I can't see how it could scale. Each node needs a table of all other nodes to calculate the cost for each path. This is a LOT of data and each node rebroadcasts it's neighbor's information so the more neighbors, the more rebroadcasting.

Maybe if you had a long string of nodes with each node only seeing 1-2 other nodes, it might scale up higher, but that's not the point of mesh networking.

It really means that under idea circumstances, overhead is going to scale up with node count by percent. So a few nodes is low overhead, but 50 nodes might be as much as 50% and so on.

This is a critical flaw in batman-adv to scale up. Requiring each node to know all other nodes. If this rule could be changed to have batman-adv only keep track of gateway nodes and nodes in the path to a gateway, then it could scale vastly further.

On Wed, Apr 26, 2017 at 9:10 AM, jens jens@viisauksena.de wrote:

i think this could scale , but you will need sufficient power / ethernet capacities at the nodes.

and the biggest problem may be that these endpoints make much noise on the layer2 level. (disovery protocolls and stuff like this)

you could easily imagine what would happen if you have a layer2 switch with 100000 ports.

i think that most of the time you will not do this, and implement some sort of routing between some clouds of batman-adv networks.

On 26.04.2017 16:20, fuumind wrote:

...

Hi list!

Been lurking for almost a year on the battlemesh list and recently joined here as well.

I'm curious about how well batman-adv scales. Would a network of 10 000 nodes work well? What about 100 000 nodes or 1 000 000?

Thanks! fuumind

-- make the world nicer, please use PGP encryption

I don't think it's CPU that's the issue, it's WLAN overhead. Too many nodes is too much overhead traffic.

The web is telling me there is an upper limit of a 'basic' cluster of about 300, and if you add some multicast filtering you can hit maybe 1500. OGM's are eventually going to saturate the network though.

That makes me think you should be looking at small clouds and some other 'routing' protocol to connect those clusters together. Maybe do an OSPF ring linking up the batman-adv clouds.

On Wed, Apr 26, 2017 at 10:06 AM, fuumind fuumind@openmailbox.org wrote:

Hi Jens!

By 'sufficient power', do you mean processing power to handle overhead traffic? I'm imagining a network which primarily connects WLANs.

Clouds of batman-adv networks sounds like a good idea. Any idea about how to implement the interfaces between these clouds? How big should they be allowed to grow before forming a new cloud?

fuumind

ons 2017-04-26 klockan 17:10 +0200 skrev jens:

...

i think this could scale , but you will need sufficient power / ethernet capacities at the nodes.

and the biggest problem may be that these endpoints make much noise on the layer2 level. (disovery protocolls and stuff like this)

you could easily imagine what would happen if you have a layer2 switch with 100000 ports.

i think that most of the time you will not do this, and implement some sort of routing between some clouds of batman-adv networks.

On 26.04.2017 16:20, fuumind wrote:

...

Hi list!

Been lurking for almost a year on the battlemesh list and recently joined here as well.

I'm curious about how well batman-adv scales. Would a network of 10 000 nodes work well? What about 100 000 nodes or 1 000 000?

Thanks! fuumind

Ok, so clouds of a few hundred nodes would work reasonably well and at the same time keep customization to a minimum.

What makes OSPF a better candidate for interfacing between the clouds? Does it keep a more limited routing table and thus minimize overhead traffic? I haven't been able to figure it out good enough yet.

fuumind

ons 2017-04-26 klockan 18:05 +0200 skrev Simon Wunderlich:

Hi,

"works well" depends much on  your scenario.

I've seen many well-working networks with 100-300 nodes. There are Freifunk  community networks with over 1000 nodes running batman-adv (i.e. "standard"  users like laptops and smartphones), but they employ a lot of filtering to  avoid too much broadcast.

Unless you have special restrictions like minimal broadcast traffic or high  capacity on the wireless links, memory and CPU, I would not recommend to plan  for 10 000 nodes and beyond.

Cheers,      Simon

On Wednesday, April 26, 2017 4:20:59 PM CEST fuumind wrote:

...

Hi list!

Been lurking for almost a year on the battlemesh list and recently joined here as well.

I'm curious about how well batman-adv scales. Would a network of 10 000 nodes work well? What about 100 000 nodes or 1 000 000?

Thanks! fuumind

Now I get it. :) That makes sense.

Thanks!

fre 2017-04-28 klockan 14:23 +0200 skrev Simon Wunderlich:

Hi,

On Friday, April 28, 2017 12:44:08 PM CEST fuumind wrote:

...

Ok, so clouds of a few hundred nodes would work reasonably well and at the same time keep customization to a minimum.

Yes

...

What makes OSPF a better candidate for interfacing between the clouds? Does it keep a more limited routing table and thus minimize overhead traffic? I haven't been able to figure it out good enough yet.

The idea would be that you do a layer 3 protocol only on a few selected  "gateway" nodes of the clouds. Those gateways would forward traffic from one  cloud to the other, without having the details (and thus not having the  overhead) of the state of the mesh.

Which technology is the best is beyond my knowledge.

Cheers,       Simon