Multiple RAID Set up on Single MOBO with Raptors and regular SATAs

[G.T]

Processor:
3.20E GHz P4 Prescott

Motherboard:
Asus P4P800-E Deluxe - 1009 BIOS

Memory:
2 Gig Mushkin HP 3200

Video Card(s):
Leadtek A6600GT TDH

Sound Card:
Creative X-FI Fatal1ty FPS running through Creative Inspire 6.1 6700 speakers

Harddisk(s): Total 520GB of Storage = Half a Terabyte;
2 x 74GB Western Digital 10k rpm SATA Raptors Striped in RAID 0
2 x 80GB Western Digital SATA Striped in RAID 0
1 X Maxtor 250GB ATA


The Mobo has Promise RAID controller chip and ICH5R southbridge RAID.

I am making sure that the info I have thus far read about which is faster for what I intend points me to believe that the ICH5R southbridge RAID would be the best place to park 2 shiny new Raptors in RAID 0 for the best speed and performance for the System adn it's general running.

Correct? / discuss please.

As part of the discussion, mainly orientated to the Raptor Owners or aficionados of them, when setting them up in RAID what are the best sizes to choose when given the Format the RAID drives options? I hit the "Desktop overall performance" option but am not 100% sure that is the best choice for myself as I was unsure about the other choices, any help would be greatly appreciated in the questions or ponderings above and any discussion topic related that proves of interest.

 

[Poisonsnak]

I agree with having your array on the southbridge controller instead of the PCI based one. I don't have any benchmarks for hard drives but following is a comparison (by nVidia so maybe take it with a grain of salt) of southbridge based gigabit ethernet and pci-based:

http://www.nvidia.com/object/IO_8451.html

when you say "...what are the best sizes to choose..." are you talking about stripe size? I think I usually pick 16k though I've heard people pick larger (~128k) for drives with fewer large files (something like 128k). Also it's good to match your cluster size when you format it to ntfs to your stripe size.

 

[G.T]

Yes, Strip sizes is the Topic I was trying to broach with so little skill, good catch and thanks. So small is better with respect strip sizings, bugger, heh.

No matter, need to move them around anyhoo, so a reformat and reinstall is no extra work and it's good practice.

 

[Steevo]

I and many other have found that 128K is the best of all worlds. Mebey a file server where files are small you would gain performance by a smaller stripe size.


Try though and see what you experiance. You might get better performance.


Now on to the RAPTOR drives. Again I post the question. WHY? Less density per inch of drive space at higher RPM is still not quite equal to higer density at lower platter speeds.

This is a dead horse that many new jockies get onto unfortunately. The difference between a 300Gb drive at 7.2K and a 74Gb drive at 10K is the 300 can SUSTAIN a higher bandwidth while the 74 can seek faster, technology that is soon becoming negligible due to NCQ active in RAID arrays being the norm. Almost SCSI inteligence in a SATA drive. But not close.

If you really want the best performance bang for the buck buy 4 300Gb Maxtor drives and RAID that at a smaller stripe size to ge the biggest baddest bench you can. Plus have 1.2TB of data storage.

 

[G.T]

Like all new jockeys at a club, it's good to ride as many rides and in as many guises as possible for the personal experience and first hand knowledge, plus how else can you determine performance without seeing it for yourself and the subtle changes tinkering with the settings can bring?

As for larger versions and more storage, all those considerations have been taken in mind, reaching a half terabyte of storage was a nice step in the right direction for now, all in good time.

 

[D_o_S]

I pick 128K stripe.

 

[Alec§taar]

I pick 128K stripe.

Some test data for you guys, regarding "stripe size" on RAID arrays, from Maximum PC magazine issue "Speed Issue" June 2004 & it may surprise you and show you it depends on what you do most in order to optimize this properly:

TEST SETUP:

Stripe sizes ranging from 16k-32k-64k-128k, with:

Small file tests sizes of - 94.5mb total testset size/composed of files averaging 20-35k each

Medium file tests sizes of - 2.27gb total testset size/composed of files averaging 4-6mb each

Largel file tests sizes of - A single 749mb sized file total testset size

* TEST RESULTSET DATA USING HDTACH & RESULTS BY STRIPESIZE (Winning scores are bolded by category):

Avg. Read speed:

16k = 97 <- tie for win
32k = 97 <- tie for win
64k = 95
128k = 95

Small File Write (in seconds):

16k = 32
32k = 30 <- winner
64k = 31
128k = 33

Medium File Write (in seconds):

16k = 71 <- winner
32k = 75
64k = 74
128k = 76

Large File Write (in seconds):

16k = 21 <- tie for win
32k = 23
64k = 21 <- tie for win
128k = 22

** ALL TESTS DONE ON A PROMISE FastTrax TX4000 ATA PCI RAID card...

(You judge based on those numbers, & what type of work it is you do most & with what filesizes types mostly - it matters, like in any optimization, what it is you are out to do and with what kind of data + equipment types)



APK

P.S.=> Those results admittedly surprised me too!

See, I too, would have (and did initially) went with 128k stripes myself as well, but ended up switching to a 32k size (minimum my controller has now on it, no 16k possible).

As mentioned, I did initially using a Promise EX8350 RAID 0 - 6 128mb ECC Ram Caching x4 PCI-e controller w/ onboard Intel Super I/O controller onboard it to offload the main system CPU from managing the RAID arrays here that runs 2 WD "Raptor X" 10k rpm SATA2 150gb disks w/ 16mb of buffer on them each in RAID 0...

However, it appears that 16k-32k seem to win the most tests for typical use-patterns by end-users (this may vary based on filesizes you typically use & work you do though, keep it in mind)... apk

 

[{JNT}Raptor]

16k Is the Overall all around best performing Stripe Size.....It's been tried/tested and proven for some time now.......Matching your OS partitions Cluster size to 16k will give you added performance as well.

Good luck.

 

[Alec§taar]

Agree/Disagree on NCQ/TCQ advantages in SINGLE-USER desktop machines

16k Is the Overall all around best performing Stripe Size.....It's been tried/tested and proven for some time now.......Matching your OS partitions Cluster size to 16k will give you added performance as well.

Good luck.

Good Point that last one... sort of TOUGH to do on a booting partition, @ least if you let the OS install CD/DvD do the formatting of the drive (Fat16/32 not sure, but NTFS doesn't matter - it defaults to 512 byte sized ones @ install from the OS install media afaik/iirc)

The ONLY way around that I could figure is to use tools like "Partition Magic", OR, formatting the disk on another system first, & then letting the OS install media install to it, w/out blowing the partitions already done formatting from another PC!

I go with 4096kb sized ones, to match the memory mgr. page size... simple reasoning enough for that imo!

APK

P.S.=> Interesting point brought up by Steevo above too regarding Command Queuing!

(That is, provided your onboard RAID firmware leverages it (NCQ/TCQ types, & iirc, not many do from say, Silicon Image or others built into today's mobos))

Heck, or your disk I/O controller period on your mobo!

Still, TCQ/NCQ can makes some sense for great transferral rates & even negates (to a degree) the rpm advantage "raptors", albeit (the one point I had a sore spot with in his evaluation above) in multi-user environs ONLY, @ least, imo!

That's where command queuing really helps, for servers, but for single-user disk-use patterns? Not so much... apk