Optimizing Network Scalability

Please use this forum for general user support and related questions.

Moderator: coordinators

Forum rules
Please include your operating system type/version, LuxRender version and Exporter version used when submitting a support post.

Make sure you have read the Release forum thread for Release and RC (Release Candidates) builds as these threads contain information on known problems and workarounds: Test Builds Forum

Re: Optimizing Network Scalability

Postby Heavily Tessellated » Sun Oct 07, 2012 11:18 pm

Hi.

Your results seem exactly what I'd expect, actually.

gecko wrote:But even if the laptop is contributing nothing, the i7-2600 is beating the crap out of the AMD FX 8120s.

Not that unusual. Historically AMDs can't hold a candle to Intels when we're talking number crunching. As a desktop user CPU they can scream, but even the current generation brand-new top-end bulldozer Interlagos still can't hold a candle to their equally powered Xeon brethren, even budget lines like SandyBridge-EPs... a quad of 6274s giving you a 64-core box will be about equal to a dual E5-2660 16-core box. I love AMD, but I still buy Intels. (gonna put two of these under my desk when I hit the lottery. And 32G 1066 RDIMMs are down to about $1k each! Hey, it's only money. :^) )

Also FWIW, since those FX linux boxes are just for render slaves and services, you might want to recompile your kernel with CONFIG_PREEMPT_NONE=y, CONFIG_HZ=100, CONFIG_HZ_100=y... disabling preemption and lowering the interrupt timer could eek out another ~8% for your renders, but don't do this if you want it to remain responsive under load. (or if you're a Ubuntu user, just install a "server" kernel) There might be other AMD specific system tweaks, and lux is an easy compile if you want to try any AMD optimizations...
User avatar
Heavily Tessellated
 
Posts: 183
Joined: Thu Oct 01, 2009 8:13 pm

Re: Optimizing Network Scalability

Postby gecko » Sun Oct 07, 2012 11:52 pm

Results are in. The AMD FX 8120's are performing at about half the speed of the i7 2600 both when running independently or running as slaves. This holds for total S/p output over 5min, as well as kS/s averaged over 5min. Network scaling is actually surprisingly linear, with each of the AMD systems performing very similarly. Network reporting on the master is very consistent with reporting from the individual slaves and only lags based on the last time of film transmission (and film transmission frequency does not appear to have a significant impact on slave or master performance). Just the highlights of an 86 data-point benchmark, let me know if you want any other details.

Looks like it's time to start overclocking my little render farm :/ If anyone has other ideas for why these AMDs are significantly under performing relative to the PassMark scores let me know.

@Heavily Tessellated - Not sure I'm up for recompiling (understand the concept, but it's a bit beyond my experience). I didn't expect the AMDs to match the i7, but I also didn't realize how far behind they were in brute computing power - when I bought them I assumed I was just looking to maximize flops, and that the AMD chips were behind on more "practical" computing exercises. Apparently I had that backwards. Nonetheless I'm still pretty sure I came out ahead on the cost per performance curve.

@SATtva - here are the kS/s averaged over 15sec at the end of a 5min benchmark.

i7 2600k (3.4GHz) running 64-bit Win 7 and Lux 1.0 (OpenCL version)
61.40 kS/s

AMD FX 8120 (3.1GHz) running 64-bit Ubuntu 12.04 LTS and Lux 1.0 64-bit (non OpenCL)
31.67 kS/s

Image
User avatar
gecko
 
Posts: 119
Joined: Sun Sep 30, 2012 2:32 pm
Location: USA

Re: Optimizing Network Scalability

Postby SATtva » Mon Oct 08, 2012 12:21 am

Really odd results. I have a feeling that GCC is producing suboptimal code for Bulldozer cores if not specifically tailored for this platform. I'm going to produce a FX-specific build for testing, i'd be glad if gecko could benchmark it. Just give me a day -- caught cold, feel bad.
Linux builds packager
聞くのは一時の恥、聞かぬのは一生の恥
User avatar
SATtva
Developer
 
Posts: 5931
Joined: Tue Apr 07, 2009 12:19 pm
Location: from Siberia with love

Re: Optimizing Network Scalability

Postby SATtva » Mon Oct 08, 2012 3:05 am

Ok, anyways... You can download the build from here. Dependencies and Lux were compiled with these CPU flags:

Code: Select all
-march=bdver1 -mtune=bdver1 -mno-movbe -mno-fma -mno-bmi -mno-tbm --param l1-cache-size=16 --param l1-cache-line-size=64 --param l2-cache-size=2048
Linux builds packager
聞くのは一時の恥、聞かぬのは一生の恥
User avatar
SATtva
Developer
 
Posts: 5931
Joined: Tue Apr 07, 2009 12:19 pm
Location: from Siberia with love

Re: Optimizing Network Scalability

Postby gecko » Mon Oct 08, 2012 6:49 pm

SATtva wrote:Ok, anyways... You can download the build from here. Dependencies and Lux were compiled with these CPU flags:

Code: Select all
-march=bdver1 -mtune=bdver1 -mno-movbe -mno-fma -mno-bmi -mno-tbm --param l1-cache-size=16 --param l1-cache-line-size=64 --param l2-cache-size=2048

Was really hoping that would fix everything . . . but sadly no, I'm still getting 32.11 kS/s averaged over 15s after a 5min run.

I did also note that the efficiency values are different - the AMD (both standalone and running as a slave) reports an efficiency of 1177%, while the i7 reports an efficiency of 1183%. Doesn't sound like a huge difference to me, but those numbers are very stable no matter how many times I run the benchmark or on which of the AMD computers. This is true for both of the builds I've tested. Is that expected?
User avatar
gecko
 
Posts: 119
Joined: Sun Sep 30, 2012 2:32 pm
Location: USA

Re: Optimizing Network Scalability

Postby Heavily Tessellated » Mon Oct 08, 2012 11:25 pm

I'm pretty sure you are getting expected results. Other than some system tweaking to squeeze a few more percent here and there. That's why so many people are mad at AMD over them. I know you want to keep trying to find where that one bottleneck is so they will be near/equal to the 2600K, like the numbers on paper say it should. This will not happen. I wouldn't spend too much time trying.

There are tens of thousands of hate pages for Bulldozer. There's plenty of bright geeks on anandtech and toms and such that will explain why they suck, to the point of boredom. (those L1 4 cycle cache latency discussions will never die.) Bulldozer will be one remembered as one of those huge letdowns that everyone waited for because the innovative design was(is!) real and the hype was so good... "Coming soon: Bulldozer! AMD's paradigm-shifting, ground-breaking, revolutionary new architecture!" So when it didn't deliver, it made it that much more painful. (never use Apple-grade hype unless you have Apple-grade fanbois - who won't complain, ever!) :^)

Not for nothing, but if you decide to OC them, keep an eye on the power consumption. 8120s are around 200w with all cores pegged, but that will shoot up to nearly 600w @ 4.6GHz. I don't know what your kilowatt/hour rate is where you live, but an extra 800w running 24/7 will make a difference. My rate is $0.23968/kWh elec+deliv, so that would be an extra $140/mo. Running both stock clocked ( w/ 16G & mid 1G GPU ) would be $39/mo. NYC electric is really high, but still, you don't want any surprises like that.
User avatar
Heavily Tessellated
 
Posts: 183
Joined: Thu Oct 01, 2009 8:13 pm

Re: Optimizing Network Scalability

Postby gecko » Wed Oct 10, 2012 7:56 pm

Heavily Tessellated wrote:I'm pretty sure you are getting expected results.

That's cool if I'm getting predicted results, I just want to make absolutely sure that's what's happening (as opposed to me doing something stupid, which has certainly happened before). As I mentioned previously, I was surprised at the difference between the PassMark benchmark and my Lux performance on the two systems, that's all. There's good reason I picked up parts for 3 complete bulldozer systems for the same price of a single i7 - this is AMD's third screw-up in a row, and as much as I appreciate Intel having competition, AMD has made a mess of everything they've touched since they bested the Pentium II chips. So I'm not shocked by the sub-par performance.

Anyway. Appreciate your concern about my electricity bill as well :) So far I've been able to pass this off to my wife as "oh, it doesn't cost that much". But I'll keep my eye on it. One of the nice things about living in the midwest though is everything is cheap, including electricity. But I do have expensive hobbies, what can I say. I already hand built a tile-inlayed wood and acrylic case for my render farm, so I think the electricity bill is the least of my concerns ;)
User avatar
gecko
 
Posts: 119
Joined: Sun Sep 30, 2012 2:32 pm
Location: USA

Previous

Return to LuxRender User Support

Who is online

Users browsing this forum: No registered users and 0 guests