RAID 0 - Real-world peformance on the home desktop?

Jimbo Mahoney · 19 Jan 2006 at 14:20

I'm always seeing people wanting RAID 0 arrays, even though they are using a fairly typical home setup of Windows + Games + MP3s.

My personal opinion is that RAID 0 offers little, if any, performance improvement on the typical home desktop system, BUT, I would like to investigate this as fully as possible, partly because I too, ironically, am tempted with the RAID 0 bandwagon...

My own belief, based on my understanding of a RAID 0 setup, is that there comes a point where the latencies of the discs are the limiting factor, something RAID 0 cannot address and in fact, may increase.

So,

Small files = latency is limiting factor

Large, contiguous, files = transfer rate is limiting

My question is, where is the turning point? Is it 500kB, or 1MB, or 100MB?

So far, I've only managed to dig up a paper from 1996:

http://archive.nlm.nih.gov/pubs/bangbuck/bangbuck.php

From this, the fastest performing system (3 disk RAID 0+1) is only 9% faster than 6 simple volumes. Most of their files are <100kB.

Does anyone have any idea as to the file size which means RAID 0 would be faster?

Are there other factors to consider? (Apart from reliability, I'm only interested in this from a performance POV. Everyone should realise the increased risk in having data striped across multiple drives...)

Jimbo Mahoney · 19 Jan 2006 at 14:43

Interestingly, I hadn't realised this before, but any files smaller than your chosen stripe size will only be written to one disk, eliminating ANY performance gain from RAID 0.

Also, a little formula I found:

Optimal Strip size = Average File Size / 2n

where n is the number of drives the RAID controller 'sees'. For RAID 0 = all drives, for RAID 5 = all drives, for RAID 10 = 50% of drives. Always round the final figure down to the nearest stripe size permissible.

EDIT - surely (Average File Size / n) makes more sense?

also

Optimal Cluster Size = Stripe/2

Let me know if these hold water.

Jimbo Mahoney · 20 Jan 2006 at 12:24

/watches tumble weeds roll by....

I see no one makes an informed decision when buying RAID 0 setups then?

/dons flame suit

Or, I'm asking dumb questions...

Jimbo Mahoney · 20 Jan 2006 at 16:21

I just wonder if there's been any benchmarks anywhere which relate to real-world usage with different file sizes, rather than just people flinging HDTach burst speeds around

Of course, having never used RAID 0 myself, I am also open to having my opinion changed.

It is true that with smaller files, you are access speed-limited though, right? Or, does 0% fragmentation mean that they are effectively contiguous and the higher throughput of RAID 0 would help?

Jimbo Mahoney · 20 Jan 2006 at 16:33

noxidjkram@hotm said:
0% fragmentation would mean the files were contiguous *if* the files you were getting were related in a way that the defragmentation process had accounted for, nothing else.

The higher throughput of RAID0 should help on any size, it will just be more noticable on the larger files due to the percentage of total time reason mentioned above.

M

Of course! (to both points).

Ta.

Does anyone know of a program which pulls up statistics of file sizes on a drive?

Obviously, average is easy to do, but can be heavily biased. I found one piece of software which sounded promising, Simpl-File Statistics, but again, it only gives average, min and max and crashed when I gave it a drive to analyse because of the System Volume Information directory. Other statistics would no doubt be useful in helping determine stripe size or whether RAID would be beneficial or not.

Thanks for the input so far. (all)

Jimbo Mahoney · 20 Jan 2006 at 16:36

semi-pro waster said:
If the argument is they are unnecessary for an average home user

I also mean most of us here on OcUK without an OBVIOUS benefit, like video editing. Gamers / overclockers / average home users, but not people who edit 100's of MBs of video, as that is almost a case-closed scenario.

noxidjkram@hotm said:
There are also those of us who must just have the fastest and the best!

I agree, but I'm questioning whether, in many cases, RAID 0 is any faster, except in HDTach. I guess there's the placebo factor to take into account, or the fact it makes you happier to know you have RAID 0 Raptors even though they don't really benefit you in the real world, which is fair enough.

Otherwise, there's me. I have come to realise (after 2 years of RAID0 Raptoring) that those extra seconds do not matter to me for my usage, and that i prefer silence.

You want to get yourself some SilentStar WC enclosures! I have my two drives (one's a raptor) inside a Dual Silentstar and they are great! I used to use SilentDrives (and still do in another rig), but was worried about the Raptor inside one of them with little airflow (I'm WCing the whole rig and there aren't very many fans around).

ROFL

- When you buy the 74GB Raptor RAID bundle, you save 1p

http://www.overclockers.co.uk/acatalog/westerndigital_sata.html

Jimbo Mahoney · 23 Jan 2006 at 11:45

G00SE said:
this thread is most interesting.

i have just gone from raid 0 ide to a single raptor (36gb) and the difference in tangable performance is not noticble!

What was your stripe size and cluster size?

marc2003 said:

Good work fella!

Jimbo Mahoney · 23 Jan 2006 at 11:57

Any further input on the Optimal Strip Size being:

Average File Size / 2n

where n = number of drives over which striping will occur? (e.g. RAID 0 with two drives, n = 2, RAID 0+1 or 1+0 with 4 drives, n = 2).

I'm sure

Average File Size / n

makes more sense?

Anyway, on my work laptop, the entire HDD is one partition, and it breaks down like this (thanks Marc2003!!!):

Distribution of sizes in C:\

Size Interval % of Files

Over 1 GB .......0.0%
256 MB - 1 GB .....0.0%
64 MB - 256 MB ....0.1%
16 MB - 64 MB .....0.3%
4 MB - 16 MB ......0.7%
1 MB - 4 MB .......2.3%
256 KB - 1 MB ......6.7%
64 KB - 256 KB .....13.3%
16 KB - 64 KB ......21.2%
4 KB - 16 KB .......20.5%
1 KB - 4 KB .......17.1%
0 KB - 1 KB .......17.8%

(no doubt the forums will mess up the table format)

You can see that ~ 75% of my files are smaller than 64kB and ~90% are smaller than 256kB.

I'd be inclined to go with a stripe size of 32kB for this system, or even 8kB since some 55% of my files are smaller than 16kB!

Maybe most people are making their stripe size too large when RAID 0'ing at home?

Jimbo Mahoney · 23 Jan 2006 at 17:30

ByteJuggler said:
Surely we should be talking *median* filesize, as opposed to *average* filesize... that's what the size interval tables effectively give you...

Yeh, sorry, that's what I mean. You're right - average is silly.

Jimbo Mahoney · 24 Jan 2006 at 12:37

ByteJuggler said:
Surely we should be talking *median* filesize, as opposed to *average* filesize... that's what the size interval tables effectively give you...

OK, specillay for you:

Stripe Size = Median File Size / n

Where n is the Stripe Width.

Do you want me to incorporate some measure of interquartile range in there too?

Jimbo Mahoney · 25 Jan 2006 at 16:31

Sadly, I think RAID 1 is only as fast as the fastest drive, ie no different from a single disk.