Republic of Folk

So I don't know how our current hosting works, but we might be able to get a better deal by using a google cloud server that turns on and off as needed.
Basically there would be a small server that acts as a controller. This controller could host anything we want except the arma 3 server itself.
There would then be a web interface which would then activate a larger server until given the command to turn off or after a certain duration.
On server start it would start the arma server(s).

Thus in effect whenever someone wanted to use the server they would just go to the web page hosted by the controller and activate the arma server.

The potential benefit to this scheme is that we could get a higher powered server for cheaper by running it only when needed. The downsides are that pricing would be variable, the
server wouldn't always be available, and we would only have CLI access to the machine.

This scheme would cost be an estimated cost of $13 (US) to $24(US) per month depending on how exactly we set it up. If it is only $12 per month I would be glad to cover it myself.

Price details:
The controller machine would be a f1-micro machine (0.8cent US per hour). (6$ / month)
The arma machine would be either a n1-standard-2 (14cent US per hour) or a n1-standard-4 (27cent US per hour). (6-12$ / month at 5 hours per tuesday+sunday sessions)
There is the additional cost of storage (80cent US per month).

Technical details:
The controller server would be hosting a simple web server which would make a REST or CLI call to the google API to activate/deactivate the arma 3 server instance.
The arma 3 server would activate on instance start (and potentially even fill its mpMissions folder from a directory on the controller server).
The arma 3 sever could potentially be put into either a mission making mode or a live mode so that we can control the admin password.
If the arma 3 application server crashes then it could just be restarted using the controller's web interface to stop and then start the machine.
We might have to route all network traffic going to the arma 3 server through the controller server because assigning a static public IP to the arma 3 server costs 6$ per month. (A used public IP costs nothing, but an IP attached to nothing costs money).
I don't know if we currently use a headless client, but there would certainly be enough computing power on a n4 server (and probably enough on an n2 server) for one to be hosted locally.


At this point I have determined that this is all possible to do. I have hosted an arma 3 server (but not the headless client yet) on google cloud services and successfully ran a mission on it.

Thoughts?
Raptoer

Just a quick reply, not in the role of a host, but as someone who's fairly familiar with Arma servers, our server setup, and what demands our playercount imposes. I'm also not touching on the budget concern except by noting that getting the money together has never been an issue during the history of FA or its predecessors.

First, performance. If there are any other groups with our numbers doing this (or using any kind of non-dedicated setup), I've not heard of them. Admittedly I've not been following the A3 server business in a few months too closely so I'd love to hear about it if someone's doing it.

Getting the A3 server to run on Google Cloud is far cry from it actually being useful for server hosting. Playercount and the number of active AI are the key factors for performance. Current playercounts in coop pushes our current dedicated server (Xeon E3-1245 v2 IIRC) to around 30-40sfps on average. That'd be the necessary guaranteed performance baseline. Virtualization always incurs some overhead and A3 is very CPU-heavy, primarily single-thread (as in if the main thread starves, nothing else happens either) workload. There can't be any latency or fluctuation in computing resource availability. The A3 server is slowly getting better with multithreading though so I wouldn't go with 1 or even 2 physical cores. A HC is less CPU-heavy but still noticeably so, I wouldn't run one in the same environment as a server without enough physical cores being available so as not to overlap.

Requirements on the networking side aren't too huge, but latency needs to be constant and minimal. The way the A3 server handles networking makes everyone suffer from any bad connections. We don't police players on this so server has to be up to snuff.

Second, platform-wise Windows (with RD access) is pretty much a necessity. BI hasn't made any promises as to simultaneous availability of Windows and Linux releases, and this has been an issue in the past. AFAIK Google's Windows support is still heavily beta.

Summa summarum: this would need testing with a realistic playercount to assess viability. A full 40+ player public server running a coop mission perhaps?

I thought we had problems with server fps, especially near the end of missions, but that might just be inevitable no matter what we do.

Now that I hear what the server uses, it sounds like this might be at best a marginal increase in computing power (gcs uses Xeon E5s), with an unnecessary decrease in price for a very large increase in complexity.

It's still a good thought, and thanks for the suggestion. The tech world changes fairly rapidly, so there's definitely a point to helping the hosts ensure that we're pursuing the best solution for Folk ARPS at a particular given time. If you (or any knowledgeable comrade, for that matter) have an idea that you think merits consideration, we're always reachable either through Skype, forum, or PM.

As for SFPS: There are some things that mission-makers can do to mitigate SFPS problems, such as taking advantage of F3's caching function or guarding against move-order spam by the AI or not placing too large a group in too confined an area... but a large portion is outside of our controls as well - ArmA 3 is constantly iterating, and not all of its iterations play well with others. Witness the workaround we're currently forced to use so our loadouts load properly.

~ Ferrard

Just taking a quick peek at the AWS instances (this is my day job but we don't normally use windows AWS instances), the i2 variant (which would be the one I'd pick if I was running an Arma server), running windows with eight cores, 61Gb RAM and a pair of SSDs in RAID0 would run to about $2 per hour. That might be slightly faster than the dedicated Folk server, but I don't know it'd run to less money over a year.

AWS has some fantastic uses, but I'm not sure that they can outdo a dedicated hardware box for gaming servers just yet.

(Mind you, the testing of that would be fun )

Sparks, Raptoer or anyone else, do you have any generally accepted wisdom, personal experience or any neat benchmarks on how much the virtualization layer and other cloud stuff affects these kinds of workloads, compared to running the same stuff natively? This is well out of my area of expertise so I'm curious about how close the performance is these days.

It's not exactly a simple question - AWS is set up for a completely different sort of thing (web applications generally) which aren't anywhere near as demanding of the system as FPS game servers like Arma would be. And you're into a mess of marketing benchmarks if you look at most of the "studies" that have been done. I've seen estimates that say that bare metal is 2.6 times faster than an equivalent AWS instance. For web apps, no big deal, it's a bullshit number (modern CPUs blew past the level of performance needed to run a web app quite a few years ago) but for games, where every cycle is in use, that number might actually be in the ballpark.

The whole point of AWS, the advantage of it, is the ability to do some crazy stuff like pushing a button and bringing up a thousand machines from out of nowhere, configuring them automatically, linking them into an application, and letting a company that's suddenly found itself with five million new users in the last half hour scale up to meet demand. So raw performance isn't actually their primary goal.

In terms of price/performance, somewhere like Hetzner that will sell you a colocated dedicated box in their datacenter are probably still the best option.

I just thought the AWS idea was nifty and in a few years, when their instances get faster, the performance thing may well no longer be an issue and it just becomes an economics thing.

Where I think things will get really interesting will be something like Arma 4 or 5, when they write in support for cloud virtualisation on the server side from the get-go. Session has ten more people than you expected? No worries, we'll just fire up another AWS instance, you'll be back up to 60sfps inside of ninety seconds...

Initially I considered AWS, but there is a little catch in their pricing. If you have any dedicated instance running then it costs 2$ per hour + the instance costs. GCS has nothing like that.
I can't find anything on the relative speed of bare metal vs virtualization but I have a hard time believing this 2.6x number. I think nobody in the cloud computing world cares about per machine performance because they live in a world where you can just add more machines to a cluster.

I use AWS at work, but we never tested its machine performance because we are bound by our database speed rather than application sever speed.

Our best performance test might just be to start up an arma 3 sever with a few hundred AIs and see how it responds. I still have both accounts set up with their servers turned off, I just need to make sure I have ASR AI installed or else it won't be a valid comparison to the current server.