Resolve Disk IO performance is sluggish compared to Scratch How to Improve

[Margus Voll]

How exactly did you test so that you can say it is io and not the lack of gpu ?

—

Margus

https://iconstudios.eu

DaVinci 9, OSX 10.7.4
MacPro 5.1 2×2,93 24GB
GTX 470 / Quadro 4000
Multibridge 2 Pro

[Rohit Gupta]

How are you monitoring stereo? Two SDI outs, or just one.

Try turning on optimized video monitoring option and see if that helps.

Also, try your image scaling to bilinear.

[Toby Risk]

Hi Rohit and Margus, thanks for the interest

Firstly I should clear up something in my previous post, comparing a 2K playback on Resolve with a 4K playback on Scratch.

Scratch did indeed play the 4K (3996×2160) stereo at 24fps realtime. After I posted the first post I manage to run exactly the same test on Resolve with the 4K files.

The same files played at 10fps on Resolve, which is double the data throughput performance of the 2K playback (half as quick, not a quarter, as suggested, as Scratch).

In terms of Disk IO this was about 800MB/s on Resolve and 1.7GB/s on the Scratch.

I only need 400MB/s for the stereo 2K, so Resolve CAN pull enough data through quick enough for the 2K stereo playback, but something is throttling it.

Regarding whether it could be a GPU issue or not. It’s possible, I don’t know enough about DaVinci’s internal workings to say otherwise. But if the GPU was topping out with a stereo 1998×1080 stream then it should have topped out at around 5.5fps with the 3996×2160 files, not 10fps. However, if BMD say that’s the issue then I’ll put one in tomorrow. My feeling is the FX6000 is more than adequate for this throughput test. This playback is without any LUT’s, grade, resizing or image manipulation whatsoever.

In regards to the SDI setup. This was to be a test to see if Resolve worked well on the Globalstor. At the moment it is our Scratch production machine and currently uses the SDI daughter board of the 6000 for the SDI output which Resolve obviously can’t use. If Resolve is proven on the machine then I’ll go buy the Decklink 3D.

So I am monitoring just on the GUI monitor and going by the fps indicator at the top left of frame. Stereo is toggled on, (I am assuming this is reading both frames as the playback speed slows down as opposed to mono where it plays realtime), but I have selected ‘none’ on the stereo viewer setup, so as not to tax any gpu in processing the gui.

If the feeling is that this is the issue then I’ll see if I can put one of my old decklinks (not 3D) and an FX4000 in it tomorrow and see what happens. I’ll also check the Bilinear setting and video optimisation setting and let you know how it goes.

My personal belief is that it’s something in the Disk I/O calls or the size of the chunks of the frames it reads from the disk. There was an issue (now fixed) with an older version of Scratch that if a file’s size, say DPX was not perfectly aligned in 8KB blocks then Scratch would read each file twice necessitating double the IO bandwidth. Maybe there’s something like that in Resolve.

(feature request) It would be nice if Resolve had a popup box with a complete system stats readout somewhere to monitor the cpu/memory/IO/GPU performance of the machine.

Thanks again and sorry for my ramblings

Toby

Colourist | Editor | Post-Production Consultant — 25 years at the post-production coalface, and still loving it.

[Peter Chamberlain]

Hi Toby, as a test, what does the BMD speed test report for your disk speed.
Can you set your timeline fps to 48 and play a 2K stream of dpx. This should indicate the frame rate from disk and through the gpu.
Remember that Resolve uses the GPU for image scaling so if you have a 4k source but monitor in HD, scaling is being performed.. thus GPU used. Comparisons of CPU usage are pointless… they are different apps with quite different processes.
Its seems you have not configured the system as per our config guides. Using a single shared GPU for mono is restrictive let alone for stereo. Also slot position for GPUs is important… the UI and processing GPUs should be in x16 slots that don’t share bw with other slots.
Resolve software scales to 8 GPUs on linux, 4 on Windows and 3 on Mac, we have many customers using these configs for stereo but I couldn’t say I know of one that has a single GPU for more than mono 2K use. Of course you can preview with a single, and this works in laptop as well but its not as per our recommendation in the config guides. I hope this helps.
Peter

[Toby Risk]

Thanks Peter

I’ll see if I can put in another GPU and run some more tests. If that fixes it then I’ll be happy.

I’ll let you know we get on.

Regards

Toby

Colourist | Editor | Post-Production Consultant — 23 years at the post-production coalface, and still loving it.

[Toby Risk]

Hi Peter

I have tried it with a 2nd GPU, an FX4000. There is an improvement, but not significantly so. The speed creeps up to 24fps and drops a couple of frames every 2 or 3 seconds, but I don’t think this was a fair test as I am limited with how much reconfiguring I can do on this machine.

I appreciate what you say about the GPU resources and bw requirements. Therefore, I am going to hold off for a bit until I can check it out thoroughly with a system I can play around without risk of losing our production machine.

I’ll let you know how it goes when I get to it.

Thanks

Toby

Colourist | Editor | Post-Production Consultant — 23 years at the post-production coalface, and still loving it.

[Sascha Haber]

Please use a GTX680 if possible.
The quadros available are helplessly outdated.
The Quadro 6000 is like half as fast as a GTX580 for CUDA processing.
But nether less ,Scratch uses OpenGL for playback, I dont think Resolve can ever come close to its performance playback wise.
There is no processing, just a polygon mapped with a texture..

A slice of color…

Resolve 9.0.4 OSX 10.8.2

Colorist / Aerial footage producer
https://vimeo.com/saschahaber