RAID 0 - Stripe Size For Best All Around Performance?

[-JJ-]

Hi there,

I am about to create a RAID 0 array with 2 (two) WD6400AAKS hdd's on ASUS P5K-E WIFI/AP (ICHR9) motherboard..

The plan is to create 1st slice with 300gb and second 2nd slice with all remaining space.

Most of the data are small files with size ~5mb-6mb.

What stripe size should I chose in order to get the best all around performance? 64kb ? I am particularly interested in fastest possible boot times. The operating system shall be WinXP Pro.

Is it right that the smaller the stripe size the better the boot performance?

[Polizei]

I've tried 2 slice sizes with my RAID0 boot array. 64k had the best performance, considering writes and reads. When I switched to 32k after a format, my read speed went up a bit, maybe 5%, but my writes took a huge hit, much slower.

I don't think boot volume's file size average 5-6MB... Windoze has much smaller file sizes. I say go for 64k.

[zanzabar]

it depends on what ur doing and what drives/controller u have, i have a stripe size of 128 right now and it seams better than the 64 for mine

but small files will be better with 64 most likly

[One_Hertz]

32 or 64 for general use. 32 if you don't do too much writing, which is what I use.

[-JJ-]

it depends on what ur doing and what drives/controller u have

I wont have no seperate raid controllers. As I already said I will be having RAID 0 on ASUS P5K-E WIFI/AP motherboard which has ICHR9 chipset

Will it make a difference, if I decide to create a RAID 0 without having no slices, just one large RAID 0 consisting of two WD6400AAKS hdd's, instead of having two slices (300gb/900gb).

Well, sure that hdtune and hdtach will give much better results on the smaller slice (lower accestimes and higher read/write mb/s) if I make one smaller slice and one larger, however will it (slicing or no slicing) affect boot times / app load times differ ?

If "slicing" is essential, than what should the maximum amount for the smaller slice?

Would it be fine, if make it like this:

Slice No.1 - 300gb,
Slice No.2 - 900gb,

?

[Polizei]

I can't imagine you installing 300GB worth of programs. 300GB and 900GB will work just fine.

[Serra]

I'd go 64k - it's a universal default for a reason. The rule of thumb is pretty much unless you have a specific reason to change it, keep the default.

[-JJ-]

I can't imagine you installing 300GB worth of programs. 300GB and 900GB will work just fine.

I agree, however the thing that cracks me up is the “cooperation” between slices in Intel Matrix Raid. For instance, copying a certain file from Slice to Slice takes much more longer time than it does in a non raid mode from one HDD to other HDD. You know what I am sayin’? Even copying from “slice” to USB flash drive is faster.

I remember first having two Caviar RE2 500gb in Intel Matrix Raid with slice no.1 – 100gb, and slice no. 2 – 900gb. Well If I had to move/copy files from no.1 to no. 2 or other way around it would take me forever, and that is the reason why I am thinking of having RAID 0 without no slices.

Data loss is not an issue. I have had raid 0 for years and never had any issues of data loss. Moreover I backup all of my important stuff.

[doakh]

Can you use different stripe sizes in raid 0. For example a 32 or 64 stripe size for your operating system and programs and another stripe size for your data. I will write larger video files so I am thinking a 128 (or larger) stripe size for these. So two partitions.

[lefy]

Can you use different stripe sizes in raid 0.... So two partitions.

sounds like you're talking about matrix raid with two volumes? if so, then yes, each RAID volume can have a different stripe size.

[IanB]

I agree, however the thing that cracks me up is the “cooperation” between slices in Intel Matrix Raid. For instance, copying a certain file from Slice to Slice takes much more longer time than it does in a non raid mode from one HDD to other HDD. You know what I am sayin’? Even copying from “slice” to USB flash drive is faster.

I remember first having two Caviar RE2 500gb in Intel Matrix Raid with slice no.1 – 100gb, and slice no. 2 – 900gb. Well If I had to move/copy files from no.1 to no. 2 or other way around it would take me forever

Well, DUH! Disk read from one "end" of the disk, disk write to other "end", followed by read of next chunk of data, then write, then read, then write... Of course it's slow, the disk heads are seeking from one end to the other continuously, and that's the slowest part of the process, tens of milliseconds between every seek.

Writing from one disk to another avoids the continual reseeking because then the reads and writes are contiguous. That's not a "bug" or deficiency in Matrix RAID, it's just unavoidable physics if you are copying files between different parts of the SAME disk(s).

[virtualrain]

I'd go 64k - it's a universal default for a reason. The rule of thumb is pretty much unless you have a specific reason to change it, keep the default.

The number of drives must play a factor.

What about 64KB divided by the drives in the array... 32KB for dual drive arrays, 16KB for four drive arrays, etc.?

[-JJ-]

Well, DUH! Disk read from one "end" of the disk, disk write to other "end", followed by read of next chunk of data, then write, then read, then write... Of course it's slow, the disk heads are seeking from one end to the other continuously, and that's the slowest part of the process, tens of milliseconds between every seek.

Writing from one disk to another avoids the continual reseeking because then the reads and writes are contiguous. That's not a "bug" or deficiency in Matrix RAID, it's just unavoidable physics if you are copying files between different parts of the SAME disk(s).

i got it .. now it seems so obvious

[AlphaHeX]

The number of drives must play a factor.

What about 64KB divided by the drives in the array... 32KB for dual drive arrays, 16KB for four drive arrays, etc.?

Yes, that is true but I think that 64 is for 2 drives, 32 for 4, etc.

[-JJ-]

just for info I will post the end results of my WD6400AAKS x 2 Intel Matrix Raid setup:

So here are two WD6400AAKS on intel matrix raid. One RAID 0 slice is 200GB (for OS) and the other one with all remaining GB's is 1TB RAID 0 slice. Benchmarks are done with HDTach 3.0.4.0 - Long bench 32mb. Both "slices" have stripe size of 64kb.

Slice No. 1 - 200GB


Slice No. 1 - 1TB

[swiftex]

Yes, that is true but I think that 64 is for 2 drives, 32 for 4, etc.

me - 5 drive array 64k all the way - use kiss principle - i find if purposes different then a separate array set-up specifically for that purpose

mind you i do not have anything near the experience some guys here have in this department

[ikjadoon]

Here is an example of how block sizes affects a WD6400AAKS's read speed, albeit a single one:



~Ibrahim~

[Serra]

The number of drives must play a factor.

What about 64KB divided by the drives in the array... 32KB for dual drive arrays, 16KB for four drive arrays, etc.?

I probably never would have noticed this if there hadn't been a thread resurrection... but yeah, I would say that offhand (with no results to look at) that would sound extremely reasonable to me.

The only thing I"d change though is that since RAID-0 inherently requires 2 disks and since 64k is the default which every manufacturer has independently determined to be the best *overall* size, it would have to go like: 64KB for dual drives, 32k for quad, 16k for octo... again, all *on average*

The actual trick is to determine what works best for the size of files you're reading/writing... which pretty much means running real-world bench tests on activities you perform (benchmarks are only going to show for the benchmark). Unfortunately, I'm also willing to bet that for most applications this difference would be so small it would be nearly impossible to accurately stopwatch it so it's a bit of a tossup either way.

[stevecs]

Generally yes, if you take your average (80% rule) request size and then you set your stripe stripe size so that stripe size * (stripe_width - parity drives) <= request size.

Now this is _NOT_ linear, this scales better when you have a higher degree of parallelism of your requests. This is common here when you run streaming tests say (your queue length is close to 1 generally). This is not what you see in heavy use environments which is one reason why you have to really watch your workloads (request size, queue depth, request type (random,sequential), et al).

Generally if you take the time to map out all the combinations of stripe size/request size the difference is minimal. It's more of a function of queue depth or request size for your performance.

I've attached an example here for a raid-6 array (24 drives, all combinations of stripe size and request sizes from 64KiB to 16GiB). This was a fast test and the only one I had handy on the USB key here as I had this argument recently at work.

[-JJ-]

more results.. this time with smaller first slice.. same 64kb stripe size..

so 2 (two) WD6400AAKS in Intel Matrix Raid, with two slices (100gb & 1,05tb).


Slice No.1 --> 100gb

Slice No.2 --> 1,05tb