Underwhelming speed increase, looking for advice

DMerz III · March 10, 2016, 10:03:42 AM

Hey Keyshot friends!

Been a heavy user now for about 18 months. My company decided to upgrade our rendering capacity, and got us another custom Boxx machine to add to our render network. So I wanted to do a test of how much faster this addition got us and report back that it was a worth while investment.

The test render was something we were working on for a client so unfortunately, I cannot share this file at this time, but I plan on doing this test over with a different demo file.

Old Config: 2 Boxx RenderPros : One set at Master & Slave the other set as a Slave 48 total cores. Render time: 53 mins.

After the additional Boxx RenderPro was added into the mix, One set to Master & Slave, the other 2 set as slaves, 3 total slaves, 72 total cores.
That's 50% more cores! Should make a significant difference with all the same settings, right? Render time: 49 mins!

Only shaved off 4 mins =(

THEE ONLY DIFFERENCE I could determine, was that when we added the new RenderPRO we updated the keyshot network from 6.1.72 to 6.1.77

Any insight as to why we're getting an underwhelming speed change after adding 50% more computing power, would be greatly appreciated.

Thanks guys!

Chad Holton · March 10, 2016, 12:22:45 PM

Hello - How many NR cores does your license allow?

DMerz III · March 11, 2016, 09:40:54 AM

96! We thought of that too, but we checked and we're covered. Definitely says we have 72 cores connected in the Network Queue as well.

However, upon further inspection, we found an additional piece of geometry on at the same time which was not turn on in the first render. We re-ran our test, and got it down to 44 mins. So, we shaved off 9 mins from 53. Not terrible, but still somewhat underwhelming after spending several thousand dollars on computing upgrades.

theAVator · March 11, 2016, 02:56:19 PM

Just curious, what were you using for the test render? Was it just a single image, animation, or VR?
Do you also have the settings you used?

DMerz III · March 16, 2016, 08:05:58 AM

We're working on a project for a client, so I can't share too much detail about what the subject actually was. Single image. There were some transparent glass parts, but that was about 10% of the overall image. The rest was metallic paint. The image was utilizing zero alpha channel. So the composition was frame to frame.

Our settings were
Advanced Control:
Samples: 100
Ray B: 16
Anti-Alias: 1
Shadow: 3
GI: 1
Pixel Blur: 1.5

TIF (8bit)

No complaints about the actual render time, more concerned with how much time was actually reduced when going from 48 cores to 72. Like I said, we went from 53mins to 44mins. Was expecting more like 35-38 mins. Real question is, when does upping the # of cores stop giving you return on speed. I'm asking because we're willing to invest in an additional 24 cores to add to the render network, but if it's not going to give me significant speed boost, I'd rather allocate that money elsewhere.

Thanks for anyone who has some experiences to share!

Chad Holton · March 16, 2016, 08:19:09 AM

What are the core speeds for each RenderPro?

DMerz III · March 16, 2016, 10:36:53 AM

We have three of these:

DUAL XEON E5-2680v3 2.5GHz, 30MB cache, 9.60 QPI (Twelve-Core)
32GB DDR4-2133 REG ECC (8 - 4GB DIMMS)
240GB SSD SATA 6Gb/s
Microsoft Windows 7 Professional Edition 64-Bit

Model Specifications
Dual Intel Xeon E5-2600v3 Processor
Intel C612 Chipset with QPI up to 9.60GT/S
Up to 256GB DDR4 2133Mhz DIMM REG ECC
1 x Intel i350 Dual-port Gigabit Ethernet Controller
2 x SATA 6Gbps Ports
3 x USB 2.0
1 x PCIE x 16 Low Profile Slot
1 x ASPEED AST2400 BMC
1 x IPMI 2.0 (Intelligent Platform Management Interface v.2.0)
1 x 350 Watt Power Supply

Chad Holton · March 16, 2016, 11:12:24 AM

Thanks for the info. Very nice setup! You mentioned that you had extra geometry in the first comparison example, removed it and gained speed for the second comparison. You're not too far from your expectation time - so, is it possible that there are still other variations in the scene making it different and slowing the rendering process?

DMerz III · March 16, 2016, 11:36:34 AM

Thanks! We've been pretty excited about our setup, we're just trying to find where we can optimize things.

The only change was we went from 6.1.72 to 6.1.77 on the Network Render. Everything was updated in tandem (Master/Slaves/Queues)

If we're close to the expectation, is there a 'predictable' time we should expect if we were to add 1 more of these machines to our render network?

Another thought we had was, what if we split 4 machines into two network queues. Had 2 RenderPros connected, and separately had 2 other RenderPros connected. It would mean that we'd need two masters and have to figure out how to send jobs to either one manually, but essentially it'll give us 2 queues rendering at the same time instead of 1 line that is just slightly faster.

Thoughts?

DMerz III · March 22, 2016, 03:16:15 PM

Another note:

Been a couple weeks now, and I have just noticed another weird thing. When you open the window that is called 'slave status', I'll see the 3 machines that are slaved for the network. All the way to the right, there is a column titled 'Total Regions' as in, these are the total regions each machine was assigned and then rendered.

Keep in mind, each of the 3 machines are thee exact same specs.
Computer1 (Master and Slave) - 1303
Computer2 (Slave only) - 1311
Computer3 (Slave only) - 275

As far as I know, these #s will reset if the machines are power cycled (turned off and then on again). And as far as I know, they were all last restarted at the same time, so unless #3 restarted recently without me knowing, this machine lags waaaaay behind the other two in regions rendered. Could this be significant?

Arn · March 31, 2016, 08:38:08 PM

Quote from: dmerziii on March 22, 2016, 03:16:15 PM
As far as I know, these #s will reset if the machines are power cycled (turned off and then on again). And as far as I know, they were all last restarted at the same time, so unless #3 restarted recently without me knowing, this machine lags waaaaay behind the other two in regions rendered. Could this be significant?

It might be good to double check the machine is performing as expected. Confirm the machine is running at the correct clock speed and utilizing all the cores and threads. You could also check the Windows logs and stability index and compare it to the other machines.

If, for instance, the machine is much warmer than the others, it might throttle itself back to prevent damage. You will not see much difference, other than reduced performance. Though there are many more reasons a machine might perform different from similar computers.