OTOY Forums

Thanks for explaining it. Any news on upping the GPU limit for 3.0 ?

BorisGoreta wrote:Thanks for explaining it. Any news on upping the GPU limit for 3.0 ?

The official word is to reconsider the limit, but I don't have any new details for it. Sorry.

Is it possible to scan nodes on both IP sub groups, 192 and 168, at the same time since I would like to have 1 node at 192 and 1 node at 168 so that the main workstation could send data at 100 MB/s to both nodes. If I have them connected with the same switch then the every node gets half the link speed.

No, that's not possible unless you configure a different sub-net in each master. It also means you would not be able to use both slaves in one master.

As an alternative you could launch the slaves directly via the command line and not via a daemon. This way you can assign the slaves to whatever master you want, but of course, would have to do that every time you want to launch/stop a slave.

BorisGoreta wrote:Is it possible to scan nodes on both IP sub groups, 192 and 168, at the same time since I would like to have 1 node at 192 and 1 node at 168 so that the main workstation could send data at 100 MB/s to both nodes. If I have them connected with the same switch then the every node gets half the link speed.

Network rendering only works on the same subnet. The master discover daemons by broadcasting messages which is per subnet. So a master can only work on a single subnet. If you do have access to multiple subnets, you could assign each master to a subnet. Likely your PC would have 2 NICs , one for each subnet. I haven't tried this before but in theory it should work.

cheers,
Thomas

Let's say that both nodes are connected to a same switch and master workstation is also connected to this switch so all are on the same subnet.

When workstation sends data to the nodes is the actual data transfer halfed from 100 MB/s to 50 MB/s because it pushes data to both nodes ?

Or is the speed still 100 MB/s because the data pushed is exactly the same ?

BorisGoreta wrote:Let's say that both nodes are connected to a same switch and master workstation is also connected to this switch so all are on the same subnet.

When workstation sends data to the nodes is the actual data transfer halfed from 100 MB/s to 50 MB/s because it pushes data to both nodes ?

Or is the speed still 100 MB/s because the data pushed is exactly the same ?

When the master distributes the render data to multiple slaves, it will update them sequentially starting with the slave with the most GPUs or with the one that requires the least amount of data being transmitted. The idea here is to have at least some slaves starting rendering while others are still updating. Data updates are also differential, i.e. when changes cause Octane to restart rendering, only those bits are sent that actually have changed. And we also use compression as much as reasonably possible. All these optimizations together make the data transfer of scenes to the slaves surprisingly fast. At least, I didn't see many complains about slow slave updates.

I see, thanks for explaining. So I wouldn't benefit from link aggregation having one dedicated 100 Mb/s LAN path to one node and another dedicated 100Mb/s LAN path to the other node if you say they are updated squentially.

But when testing real world situations I can clearly see that both nodes kick in approximately at the same time as if they are updated simultaniously so I don't experience this sequential update.

Not sure how it relates to this discussion but thought I'd weigh in.

We use Butterfly Net Render to manage our NETWORK rendering. I can load up scenes inside BNR and have it manage the NODES for rendering. A lot of times I will have one GPU on my local machine being used while my local LW instance uses the other. As long as other machines are licensed I could have other nodes with up to 12 gpu's on each machine being used and managed by BNR. We have not gotten that far yet...but it will be possible.

I think there are FREE LWSN render managers that could do similar things for you. I think this could give you what you are wanting...being able to use as many local GPU's on your machine for yourself...while giving access to the remaining ones for LWSN with an Oactane.cfg file.

It works with multiple LW Layout instances, you can assign different nodes to different instances so they can render different tasks, they can also be on different networks if you enable auto configure master port.