These are chat archives for buddhi1980/mandelbulber2

7th
May 2017
Martin
@acidhunter
May 07 2017 09:40

nice to see the latest opencl work^^

i can supply gpu servers for testing

cuda inside docker is working, opencl is the next thing.
when i got it working i push the image to dockerhub
Krzysztof Marczak
@buddhi1980
May 07 2017 15:20
Good to know. Every machine for testing will be appreciated. I remember problems encountered when I developed old Mandelbulber with OpenCl. Big viariety of GPUs and systems made a lot of problems.
Martin
@acidhunter
May 07 2017 17:03
i tried to get opencl running with nvidia-docker (https://github.com/NVIDIA/nvidia-docker)
Krzysztof Marczak
@buddhi1980
May 07 2017 17:04
what is the result?
Martin
@acidhunter
May 07 2017 17:05
the container is based on ubuntu trusty, the latest nvidia driver on trusty is nvidia-375
to get it working the driver must match to the version on the host
on my arch workstation is nvidia-378 installed
Krzysztof Marczak
@buddhi1980
May 07 2017 17:06
why it has to match?
Martin
@acidhunter
May 07 2017 17:06
i get errors with that ubuntu container
other tests with arch containers, including the same versions in container/host did work partially
i had success running gpu accelerated applications inside docker, deepdream and tensorflow too
what would be the best way to supply hardware for testing?
i think about containers with root access on these machines
Krzysztof Marczak
@buddhi1980
May 07 2017 17:11
I have no idea
Martin
@acidhunter
May 07 2017 17:12
i try to get a second gpu. then i can setup a KVM with gpu passthru
Krzysztof Marczak
@buddhi1980
May 07 2017 17:12
I have no experience with dockers. I would like to focus on writing the code. Testing I'm letting to another people like you :)
Martin
@acidhunter
May 07 2017 17:13
i do some research with docker the next days
Krzysztof Marczak
@buddhi1980
May 07 2017 17:13
But now is too early for testing, because code still doesn't exist
Martin
@acidhunter
May 07 2017 17:13
id do this in spare time at work
Krzysztof Marczak
@buddhi1980
May 07 2017 17:14
I have just finished part of code which prepares OpenCl for rendering, but it's only 1% of total work
Martin
@acidhunter
May 07 2017 17:14
ok, then i have some time^^
i need the docker thing for some other tasks for my work
so there is motivation to get this done
yesterday i got a basic kubernetes setup
i finish this the next weeks
with documentation for new users
everyone can get 300$ credit on google compute for free
Martin
@acidhunter
May 07 2017 17:19
with my documentation and dockerfiles its really easy to get up a cluster, nearly unlimited in size
Martin
@acidhunter
May 07 2017 17:26

is it possibly to render in tiles?
i think this could give better performance

example:
a settings file will distributed to 4 worker, every worker renders 1 tile.
later the finished tiles get pushed to a server that merges the tiles to 1 picture

Krzysztof Marczak
@buddhi1980
May 07 2017 17:27
Now it's divided to lines, which gives the same result as dividing by titles
Martin
@acidhunter
May 07 2017 17:27
some days ago i tried to network render at 20k x 20k
i took 5 minutes to start the rendering
if every worker is standalone i think this could give a sppedup
Krzysztof Marczak
@buddhi1980
May 07 2017 17:28
Ok, I agrre with this case, but it's not typical usage. If the image will be divided to tiles, then would be difficult to balance it
every tile will have different rendering time
Martin
@acidhunter
May 07 2017 17:29
on small images the actual way of doing network rendering is nice
but if u scale up to 100 or more workers it would get rrealy slow
Krzysztof Marczak
@buddhi1980
May 07 2017 17:30
And now every worker is independent. Can work even with very slow internet connection, because doens't need requests from server all the time. Every worker can decide what to render next.
Why you think that it would be slow?
Martin
@acidhunter
May 07 2017 17:31
a 20k x 20k image needs some minutes to get the network clients start rendering
Krzysztof Marczak
@buddhi1980
May 07 2017 17:31
The problem is lack of memory in this case
Martin
@acidhunter
May 07 2017 17:31
yes that is the point
sry its difficult to explain, my english sucks a bit
beacause that i ask if it would be possible to render tiles
if a settings file get uploaded to the nodes, every node is only rendering a part and uploads the image afterwards to a ftp
so every node works alone
Krzysztof Marczak
@buddhi1980
May 07 2017 17:34
What I can change is another allocation of memory for NetRender clients. Now it allocates memory for full image. I can change this to allocate only that parts which are actually rendered by given client.
It should solve this problem.
Martin
@acidhunter
May 07 2017 17:34
i dont know if this works
just an idea
Krzysztof Marczak
@buddhi1980
May 07 2017 17:38
I don't like idea with tiles, because it will be difficult to balance between nodes. Some of nodes can finish rendering very fast and some after very long time. Then it will be unefficient.
Martin
@acidhunter
May 07 2017 17:38
i dont mean to implement this in netrender
Krzysztof Marczak
@buddhi1980
May 07 2017 17:39
If you just ask to rendering parts of image, it's only partially implemented in the code, but not finished.
Martin
@acidhunter
May 07 2017 17:40
another example,
uploading a settings file to 100 servers, every server renders a defined part and uploads it afterwards to a ftp or something else
so it would be possible to distribute it to a cloud provider to render images with unlimited size
after all nodes are finished the tiles can merged
Martin
@acidhunter
May 07 2017 17:47
output could be something like this