Questions about optimizing game server performance on a VDS

[Patriot]

When to Increase the amount of Physical Memory?
My 3-core VDS currently has 3 Gb of RAM with 50% of that used by three L4D2 game servers and the OS (Win 2003 Server). Each game server currently uses close to or just over 300Mb or RAM. If memory use for each incremental L4D2 server is linear, I should be able to add another 5 servers before reaching 100% use of physical memory.

Other posts have talked about keeping CPU load below 50% on average. At what point will memory usage negatively affect game server performance, assuming a CPU load below 50%?

Dedicated IP or One IP with multiple Ports for a Game Server?
Is there any performance impact on a game server when using a dedicated IP versus sharing an IP and using different ports? How many ports can be used for game servers on a VDS? If using multiple ports, does it matter if the port numbers are in sequential order (i.e. 27015, 27016, 27017, etc.) or not?

Maximum FPS for L4D2 servers?
My three L4D2 servers all run at 400+ FPS on Windows 2003 Server. Given that L4D2 is somewhat unique, can the FPS be further increased with Windows 2008 and HPET as with other Source games?

Thank you,

Patriot

[Edge100x]

When to Increase the amount of Physical Memory?

Generally you should upgrade your memory when you have less than around 300 MB available. This minimum amount is needed for the disk cache to be effective and for Windows to be able to satisfy temporary memory requests.

At what point will memory usage negatively affect game server performance, assuming a CPU load below 50%?

The two are mostly independent.

Is there any performance impact on a game server when using a dedicated IP versus sharing an IP and using different ports?

No.

How many ports can be used for game servers on a VDS? If using multiple ports, does it matter if the port numbers are in sequential order (i.e. 27015, 27016, 27017, etc.) or not?

You can use as many ports as you'd like, but keep in mind that our default DDoS mitigation rules only cover common ones, like 27015 to 27030. If you use multiple ports, just make sure that your servers don't step on each other, because some games (like Orangebox) use multiple ports per instance. Check which ports each instance is using with "netstat -ano" output.

My three L4D2 servers all run at 400+ FPS on Windows 2003 Server. Given that L4D2 is somewhat unique, can the FPS be further increased with Windows 2008 and HPET as with other Source games?

Yes, I would expect it to be able to be increased to 1000 under Windows 2008, as with other Source-based games that use the older tickrate/FPS model. However, L4D2 has such a low and locked tickrate (30) that the FPS doesn't matter very much; you may actually have better results (lower CPU usage, for instance) from running at a lower one.

[Patriot]

Edge100x,

Thank you for such a quick and thorough reply! The service at NFO is really outstanding.

Best regards,

Patriot

[Patriot]

Edge100x,

I have a a followup question about memory resources and use after taking a closer look at Task Manager.

Attached is screen showing resource use on my 3-core VDS with 3 full L4D2 servers running. Can you explain how to interpret the page file use versus physical memory (total, available, and system cache) and offer guidance on how to optimize memory resource usage, such as with settings in the server OS or hardware upgrades (ssd, additional physical memory, etc.)?

Thank you,
Patriot

[.=QUACK=.Major.Pain]

Your graphs look ok.

We run ours pretty hard and have no issues.

3 core VDS with 1gb upgraded memory.

Runs: 4 L4D2, 1 TS3, and 5 Procons.

[Patriot]

Your graphs look ok.

We run ours pretty hard and have no issues.

3 core VDS with 1gb upgraded memory.

Runs: 4 L4D2, 1 TS3, and 5 Procons.

Hi Major,

Thank you for the feedback. It's helpful to know what others are running on comparable servers.

Not to get off-topic here, but if you run L4D2 versus servers with SourceMod plugins, have you ever experienced an issue with missing health packs and ammo during the first round when joining from a lobby?

Regards,
Patriot

[TimeX]

Not too long ago I talked to a customer at great length trying to troubleshoot that very problem, but I don't recall the outcome. I'm not sure if that was you perhaps, but I can spend some time later tracking it down again if it will help.

[Edge100x]

In terms of your resource usage, I agree that you appeared to be doing very well at the time that screenie was taken. Your overall CPU usage and CPU usage per-core both appear to have been low, and you had ~1.2 GB of available memory. Your peak commit charge number tells us that at one time your servers had requested ~2.61 GB of memory, which got close to your 3 GB limit (if all of it was in physical memory), but it did leave some buffer.

[.=QUACK=.Major.Pain]

We run Win2008 compared to your Win2003, and we don't run any plugins on our L4D2 servers. We found people just prefer to play the custom maps and the plugins just cause extra load.

[Patriot]

We run Win2008 compared to your Win2003, and we don't run any plugins on our L4D2 servers. We found people just prefer to play the custom maps and the plugins just cause extra load.

Major,

Thank you for the additional details about your servers and OS.

Patriot

[Patriot]

Not too long ago I talked to a customer at great length trying to troubleshoot that very problem, but I don't recall the outcome. I'm not sure if that was you perhaps, but I can spend some time later tracking it down again if it will help.

TimeX,

Thank you for your offer to track down the issue. I don't think it is necessary at this point, however.

Patriot

[Patriot]

1. How are you assessing per-core CPU usage from the screen shot?

2. Could you please comment on Page File usage and how it should be interpreted?

[Patriot]

Edge100X,

On Windows 2003, what it's calling "PF Usage" is mislabeled; it's actually the commit charge, which is the total amount of space that all processes on the server have requested. Not all of this memory will actually be used by the processes after they request it, so you may not need as much memory as the commit charge would seem to indicate.

Thank you so much for the clarification. I was thinking "PF Usage" represented the virtual memory paging file, which is hard disk access and much slower than physical memory access. Hence my mentioning an SSD, thinking if the server is using virtual memory better to have it on an SSD.

Patriot

[Edge100x]

I accidentally edited your post instead of replying before, it seems! I was likely trying to edit my own, right after posting it.

I'll edit yours back. Here's what I said in my accidental edit, that you just referenced:

1. How are you assessing per-core CPU usage from the screen shot?

Windows breaks it out per-core there. If you don't have processes assigned to specific cores, we may be seeing some of the usage of a particular process assigned to multiple cores and mis-represented low numbers as a result, but your overall usage is lower than a single fully-utilized core would represent (33%) as well.

2. Could you please comment on Page File usage and how it should be interpreted?

On Windows 2003, what it's calling "PF Usage" is mislabeled; it's actually the commit charge, which is the total amount of space that all processes on the server have requested. Not all of this memory will actually be used by the processes after they request it, so you may not need as much memory as the commit charge would seem to indicate.

[Edge100x]

Thank you so much for the clarification. I was thinking "PF Usage" represented the virtual memory paging file, which is hard disk access and much slower than physical memory access. Hence my mentioning an SSD, thinking if the server is using virtual memory better to have it on an SSD.

Nod, it is confusing that they refer to it like this.

You are right that "hard faults" (when the OS has to retrieve a page that was previously pushed to disk) do cause a serious performance hit. It's more difficult to see the count of these on Windows 2003 than in Windows 2008+ because there's no easy system counter for them, but you can tell approximately when they start because you will see available RAM drop to under 200-300 MB, simultaneous with a high commit charge (higher than the total RAM in the system) and heavy disk access.