Friday Facts #315 - New test servers

[ethernal]

RAM performance is affected by AMD Infinity Fabric clock. Basically if it's not on the same "tick" level (I think /4) the Infinity Fabric needs to overclock and slows down the whole CPU/RAM etc. as it needs to find the middle ground for both.

More info here and in other sources (google them please):
https://www.gamepressure.com/newsroom/h ... ance/z6e01

[Bilka]

Does the massive ammount of cores in your new CPUs imply, that the game now does significantly benefit from more than four cores?

No. Compilation is multithreaded compiler-side, and the tests are multithreaded by just running multiple instances, usually 1 fork per processor thread. You can see that here: https://www.youtube.com/watch?v=LXnyTZBmfXM

[BlueTemplar]

That is actually what I hoped they'd do to "discourage" turret creep. Let them have a bit of OOMPH when placed. Let it be folded up like those outrigger leg supports used on heavy vehicles like cranes. Have them appear all folded up, then let us see them unfold their legs from an upright position to down on the ground, then lift itself up to operational height. It takes negligible time when turrets are placed for defense, but they are vulnerable if deployed under fire.

Traditional turret creep works by building lots of turrets starting outside the enemy's aggro range where new turrets tend to be less far away from the enemy than older turrets.
One common property of turret creep is, that new turrets are protected by old turrets while they are build, so build time does not really matter that much (but squishiness of unfinished turrets does).
Multiple working turret creep patterns have been independently invented and successfully used by lots of players in traditional RTS games like Dune 2, C&C, and Warcraft 2.

Well, there's also "turret rush", when you run right up to a nest, drag a line a turrets, then drag-fill them with ammo, then switching to equalizing bullets/repair/grenades/fish for tanking.
It's often more resource (and time!) efficient, as you waste the minimum killing biters rather than the actual nests.
(A speed-unupgraded turret goes through its worth of piercing rounds in only 3.57... seconds, so losing a few turrets via rushing is less bad than it looks like.)

[rantoc]

I dont want to be a bish and spoil the fun with new hardware but - If it's write heavy and hw at that price range you should have considered the 970 pro's that have better endurance than the 970 evo's. Also the new line of cpu's is around the bend (halfpriced intel hedt's and tr3's along with 3950x's) so unless you need it _now_ consider returning it imo

That said - Awesome job with the game!

Regards / Mike

[Rseding91]

I dont want to be a bish and spoil the fun with new hardware but - If it's write heavy and hw at that price range you should have considered the 970 pro's that have better endurance than the 970 evo's. Also the new line of cpu's is around the bend (halfpriced intel hedt's and tr3's along with 3950x's) so unless you need it _now_ consider returning it imo

That said - Awesome job with the game!

Regards / Mike

We do need it now. We needed it about 7 months ago when 0.17.0 shipped and will continue to need it for the rest of Factorio development.

[Unclebod]

Also the new line of cpu's is around the bend (halfpriced intel hedt's and tr3's along with 3950x's) so unless you need it _now_ consider returning it imo

After having computers as a hobby for about 35+ years, I'd say that there is always something better around the bend, if you look 6 months ahead. If not something new, you will in most cases be able to get the same thing for a better price.

[Krazykrl]

RAM performance is affected by AMD Infinity Fabric clock. Basically if it's not on the same "tick" level (I think /4) the Infinity Fabric needs to overclock and slows down the whole CPU/RAM etc. as it needs to find the middle ground for both.

More info here and in other sources (google them please):
https://www.gamepressure.com/newsroom/h ... ance/z6e01

Most Ryzen3k chips can get pretty close to 1900Mhz fabric clock (fclock). Increasing the fclock grants pretty substantial performance gains, as long as you keep the fclock:uclock:mclock at a 1:1:1 ratio. Your ram should always be at a 1:1 (1:2 due to DDR, i.e. 1800fclock:3600RAM), not having it at a perfect ratio actually adds pretty substantial memory latency. If you do not use the 1:1 fclock ratio, to gain similar performance; you would need to tighten up timings, or go sub-ambient on the CPU, and practically shift that extra fclock into cpu clock; since ryzen does not do low temps with high fclocks that well.

Source: ryzen-3000-memory-benchmark-best-ram-fclk-uclock-mclock

[NimbleCZ]

@Bilka Did you try to use the fast RAM in the Intel PC? It would probably help as well.

[Oktokolo]

https://www.youtube.com/watch?v=LXnyTZBmfXM

I forgot about that video. Tests seem to be optimized for low memory use by using tiny map sizes. That way it indeed makes a lot of sense to run a lot of them in parallel.

[brucethemoose]

We now have 10 servers, some run tests, some do the deployments, etc. We are hoping to reduce this number in the future, as we can have 1 beefy PC do the work that is currently handled by multiple older/slower ones.

This is Prague. There is no server market or common set of companies which sell server hardware to buy from All of those machines are just desktop systems we've built from parts purchased online. If I compare the cost of the hardware vs what I know the same performance hardware would cost in blade form we got it for about 1/3rd the cost. It's a cheap game to play.

I've used server CPUs and I've used workstation CPUs and workstation CPUs win when it comes to the workload that we have. We've tested ourselves and had people test the same tests on their 48 core server systems with terrible results.

Sorry for necroing, but I'm just now starting Factorio, catching up on recent dev blogs before I go to sleep, and just registered, so I can't PM

I'm not sure how much consolidating y'all want to do, but Newegg's Czech site sells a 7702P + mobo combo for a almost reasonable price with shipping:

https://www.newegg.com/global/cz-en/sup ... 6813183689

Thing is, it's regular ATX instead of some huge server form factor, so it'll fit in a regular PC case and work with a small ATX power supply. And in spite of SuperMicro's claims, I believe it'll take regular DDR4, though I'm not 100% sure on that. So what you could theoretically do is rip out the mobo from one of your old machines, slap that in, maybe buy some more matching RAM, and have it do the work of several machines.

You could probably overclock the i9 pretty hard and get similarly great results, but I don't know how expensive electricity is in Prague. It doesn't seem to hot in the winter...

I'm not sure what test you were running on the 48 core (2P Xeon? Or, god forbid, 4P Opteron?) server, but some workloads don't like 2P CPU setups. A single socket like that could be less problematic, though it might not matter if you're talking about that instanced test.