS/tcng/HTB and friends, so please forgive me if my question

might be silly...

Quoted 6 lines

After having read lots of HowTo documents I'm totally confused...

The Challenge:
=======
I'll have to deploy several "mirror" download servers (Linux) which must be able to handle a huge number of HTTP download requests (about 10k to 20k

unicast requests per server within 120 min.) which will hit my servers . The

Quoted 1 line

mirror servers will be connected with 100MBit to the Internet backbone.

Each

Quoted 1 line

"client" will download between 7MB up to 30MB per request. The download

mirror servers will be directly attached to the Internet (no FW or router between, which could handle QoS for me). Of course will these download mirrors be specially hardened :-)

Quoted 3 lines

The Idea:
====> I can not handle all the traffic at once without QoS, so I thought it

should

Quoted 1 line

be possible to establish a traffic control mechanism to guarantee

bandwith for each client download (e.g. 256kbps CIR up to 512kbps ceil). This should

Quoted 2 lines

allow me, to "serve" all clients within 120 min, before a new download wave will hit my servers. If the available bandwith for HTTP traffic is used,

clients shall not be able to start a download, until some bandwith is freed

Quoted 7 lines

by another client who finished a download.

Note: I would rather go with "tcng" than with the old "tc".

The Problem:
======
While playing with tcng and HTB I managed to shape traffic (to limit the

total bandwith for HTTP traffic to e.g. 50Mbps) but I did not manage to write a tcng configuration which guarantees a certain bandwith per client request. I must be able to assign (and guarantee) a defined bandwith to a client request, to make the downloads more predictable. My experiments so far only allowed me to limit e.g. HTTP traffic, which then was equally distributed to the number of clients, but this led to the usual "traffic jam" when too many clients initiated a simultanes download.

Quoted 2 lines

Can anybody please assist me, or does anybody have a working "tcng"

config file to do the job. Additional readings/comments/links (beside those listed

Quoted 7 lines

in the LDB/Linux Traffic Control Howtos are also welcome.)

Many thanks in advance and
best regards from Austria

Thomas

I think you might want to check Apache's mod_bandwidth / mod_throttle instead of QoS.

Regards,

Sylvain

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#3 lore raw
In reply to Thomas Mandl
Quoted 38 lines · Thomas Mandl

Hello List,
I'm new to QoS/tcng/HTB and friends, so please forgive me if my question might be silly...
After having read lots of HowTo documents I'm totally confused...

The Challenge:
=======
I'll have to deploy several "mirror" download servers (Linux) which must be able to handle a huge number of HTTP download requests (about 10k to 20k unicast requests per server within 120 min.) which will hit my servers . The mirror servers will be connected with 100MBit to the Internet backbone. Each "client" will download between 7MB up to 30MB per request. The download mirror servers will be directly attached to the Internet (no FW or router between, which could handle QoS for me). Of course will these download mirrors be specially hardened :-)

The Idea:
====> I can not handle all the traffic at once without QoS, so I thought it should be possible to establish a traffic control mechanism to guarantee bandwith for each client download (e.g. 256kbps CIR up to 512kbps ceil). This should allow me, to "serve" all clients within 120 min, before a new download wave will hit my servers. If the available bandwith for HTTP traffic is used, clients shall not be able to start a download, until some bandwith is freed by another client who finished a download.

Note: I would rather go with "tcng" than with the old "tc".

The Problem:
======
While playing with tcng and HTB I managed to shape traffic (to limit the total bandwith for HTTP traffic to e.g. 50Mbps) but I did not manage to write a tcng configuration which guarantees a certain bandwith per client request. I must be able to assign (and guarantee) a defined bandwith to a client request, to make the downloads more predictable. My experiments so far only allowed me to limit e.g. HTTP traffic, which then was equally distributed to the number of clients, but this led to the usual "traffic jam" when too many clients initiated a simultanes download.

Quoted 3 lines

Can anybody please assist me, or does anybody have a working "tcng" config file to do the job. Additional readings/comments/links (beside those listed in the LDB/Linux Traffic Control Howtos are also welcome.)

Wang Jian posted a perflow queue recently -

http://mailman.ds9a.nl/pipermail/lartc/2005q2/015381.html

It is the sort of thing you need if you can't do it with apache as recommended.

Andy.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#4 lore raw
In reply to Thomas Mandl

Hello Everybody,

I forgot to mention, that my HTTP traffic will not be handled by an apache web server, and the HTTP protocol can be replaced in the near future by a different protocol. Thus apache's mod_bandwith isn't an option for me right now.

I'm not sure how Wang Jian's post can help me. I thought the existing Linux 2.6.11 kernel and tcng already can do what I need? All I have to do is write a proper tcng config file...

But the proper tcng file is the challenge...

regards
Thomas
-----Ursprüngliche Nachricht-----
Von: Andy Furniss [mailto:[email protected]]
Gesendet: Mittwoch, 27. April 2005 23:23
An: [email protected]
Cc: LARTC
Betreff: Re: [LARTC] urgent question about tcng!

Quoted 15 lines · Thomas Mandl

Hello List,
I'm new to QoS/tcng/HTB and friends, so please forgive me if my question might be silly...
After having read lots of HowTo documents I'm totally confused...

The Challenge:
=======
I'll have to deploy several "mirror" download servers (Linux) which must be able to handle a huge number of HTTP download requests (about 10k to 20k unicast requests per server within 120 min.) which will hit my servers . The mirror servers will be connected with 100MBit to the Internet backbone. Each "client" will download between 7MB up to 30MB per request. The download mirror servers will be directly attached to the Internet (no FW or router between, which could handle QoS for me). Of course will these download mirrors be specially hardened :-)

[...snip...]

Quoted 5 lines

Can anybody please assist me, or does anybody have a working "tcng" config file to do the job. Additional readings/comments/links (beside those listed in the LDB/Linux Traffic Control Howtos are also welcome.)

Wang Jian posted a perflow queue recently -

http://mailman.ds9a.nl/pipermail/lartc/2005q2/015381.html

It is the sort of thing you need if you can't do it with apache as recommended.

Andy.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#5 lore raw
In reply to Thomas Mandl
Quoted 7 lines · Thomas Mandl

Hello Everybody,

I forgot to mention, that my HTTP traffic will not be handled by an apache web server, and the HTTP protocol can be replaced in the near future by a different protocol. Thus apache's mod_bandwith isn't an option for me right now.

Quoted 5 lines

I'm not sure how Wang Jian's post can help me. I thought the existing Linux 2.6.11 kernel and tcng already can do what I need? All I have to do is write a proper tcng config file...

But the proper tcng file is the challenge...

I don't know tcng, but the reason I suggested perflow is that you want each flow to have a ceil - unless you make a class and rule to match each flow I can't see how you can do this. Also iptables could limit the number of connections - tc can't, perflow can.

Andy.
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#6 lore raw
In reply to Thomas Mandl
Quoted 5 lines · Andy Furniss

I don't know tcng, but the reason I suggested perflow is that you want each flow to have a ceil - unless you make a class and rule to match each flow I can't see how you can do this. Also iptables could limit the number of connections - tc can't, perflow can.

Forgot to say you can also use iptables to limit rate per connection.

Andy.
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#7 lore raw
In reply to Thomas Mandl

Sorry for the stupid question, but how would I rate limit connections using iptables?

regards
Thomas
-----Ursprüngliche Nachricht-----
Von: Andy Furniss [mailto:[email protected]]
Gesendet: Dienstag, 03. Mai 2005 16:43
An: Andy Furniss
Cc: [email protected]; 'LARTC'
Betreff: Re: AW: [LARTC] urgent question about tcng!

Quoted 5 lines · Andy Furniss

I don't know tcng, but the reason I suggested perflow is that you want each flow to have a ceil - unless you make a class and rule to match each flow I can't see how you can do this. Also iptables could limit the number of connections - tc can't, perflow can.

Forgot to say you can also use iptables to limit rate per connection.

Andy.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#8 lore raw
In reply to Thomas Mandl
Quoted 3 lines · Thomas Mandl

Sorry for the stupid question, but how would I rate limit connections using iptables?

I was thinking of the patch-o-matic-ng patches connlimit and connrate, though I've never used either and they may or may not be expensive for many connections compared to perflow.

Connrate lets you mark packets if they are above limits - you could then drop those later in a filter table or with a TC filter/queue.

Quoted 18 lines

regards
Thomas
-----Ursprüngliche Nachricht-----
Von: Andy Furniss [mailto:[email protected]]
Gesendet: Dienstag, 03. Mai 2005 16:43
An: Andy Furniss
Cc: [email protected]; 'LARTC'
Betreff: Re: AW: [LARTC] urgent question about tcng!

Quoted 4 lines · Andy Furniss

I don't know tcng, but the reason I suggested perflow is that you want each flow to have a ceil - unless you make a class and rule to match each flow I can't see how you can do this. Also iptables could limit the number of connections - tc can't, perflow can.

Quoted 5 lines

Forgot to say you can also use iptables to limit rate per connection.

Andy.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc