Gaming on RAMdisk: Some Benchmarks

[XLNC]

"8GB or 16GB?" seems to be a popular question these days. With RAM price so low, I was very intrigued by software RAMdisks for gaming use after seeing this old thread. However, most forum threads I found about this topic ended with someone saying "try it and let us know." Well, I went ahead and tried it with the ever popular Crysis 1. Results are below.

System:
Core i5 2500K @ 4.5 Ghz
Asus P8P67 Pro
MSI Twin Frozr II GTX 570
16GB Ripjaws DDR3 1600 RAM (8GB setup as RAMdisk)
Samsung 470 128GB SSD
Samsung F1 1TB HDD
Dell U2410 @ 1920x1200

Setup:
I installed the game on my SSD first. Then, I used Dataram RAMDisk ($15) software to setup an 8GB RAMdisk. This software supports backing up and restoring contents at shutdown and startup. Using symbolic links, I redirected the entire Crysis folder to the RAMdisk. Additionally, I repeated this with my Samsung F1 HDD just for comparison.

Testing Method:
On each drive, I launched the game and loaded my most recent checkpoint. I timed it from hitting "Load" button to seeing "Press any key to continue." I repeated the test three times. Between each test, I rebooted the system so none of the game files were cached in memory. Load times were mostly consistent between reboots. Averages are shown below.

Crysis "Recovery" Level Load Times (average, in seconds):
HDD : 41.7s
SSD : 37.4s
RAM : 26.9s

So the RAMdisk was faster as expected, but "only" 28% faster than an SSD, and 35% faster than the F1 1TB. I was hoping for 100-200% faster. This is still a huge advantage if you're playing multiplayer, but without looking at the stopwatch, it was hard to tell the difference between SSD and RAMdisk. However, the HDD felt awfully slow since it took an extra 5 seconds after reaching 100% to be ready for play. I also ran the famous "Island" benchmark to see if the RAMdisk had any impact on frame rates. Here are the results below.

Island Benchmark - 1900x1200, No AA, 3 Loops, High Quality
Min FPS (same for all 3 loops)
HDD : 7.94
SSD : 7.71
RAM : 11.75

Avg FPS (Loop 1 - Loop 2 - Loop 3 - Avg)
HDD : 45.19 - 65.63 - 65.65 - 58.82
SSD : 48.11 - 65.75 - 65.62 - 59.83
RAM : 50.15 - 64.92 - 64.87 - 59.98

Subjective Impressions:
The RAMdisk seemed to help somewhat in frame rates, but again, not significantly. Besides numbers, I didn't notice a huge subjective difference in this game. Any improvement I perceived can easily be attributed to the placebo effect. I noticed occasional texture pop-in from all three drives, the RAMdisk was not immune. You can hardly conclude anything from two tests on a single game but embarrassingly, Crysis 1 is the most recent PC game I own. All I have left are ancient Source based games. Will test newer games when I get them and update this thread.

Hope this post helps anyone contemplating a RAMdisk. Comments?

 

[ncalipari]

ramdisks needs optimization. Depending on how you implemented it, it is very possible that the game is copied from RAM to RAM before being loaded, that is why you don't see the 100-200% speedup in loading.

On linux this test would be much better.

You should find some tests that are disk bounded rather than CPU bounded. Like restart of windows from ramdisk.

 

[Dkcode]

Nice write up!

Thanks for putting the time and effort in.

 

[XLNC]

You should find some tests that are disk bounded rather than CPU bounded. Like restart of windows from ramdisk.

I did test Linux virtual machines from the RAMdisk, and it was quite a bit faster. Though, the boot times were not very consistent between retries so I stuck to gaming only benchmarks.

As for copying from RAM to RAM, that's basically the best I can hope for with my limited understanding on the subject. I'm starting to wonder though if the symbolic link is some kind of bottleneck, since it's redirected from the SSD. I'll try installing the game directly to the RAMdisk to see if it makes a difference. If you have any info on RAMdisk optimizations, I'd love to read up to learn more.

 

[Nothinman]

As for copying from RAM to RAM, that's basically the best I can hope for with my limited understanding on the subject.

Yes, unless Windows has a way for the RAM disk driver to tell the kernel "this is memory, use XIP and use it directly", and the driver actually uses it, then it will be paging everything from the disk into main memory before use.

I doubt the symlink has any appreciable affect on things since it just adds a single lookup which would be cached after the first time.

If you have any info on RAMdisk optimizations, I'd love to read up to learn more.

Since Windows and that RAM disk driver are both closed source I doubt there's much you can do.