[LARTC] NEW imq driver

[LARTC] NEW imq driver

[Roy]

Finaly I made imq driver stable it did not crashed for all 5 hours under
high load, soo looks stable.
(old one was crashing after 1-5 min for me)

no need to patch anything just compile and insmod, should work with any
kernel probably must be > than 2.4.20

This is completely diferent code than old imq.
you can find it on my server http://pupa.da.ru
please tell how it works for you and how stable it is.

_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

Re: [LARTC] NEW imq driver

[rubens]

> Finaly I made imq driver stable it did not crashed for all 5 hours under
> high load, soo looks stable.
> (old one was crashing after 1-5 min for me)

It seems to capture ingress and egress traffic of all interfaces; wouldn't
this count packets twice ?

If the machine is doing SNAT or DNAT, what IP addresses would be seen by
the qdisc ?


Rubens


_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

Re: [LARTC] NEW imq driver

[Michael Renzmann]

Hi.

Roy wrote:
> Finaly I made imq driver stable [...]
> This is completely diferent code than old imq.

May I then second the proposal to give the driver another name? How 
about IMQ2, IMQng (next generation) or something like that?

Bye, Mike

_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

Re: [LARTC] NEW imq driver

[Roy]

Seems I was to fast to declare success,
my version is not much more stable than the original one,everything depends
on dropped packets.
 This is even not imq fault afterall, can be prowed in other way also:

atempts to police outgoing trafic it will be ok until you dont touch localy
generated packets
if you try to drop them you will be sorry, because kernel will resend then
together with new ones
of cource policer will drop them too, but linux kernel keeps resending then
thus increasing rate progresively.

I noticed that with my trafic counter. internal trafic grew to enormous
levels 10X more than it can be. In reality there was almost no output at
all.
so DONT USE POLICERS ON EGRESS. on low trafic it is harmless but on 100mb/s
it probably can kill computer (not tested).

Seems imq have similar problem  even if driver itself have no leaks kernel
consumes all resousces on resnending droped packets so that computer stops
responding


for now I dont have good idea how to fix it so I will try to avoid localy
generated trafic
so it will me possible to shape ingress and forward, egress will be left for
real device.
maybe later I will find how fix that



>
> It seems to capture ingress and egress traffic of all interfaces; wouldn't
> this count packets twice ?

No, ingress is for local and egress for everything so everything should be
ok (in theory)

> If the machine is doing SNAT or DNAT, what IP addresses would be seen by
> the qdisc ?
>

I made driver see the final destination address because it is more usefull


> Rubens
>
>
> _______________________________________________
> LARTC mailing list / LARTC@mailman.ds9a.nl
> http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
>

_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

Re: [LARTC] NEW imq driver

[Andy Furniss]

Roy wrote:
> Seems I was to fast to declare success,
> my version is not much more stable than the original one,everything depends
> on dropped packets.
>  This is even not imq fault afterall, can be prowed in other way also:
> 
> atempts to police outgoing trafic it will be ok until you dont touch localy
> generated packets
> if you try to drop them you will be sorry, because kernel will resend then
> together with new ones
> of cource policer will drop them too, but linux kernel keeps resending then
> thus increasing rate progresively.
> 
> I noticed that with my trafic counter. internal trafic grew to enormous
> levels 10X more than it can be. In reality there was almost no output at
> all.
> so DONT USE POLICERS ON EGRESS. on low trafic it is harmless but on 100mb/s
> it probably can kill computer (not tested).
> 
> Seems imq have similar problem  even if driver itself have no leaks kernel
> consumes all resousces on resnending droped packets so that computer stops
> responding
> 
> 
> for now I dont have good idea how to fix it so I will try to avoid localy
> generated trafic
> so it will me possible to shape ingress and forward, egress will be left for
> real device.
> maybe later I will find how fix that

Which queue do you use to drop the packets?

Andy.

_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

Re: [LARTC] NEW imq driver

[Philip Thiem]

[-- Attachment #1: Type: text/plain, Size: 1108 bytes --]

[snip]

> I noticed that with my trafic counter. internal trafic grew to enormous
> levels 10X more than it can be. In reality there was almost no output at
> all.
> so DONT USE POLICERS ON EGRESS. on low trafic it is harmless but on
> 100mb/s it probably can kill computer (not tested).
>
> Seems imq have similar problem  even if driver itself have no leaks kernel
> consumes all resousces on resnending droped packets so that computer stops
> responding

[snip]

Just curious, I suppose you had the same setup on the original IMQ? If not,
my point is moot.  But if so and if the problem could be verified, it would
appear to me that there is less of a problem with the old patch(except
for features), and more of a problem with the inherent nature of what
is being attempting.  Meaning that a set of assumptions made by the network
layer developers is being invalidated.  For example, local outbound traffic
being policed instead of shaped.


Philip Thiem -- Icequake.net Administrator
Isn't it obvious lumberjacks love traffic lights?
GPG Pub Key Archived at wwwkeys.us.pgp.net

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]