Setting up a community dedicated-monster-of-a-server

pfff @cola_colin forget me much ?

of course I'm down for this!!
i'll chip 20$ in!

@9'35
So lol
much color confusion

I think we should push for a system with more RAM like 64gb or 128gb for massive matches.

With 20+ people queing commands and spamming units RAM will go very quickly.
And swapping will kill the CPU too, as it has to redirect info from RAM to HDD and then Load data From 2 Locations instead of 1, so no matter how fast the CPU is once swapping occurs the game will suffer dramatically.

@cola_colin It's super easy to proclaim yourself a business (or small society) and then have access to buisness-class internet plans (symmetric up-down bandwidth with crazy speed).

The investment for the PC would be the easy part. maintaining donations for the internet subscription the hard part.

but it's reasonable 45-65 € per month for 200-500M symmetric.

worth it? yes.

You can't get better internet without putting better cables into the ground. That's not about a subscription. More about a huge investment into fibre cables.

Same reason I'm forced to use satellite essentially. :/

I have fibre leading to my house as of T-9 days.

I have a mobie phone with 3G

@maxpowerz

well you lucky ***

admit it, my plan is a better deal for all of us.

@cola_colin http://blog.gsmarena.com/sk-hynix-develops-128gb-ddr4-ram-module/ : D

Gloater

With 3g Amazingly i was able to play PA online with minimal (Constant) lag and almost no dropouts (everygame)

But seriously how many people are we looking to get in one match??
We do have to keep it realistic to some degree.

@zihuatanejo nobody noticed my disappearance from the forum for three days? internet was out as one of the 4 internet providers was setting their fiber cables into the ground in my street and accidentally severed me and a bunch of neighbour’s connection. it's likely to happen again as the three others come in the street over the next couple of days.

@maxpowerz personally if I were to run a server I wouldn't do it unless I’d found a way to cap the joinees. I want to be able to test what runs and what doesn't and keep it in the "run" zone. even if that means implementing a waiting line or whatever other method of filtering.

a single machine (as far as i understand) runs 1-2 games currently. there's only two threads per server so it'd be a waste of time to use anything other than racks right now.

which is why I'd really love for uber to work on multithreading because then we're not talking the same numbers of people you can run on a single machine. and it could go up exponentially really fast.

Neither throwing in more memory nor faster CPU clock will help you for now.

In order for the server to run significantly faster, both memory space efficiency, and parallelism has to improve first.

More memory can help you avoid swapping from masses of units with individual goals (hint: shouldn't happen any more since the area patrol revamp), but won't help you much otherwise since the performance was already subpar even before swapping started.

Higher clocked CPUs won't help you much either. Best you can do, is double the clocks with heavy water cooling and hand selected CPUs. Won't be enough since a 100% performance boost isn't going to cut it. Increasing parallellism and enhancing throughput by reducing the working set with the goal of levering cache eviction is what's going to help you. You *could* try to aim for super clocked RAM (low latencies, higher burst bandwidth) instead, I assume that PA server executable is currently neither ALU nor FPU bound, but by memory bandwidth respective memory latency. Not surprising, considering the enormous memory footprint and the data-structures used (at least the ones we know about).

What lead me to the latter assumption? Well, since Uber decided to tessellate world boxes of up to 3000 (max. size planet + "orbital box"?) units into 4 unit wide boxes for pathing and LoS tests, that means up to 10 recursive steps in dereferencing a specific box width only mediocre caching effects and little to none chance for neither the compiler nor the processor to take shortcuts. Would be possible to bypass that by using a hybrid between a double linked 3D grid (only nodes on same level are linked) and an octtree with an clever LRU list, but that would currently hurt even more since it would cost another ~300-500MB/planet for 8 additional pointers per intermediate and leaf node.

Sooo... We'd get to have super big servers at the cost of maybe $10 a month? Hmmm... I'll think about it.

Ha ha. That just happened to me. Some dude was puting in a fence and hit the cables. On 16 hour downtime though.

So we just played a 10 planet 32 player game (http://pastats.com/chart?gameId=226240) on this amazon server: c3.4xlarge (55 ECUs, 16 vCPUs, 2.8 GHz, Intel Xeon E5-2680v2, 30 GiB memory, 2 x 160 GiB Storage Capacity)

Amazon has some graphs:

At the end of the game the server has used 25gb of ram, the sim speed was lowest at ~2 fps so 1/5 of normal speed.
Interesstingly enough the server used up to 9 cores (peak of 860% for the server according to top). So it does multithreaded a lot of things it seems. Just the simulation seems to be single threaded :S

A fast 4 core cpu might not be the optimum, more like a fast 8 core cpu xD Maybe even more cores than that. Not sure.

Also can somebody make sense of those network graphs? 250000000 bytes per second seems ... unrealistic? maybe that's mbit?

2000kbs per person at start, ended at 8000kbs per person

edit forgot a 0
those numbers do look a bit big