Performance of Server 2008 Standard vs. WinXP vs. Ubuntu Server 8.04LTS

[mr.g-c 6]

Having all systems on one Server, here are my short conclusions performance-wise to help with people chosing the "right" OS for their Arma Server.

Arma-Server V1.16 on the Windows Machines, V1.14 on Linux.

One and the same mission with equal steps and 100% reproducibility everytime.

It goes like this:

Mission start, one player connected, taking FPS and CPU Utilization.

Then 16 Players after 1-1.5 hours of constant Gameplay, FPS and CPU Utilization.

1. Mission Start, one Player connected, 5 Mins waiting then following Data:

- Windows XP SP3 x86 : 09-16% CPU Utilization, Server-FPS 47

- Windows Server 2008 Standard x86 SP1: 22-28% CPU Utilization, Server-FPS 47

- Linux Ubuntu x64 8.04LTS: 12-15% CPU Utilization, Server-FPS 48-61(link!)

2. Mission played for 1-1.5 hours and 16 People playing:

- Windows XP SP3 x86 : 86-94% CPU Utilization, Server-FPS 46-47

- Windows Server 2008 Standard x86 SP1: 99% CPU Utilization, Server-FPS 42-45

- Linux Ubuntu x64 8.04LTS: 95-100% CPU Utilization, Server-FPS 46-51

Conclusion: Win XP runs best. Who needs bloated Win2008 Server anyway for only running a lame Gameserver? I can't see the need of a Domain-Controller or whatever role here and WinXP is thanks to NT-Kernel really Stable.... I remember Uptime for ~100 Days and more here at home with it.

And you can make it almost as secure as you can make a Windows Server.

[domokun 515]

Very interesting. Thanks for this research.

BTW what kind of hardware makes up your servers?

[mr.g-c 6]

CPU: Intel E8400 @4.0GHZ

4GB DDR2- 1066 RAM

Rest is merely irrelevant...... :P

[Auss 208]

Why would you want to run a game server on a domain controller..LOL

I actually came to the same conclussion you did many months ago

http://forums.bistudio.com/showthread.php?t=70187

[sickboy 13]

Interesting. I'm a huge fan of Windows Server 2008, evenmore about the R2 edition (windows 7 server), though not specifically for Game Servers.

Though all my game servers run on them. I'm running Hyper-V hosts, with Virtual Machines inside that run the gameservers.

Piece of cake for me to setup a VM with Windows XP and run ArmA inside it, after seeing your results, I think I will, even only for testing :)

It would be interesting to know what was causing the higher cpu load aswell as the lower FPS on the 2008 installation.

For all other purposes than gaming, I see improved performance compared to Server 2003 for instance.

Nice overclock btw :)

If you are still able to test with the 2008 Server OS, I would love to see the result after you disable the different tcp/ip enhancements 2008 has enabled by default. (With certain hardware, these features completely mess up network traffic :P):

(Run from Elevated Command prompt)

netsh int tcp set global rss=disabled
netsh int tcp set global autotuninglevel=disabled
netsh int ipv4 set global taskoffload=disabled

optionally you could try this one too:

netsh int tcp set global chimney=disabled

You can review the default settings if you like with these commands:

netsh int tcp show global
netsh int ipv4 show global

[Auss 208]

Nice overclock btw

Agreed. The E series is a wonderful cpu to OC. I have my gaming rig E8500 running stable @ 4.22 Ghz, Stays cool even on hot days, I highly recommend them.

I'd like to see your results if u could setup an XP box in a VM session

[sickboy 13]

Agreed. The E series is a wonderful cpu to OC. I have my gaming rig E8500 running stable @ 4.22 Ghz, Stays cool even on hot days, I highly recommend them.

I'd like to see your results if u could setup an XP box in a VM session

NP. Might have some time in the next days :)

[xeno 234]

@mr.g-c

Is that 32 Bit or 64 Bit 2008 server ?

Am running a 2008 64 Bit server and I hardly get more than 15 % cpu load, even after playing for a longer time (server cpu is a Core i7 920).

The only thing I'm facing is that 1.16 brought back some lags, dunno why yet, 1.15 was fine. And a RPT error message I haven't seen before: "Warning: Delete out of order".

Other than that 2008 is just nice.

Btw, who will run XP if he has more than 4 GB RAM ? ;)

Xeno

[mr.g-c 6]

Hi, Both Windows are X86 (32Bit)and running in Virtual-machines as Guest systems from within the Host via KVM - Kerne Based Virtual Machine .

Its Fully Hardware Virtualized - No Performance loss whatsoever at all, people who know KVM and the Processor-Specific technology behind it, knows this.

Of course i tested this all at Home on my Server here to make sure there is really no performance loss at all, by installing the system one time "real" and one time as "Virtual machine" and tested it woth Super-Pi and SiSoft Sandra Benchmarks.

Btw, who will run XP if he has more than 4 GB RAM ?

I only gave the Guest OS' 2GB of the available RAM for testing, since it never needs more, as the Arma Server hardly gets above 500-600 MB RAM allocation.

My Servers Host OS is Ubuntu Server 64Bit.

Am running a 2008 64 Bit server and I hardly get more than 15 % cpu load, even after playing for a longer time (server cpu is a Core i7 920).

Yes because you have 8 threads with the Core I7.

When i have written down my CPU Utilization, it was meant from the Process itself (not the overall CPU Utilization!) and for One Core/Thread of course, i hope you know that Arma-Server only uses Single-Core?

It the same when i run Arma or Arma-Server on Quadcore and say "I never get more than 25% CPU Utilization", or on Dualcore and say "I never get more than 50% CPU utilization".... get my point? :p:D

You might set affinity to one core/thread and then observer that thread/core alone to see how much the Arma Server utilizese it, or you simply multiply Utilization of Arma-Server.exe Process times the cores/threads you have...

The only thing I'm facing is that 1.16 brought back some lags, dunno why yet, 1.15 was fine. And a RPT error message I haven't seen before: "Warning: Delete out of order".

Same here, although i'm convinced they were introduced with 1.15b, i noticed them quite often when playing on 1.15b servers which were running on 1.14 before.

I call them "Random Lags", they appear out of no reason in middle during game and suddenly all get a yellow/red chain for some seconds, but there is no "real Lag" like it was with the older Versions, because everything is still moving around, IE no "non-moving" vehicles/persons with spinning wheels or walking animations.

@ Sickboy, yes i know of the tried Optimizations with Networking in Server 2008 since i use it here at home also.

But i have to say Network Performance is absolutely perfect so far, we get full Host-OS speed with the Virtualized Network (Virtual Bridges with TAP/TUN Devices) and para virtualized "Virtio" Network Drivers.

Also do you really think changes on this would reduce the CPU Utilization of the Arma-Server.exe Process itself?

Because if those default Network settings drain on CPU, they would rather cause Utilization on other Windows-Core processes, right?

Agreed. The E series is a wonderful cpu to OC. I have my gaming rig E8500 running stable @ 4.22 Ghz, Stays cool even on hot days, I highly recommend them.

Absolutely! Even more when you have your server standing in a Datacenter, since you can install then even the loudest Cooler, no-one cares :p

4GHZ for the E8400 we use is already quite much i think, but it runs absolutely stable (tested 48Hrs with CPU-Burn for Linux, worse than Prime in case of Heat emitation/Energy consumption).

Edited April 30, 2009 by mr.g-c

[sickboy 13]

...

All guest drivers/services etc up2date and working in all Windows versions?

I have no experience with KVM, Hyper-V I do, but you need specific drivers to take advantage, at least for storage and network. (Which im sure you know :))

The problem with the network optimizations is that I have seen it cause the most weird issues with networking, or even general system performance I have ever experienced. Sometimes reproducable, other times not.

The general consensus around the web afaik is this:

"If you have unexplained network issues, or weird system behaviour, or performance issues; Better disable the network optimizations in 2003 R2 and newer, big chance it was the cause" :)

I'm interested to know what the harddrive performance is under xp and 2008, under KVM, and what the "system" process load is during these tests (in the VMs).

Though i'm sure you have better things to do :)

Edited April 30, 2009 by Sickboy

[mr.g-c 6]

All guest drivers/services etc up2date and working in all Windows versions?

Sure.... KVM/QEMU emulates different devices such as Network Cards/Graphics Cards/Sound Cards, whatever and you can chose them, but most of them are "older ones" and Windows has drivers for them always included.

I have no experience with KVM, Hyper-V I do, but you need specific drivers to take advantage, at least for storage and network. (Which im sure you know :))

Sure i know and i use them... its paravirtualized Network Cards and Block Device (called Virtio, included in kernel 2.6.25 and upwards).

Speed is i would say almost 100% as if it were all "real devices"... in my thorough testings here at home, i could not find any downsides at all with these combos, besides the "thinking-mess" with the virtualized, and thus non real existing network(s) in the host OS , lol....

I had written down small topology plans, to get/remain a overview which VM is connected over which Virtual bridge to which virtual switch and to which physical ethernet card (if you have multiple ones on the MB). :rolleyes:

The problem with the network optimizations is that I have seen it cause the most weird issues with networking, or even general system performance I have ever experienced. Sometimes reproducable, other times not.

The general consensus around the web afaik is this:

"If you have unexplained network issues, or weird system behaviour, or performance issues; Better disable the network optimizations in 2003 R2 and newer, big chance it was the cause" :)

This is interesting of course... Sure as soon as we run (hopefully not) into any Network performance issues, i'll remember back to your advice... but as long as it works i will not change it.

I'm interested to know what the harddrive performance is under xp and 2008, under KVM, and what the "system" process load is during these tests (in the VMs).

Though i'm sure you have better things to do :)

With the default emulated IDE-Controllers, there is only Multiword DMA available so far, so there was of course some CPU load when extracting/coping larger files.

However the speed is awesome! I was installing ACE 1.07 Full on the Virtual Machine XP and it was certainly way quicker installed than on my Real Vista Computer here at home (having 3.4GHZ Q6600 4GB Ram). Also as i moved the whole Arma patches (over 1GB) from one virtual image to a second one (with Server 2008 running this time) i got ~26-28mb/sec, which is quite nice, given the fact that basically both Images are stored on the same physically existing Disk.

All in all i'm quite happy with KVM and the Full Hardware Virtualization, i would have never thought that its running that stable/great and performant...

I couldn't believe my eyes as i started XP VM first(after all install was complete) and it took only 3-5 Seconds to boot until login-screen (No Joke!):eek:

EDIT: To come back to the actually topic and forget that VM mess :P :

In my testings Win XP was almost 2 Seconds quicker in calculating 1M Digits in Super-PI test than Win 2008 or Vista (i done it here at home only)- both in VMs and in the real installed systems.

But why the CPU Utilization of Arma-Server.exe is so worse in 2008 than in XP, i don't know really.

I think its the same as i noticed my switch on my Gaming Rig from XP to Vista, in which i got considerable worse FPS (10-20%) in Vista, than i had in XP. For single-Core optimized Software, XP seems to be still the system of choice if every percent speed is important.

I think we will stick to Server 2008 though, it simply feels better

Edited April 30, 2009 by mr.g-c

[HOPEnSPIRIT 10]

I have one problem with this test AI.

What mission was running was there any AI?

As AI takes up CPU and you could have the same mission on both servers but usage will be different depending on how many AI are active and what are they doing.

My suggestion would be to run Sector Control with Purely human players; have everyone connect load the mission and then monitor CPU, then see what the results are.