Per-flow IPv4 ECMP

Per-flow IPv4 ECMP

[Matthew Dupre]

Hi,

I'm interested in the Linux kernel's support for per-flow IPv4 ECMP (i.e. consistent path selection based on a hash of the connection tuple).  I'd been led to believe[1] that this depended on the route cache, which was removed in 3.6.

However, I tested a route with multiple next hops on a 3.10 and 3.13 kernel, and ECMP was per-flow!  Obviously I'm pleased that this was the case, but I'd like to understand why this is supported, and whether I can rely on it in future.

Could anyone give me a little clarification on whether this is now supported by some means other than the route cache, and whether that support is intended to be continued?

Thanks,
Matt

[1]: http://bird.network.cz/pipermail/bird-users/2014-February/008873.html

Re: Per-flow IPv4 ECMP

[roopa]

On 9/28/15, 1:57 AM, Matthew Dupre wrote:
> Hi,
>
> I'm interested in the Linux kernel's support for per-flow IPv4 ECMP (i.e. consistent path selection based on a hash of the connection tuple).  I'd been led to believe[1] that this depended on the route cache, which was removed in 3.6.
>
> However, I tested a route with multiple next hops on a 3.10 and 3.13 kernel, and ECMP was per-flow!  Obviously I'm pleased that this was the case, but I'd like to understand why this is supported, and whether I can rely on it in future.
>
> Could anyone give me a little clarification on whether this is now supported by some means other than the route cache, and whether that support is intended to be continued?
>
This is being worked on currently by Peter Nørlund
https://lwn.net/Articles/657431/

Re: Per-flow IPv4 ECMP

[Peter Nørlund]

On Mon, 28 Sep 2015 06:48:09 -0700
roopa <roopa@cumulusnetworks.com> wrote:

> On 9/28/15, 1:57 AM, Matthew Dupre wrote:
> > Hi,
> >
> > I'm interested in the Linux kernel's support for per-flow IPv4 ECMP
> > (i.e. consistent path selection based on a hash of the connection
> > tuple).  I'd been led to believe[1] that this depended on the route
> > cache, which was removed in 3.6.
> >
> > However, I tested a route with multiple next hops on a 3.10 and
> > 3.13 kernel, and ECMP was per-flow!  Obviously I'm pleased that
> > this was the case, but I'd like to understand why this is
> > supported, and whether I can rely on it in future.
> >
> > Could anyone give me a little clarification on whether this is now
> > supported by some means other than the route cache, and whether
> > that support is intended to be continued?
> >
> This is being worked on currently by Peter Nørlund
> https://lwn.net/Articles/657431/

Hi,

AFAIK if you create a socket on the machine having the multipath route,
each socket will seemingly be mapped to particular path, and it will
behave as per-flow. But if you use the machine as a router, each packet
is forwarded independently, potentially hitting different paths each
time.

Best Regards
  Peter Nørlund

Re: Per-flow IPv4 ECMP

[Jean He]

Peter Nørlund <pch <at> ordbogen.com> writes:

> > This is being worked on currently by Peter Nørlund
> > https://lwn.net/Articles/657431/
> 

Hi Peter,

Following up on this thread, has the new patch for supporting per-flow IPv4 
ECMP landed? 

Which version should I be able to use that has this feature?

Thanks,
Jean