RAID 0 for games - results inside.

[Jimbo Mahoney]

UPDATE - Due to problems with the array and the small performance increase it offers in game loading times, I've gone back to separate disks.


I see this topic get covered a lot with no quantitative results being given. People give responses like 'it feels quicker', which is less than convincing.

I've just received a couple of 150GB Raptors to replace my trusty 36GB Raptor and 200GB standard WD drives. My machine is only used for games (it's an expensive console basically!) and all my important documents / music etc. is on another machine, as well as being backed up.

The choice of how to arrange the two Raptors, for me, was this:

Raptor 1 - Windows in first partition, downloads in second partition.

Raptor 2 - Swapfile in first partition, games and stuff in second partition.

OR

RAID 0 with four partitions.

I decided to test my main use of the system, which is games, comparing the old WD200 drive with the new 150GB Raptor and also against a variety of stripe sizes in RAID 0. The cluster size used was 64k in all tests. I did a quick test with default cluster size (4k?) and it made no difference to benchmarks or game loading.

I used three different games to test their loading times: Doom 3, Left 4 Dead and World of Warcraft.

The system specs are:

E7300 @ 4Ghz (10 x 100)
4GB budget 6400 RAM @ 5-5-5-15, 1:1
DFI P45 DK (ICH10)
Windows Vista 64

To make it a fair test, SuperFetch was disabled and the system rebooted between each run. All disks were defragmented using JKDefragGUI.

First up, a quick HDTach and HDTune of each configuration to be sure everything is as it should be:



Single Raptor:

RAID 0, 128k Stripe:

RAID 0, 32k Stripe:

RAID 0, 8k Stripe:

RAID 0, 4k Stripe:

We can see that RAID 0 does indeed offer double the throughput in benchmarks, as expected. Stripe size does not have a dramatic impact on the results, but 128k seems to be a little slower.

Next up, the games...

 

[Jimbo Mahoney]

To test game loading, the games were started and the system allowed to settle to minimum disk activity. A stopwatch was used to time the period from the moment the load button was pressed to the game appearing.

A Doom 3 quick save was used, WoW was loaded using a character in Ironforge and Left 4 Dead was started using the No Mercy campaign in Single Player mode. Times are given to the nearest second, as this is probably the accuracy of the stopwatch method .

Game loading times were as follows:

These were about the results I was expecting after seeing various reviews, such as:

http://www.anandtech.com/storage/showdoc.aspx?i=2969&p=8

http://www.anandtech.com/storage/showdoc.aspx?i=2101&p=10

I believe they were using 64k Stripes for those tests (possibly a 128k Stripe for one of them).

From these results, we can summarise thus:

Doom 3 load times were reduced by 4 seconds, or ~22%

WoW load times were reduced by 2 seconds, or ~ 20%.

Left 4 Dead load times were reduced by 3 seconds, or ~ 13%.


We can also see that a large stripe size has no effect on load times, even increasing them (although this is probably within the margin of error of the timing method used.

Using the "defrag -a -v" command, the mean file size of these three games was ~ 1MB, however, further analysis using JDiskReport showed that 71% of the files were less than 16kB in size. This explains why a large stripe size has no effect - most files are too small to be split across the array and are therefore only being read from one drive.

I think it's safe to say that RAID 0, as expected, is a little disappointing for game loading time. Only you can decide whether the 10 - 20% decrease in load time is worth it for the increased data risk (and possibly cost, unless you were going to buy two identical drives in the first place).

As for me, I've chosen to use a 4kB Stripe with 64kB Clusters (except on the Windows partition). It's said that a stripe size too small can cause a reduction in performance as the system has to work harder reconstructing the data, but I've seen no evidence of that in my tests, and I'll choose evidence over hearsay anyday...

 

[Jimbo Mahoney]

Finally, for the sake of comparison, here is a Raptor vs standard drive (in this case, a 200GB WD):

There is more evidence of performance increase here than in a single vs. RAID comparison, but there are plenty of Raptor vs. normal drive reviews out there.

 

[Jimbo Mahoney]

Just an update:

Due to numerous errors on the RAID array, meaning I've had to back up and restore my data a few times, I've gone back to separate disks.

I found the array very sensitive to hard resets too.

10 - 20% decrease in game loading times for an increased risk of data loss? Not worth it for me.