Help settle a dispute....

[Nedkt]

Ok, three questions from another forum:

1. For a RAID 0 Raptor array, at what point in the array size does extra drives start to impact random access times in a significant fashion? One person claims 2 drives is the sweet spot, one claims 4 drives is the sweet spot.

2. For the same RAID 0 array, is a high quality dedicated PCI-e (or PCI-X) controller with 128 meg onboard RAM going to give a "snappier" single-user experience or will using the latest Intel Southbridge and plugging direct to the mobo? The claim is that the PCI-e/X card will add so much latency that single-user use will actually take a hit compared to running the array on the southbridge, except in cases of dealing with very large files on a regular basis.

3. Does the HD cache have any impact on RAID performance? The claim is that once arrayed, the cache ceases to be used on the HDs.

 

[libelle]

dont consider me, in any way, fluent in the latest tech info, i'll just speak in terms of experence. back in 01/02, THG had a RAID review, when all these mobo's were coming out with on-board raid.
let me tell you, its the best thing for "snappy" speed because HDD's are the most lagging part of a system

i have a clear memory on the speed reviews and a second RAID 0 drive provided around %40 more HDDs access speed while a third was around %15 on top of that and like %10 for the fourth.
but remember, the more drives you add, the greatter the chances of the RAID crashing when a HDD dies.

in '00, i had taken a Promise Ultra 66 IDE expander card and modified it into a RAID card. even then, with the PCI bottleneck, it was still faster than a single drive. but when i went to this on-board RAID controller... i could tell a signifigant difference.

the last i know the least about, whether a disk head's throughput is incresed/decreased comparatively to the disk controller's data handling.
imho, a hdd cache is just a buffer to keep the data at a more steady flow, as to ease the controllers workload... whatever that may be on a wait for data receive.

 

[TabrisDarkPeace]

Ok, three questions from another forum:

1. For a RAID 0 Raptor array, at what point in the array size does extra drives start to impact random access times in a significant fashion? One person claims 2 drives is the sweet spot, one claims 4 drives is the sweet spot.

2. For the same RAID 0 array, is a high quality dedicated PCI-e (or PCI-X) controller with 128 meg onboard RAM going to give a "snappier" single-user experience or will using the latest Intel Southbridge and plugging direct to the mobo? The claim is that the PCI-e/X card will add so much latency that single-user use will actually take a hit compared to running the array on the southbridge, except in cases of dealing with very large files on a regular basis.

3. Does the HD cache have any impact on RAID performance? The claim is that once arrayed, the cache ceases to be used on the HDs.

1 - There is no sweet spot for RAID-0, all the read ahead, and prefetching tha the OS Disk Cache does will mean HDD1 will seek, then read, while doing this HDD2 should have already seeked, and is reading, HDD3... and so on. As there is no XOR going on (this is reads anyway) the sweet spot is whatever floods the bus the HDDs are on.

Thus 8 - 12 HDDs would be the 'sweet' spot for cheap RAID-0, with a Areca controller maybe even more.

However if you mean sweet as in price/performance, then 5-6 HDDs would be the figure you're after.

2 - A PCIe or PCI-X RAID controller card with 128 MB of cache is going to give a snapper feel than the Intel SouthBridge implemented RAID. Using a Dell 2850, 1950 or similar like a workstation proves this point, but it depends on workload, etc (Most people would notice some improvement, but if both systems have 4 GB of RAM, then the delta would be reduced).

3 - When in RAID-0 the HDD cache is still used on every drive, for both instructions (eg: seek this, do that) and data storage buffer (eg: writes queue up, NCQ takes effect, then writes are performed in the best possible way)

However with a 128 MB controller, and 4 x 16 MB cache HDDs the caches DO NOT ADD UP IN AN EXCLUSIVE CACHE WAY, each cache would be INCLUSIVE (much like Intels inclusive L1/L2 cache, vs AMDs exclusive L1/L2 cache).

Tip: When it comes to RAID suggestions, most people are full of ****.

If you want to calculate optimal stripe sizes, let me know the drives, the number of drives, etc and just fire me a PM. (Private Msg).

8) - TabrisarkPeace

Oh, and to add some credit to my post:

(You'll need to 1:1 zoom that image using MSIE v7 or Firefox)

 

[libelle]

obviously so, the more drives you add make up a faster system, up to what the controller's throughput cap is. i said the second drive is the "sweet spot" because it produces the greatest single gains over the more drives.

agreed on everything else though
how about explain my partition problems, Tabris

 

[TabrisDarkPeace]

What "partition problems" ?

I'd recommend the first partition be 10 to 40% the total capacity of the array, and that it hold your primary OS and pagefile.

Unless you want to run multiple OS's, each needing different page/swap space (eg: Linux uses a partition as swap, but multiple Linux OS's can share the same swap partition - - Unless you're using Xen, but I'd need to check on that).

The first 33% of a HDD is the fastest DTR and seek time.
The next 33% is OK
The last 33% is really slow, half as fast as you'd expect.

This applies to arrays of drives aswell, the first 33% of the array will massively outperform the last 33%.

So I'd arrange your partitions and pagefiles / swap-space / heaviest DTR + seeking I/O dependant files based on that fact. (eg: MP3's use like 16 KB/sec to stream, so they'd suite the last 3% of the drive or array fine).

You're smart, I am sure you already have something similar in mind.

I'll be back in 20 hours or so.