Re: [DCP-o-matic] Benchmarks/network coding

Attached two pictures with recent benchmarks, using 1.69.xx. One single machine testing, one network. The single machine lists only the peak performance runs, as most machines were tested with different encode thread numbers to find the best setting. It can be seen that 'overthreading' will help a lot on current multicore CPUs. Be aware, though, that this might crash the software if running under WIN32 OS. I can easily be seen that the price/performance sweetspot is with the 6core 3930k/4930k machines. They can even be overclocked to 4.5GHz. This CPU costs only around 500€ or so. This is also reflected in the CPU Passmark list: http://www.cpubenchmark.net/high_end_cpus.html I was always skeptical about network encoding with regard to achieving very high conversion rates, because I thought a typical network (even gigabit) would saturate too quickly, already around 10fps or so. I was wrong. I set up a test at a clients site with a couple of (mostly) older Macs, connected through a simple Gigabit switch. Nothing really fancy. Two iMac-i5 were the most beefy machines. The network still outperforms the fastest (and most expensive) dual Xeon machines. These tests also show that the J2k coding on the Mac is at least as efficient as under Windows. Although I still have to perform the identical test on a Mac with OSX vs Bootcamp one time. The network test was done with all machines running OS X, adding one machine by one and running a BigBuckBunny encoding for each set. And of course this was using Carls standard Bunny benchmark metadata, so 2k - the network load may look VERY different when using 4k... It's interesting to see that the 'Master', an outdated Xeon 3530, is able to supply so many frames to the clients and is still encoding many frames itself. Only with the last iMac-i5 added to the network I could see the occasional drop in CPU load on it, but very short and rare. I guess I could have gotten near 20fps by adding another. The number is missing in the table, but aggregated single CPU fps for all machines used is around 19 - that's just 2fps loss in the network coding performance. Good work, Carl! Also, not a single run exhibited problems. No crashes, no lost frames, etc. I also did a SINTEL run on the same network, giving the same performance (17.3fps) and taking just 20min runtime. Carl - what is now done on the render clients - only J2k, or also colour conversion, scaling, etc.? _______________________________________________ DCPomatic mailing list DCPomatic(a)carlh.net http://main.carlh.net/cgi-bin/mailman/listinfo/dcpomatic