RAID 0 not scaling well with HDD's?

[netboi]

I am testing a SAS3082E-R (LSI1068e) HBA setup with RAID 0. with four seagate 250gb hard drives. Using HDtune, the throughput measured is only 180-200 Mbps. Without RAID 0, single 250gb throughput is 125Mbps.

Using two samsung 320gb, using HDtune throughput is only 150Mbps. Is this typical or is something wrong with my testing? Will using 2 SSD allow me to reach around 400Mbps using RAID 0?

link for the card: http://www.scsi4me.com/ibm-serveraid-br10i-8-port-pci-express-sas-raid-controller.html

My setup:

AMD X2 235e 2.7ghz
MSI K9A2 platinum
3gb DDR2

 

[engiNURD]

is write-back cache enabled?

SSD's scale extremely well in RAID0. As for the 400MB/s, depends on the SSDs used, but most likely yes. HDDs are known to scale poorly in RAID0 (compared to SSDs).

btw, PM a mod to move this to the storage subforum.

 

[Dangman]

btw, PM a mod to move this to the storage subforum.

Moved

 

[netboi]

is write-back cache enabled?

SSD's scale extremely well in RAID0. As for the 400MB/s, depends on the SSDs used, but most likely yes. HDDs are known to scale poorly in RAID0.

btw, PM a mod to move this to the storage subforum.

write-back caching cannot be enabled for some reason. Error message says device doesn't support changing write-back settings. Im using server 2008 R2. Guess time for SSD? expensive

 

[Red Falcon]

Probably because you are using SATA drives and not SAS drives is the reason the writeback cache can't be enabled.

Also, HDDs scale well with good returns with up to 4 HDDs in RAID0, however, anything more barely gives any increase in speed.

Using my Sil 3114 PCI controller, along with 4 old Samsung 80GB drives in Software RAID0, they get 158MB/s read speeds and 100MB/s+ write speeds. Not bad for drives that are 8 years old. But with a single drive, I get around 85MB/s read speeds and 70MB/s write speeds. If you multiply each of those by 4, it should be much greater than what I am getting in RAID0.

Just remember, anything beyond 4 HDDs will only net about 5% more performance per disk aggregated.

 

[odditory]

Also, HDDs scale well with good returns with up to 4 HDDs in RAID0, however, anything more barely gives any increase in speed.

Using my Sil 3114 PCI controller, along with 4 old Samsung 80GB drives in Software RAID0, they get 158MB/s read speeds and 100MB/s+ write speeds. Not bad for drives that are 8 years old. But with a single drive, I get around 85MB/s read speeds and 70MB/s write speeds. If you multiply each of those by 4, it should be much greater than what I am getting in RAID0.

Just remember, anything beyond 4 HDDs will only net about 5% more performance per disk aggregated.

Beyond 4 HDD's not scaling simply isn't true, it all depends on the controller. With controllers like those based on the SAS2008 and SAS2108, together with a SAS2 expander you'll still see near linear scaling up to about 16 or 17 spinning disks in software RAID0, then more gains up to about 18 or 19 drives and it tails off drastically after that. With the next generation of SAS2 controllers, and subsequent generations of harddisk controllers the bar will go even higher. Same holds true for fast SSD's but the curve looks a bit different because they're more efficient with less latency per I/O.

 

[Red Falcon]

Beyond 4 HDD's not scaling simply isn't true, it all depends on the controller. With controllers like those based on the SAS2008 and SAS2108, together with a SAS2 expander you'll still see near linear scaling up to about 16 or 17 spinning disks in software RAID0, then more gains up to about 18 or 19 drives and it tails off drastically after that. With the next generation of SAS2 controllers, and subsequent generations of harddisk controllers the bar will go even higher. Same holds true for fast SSD's but the curve looks a bit different because they're more efficient with less latency per I/O.

Can you show me some graphs or charts proving this?

I've used a lot of hardware in my time and with RAID0, anything over 4 drives doesn't scale well at all with minimal returns per HDD added.

I'm aware that SSDs scale differently due to their different nature and higher I/Os with lower latency.

I really need to see some proof though that it is the controller that is the limiting factor. I'm not saying you're wrong, I just need some proof.


Also, when you mean software RAID and using a controller, I think you mean FlexRAID as software RAID is not reliant on the controller as much as the basic throughput of the controller, ports, bus, and CPU speed.

 

[sub.mesa]

I see no benchmarks i can comment on.

Try benchmarking with CrystalDiskMark or AS SSD and post the screenshot here. HDTune may yield incorrect results when used on RAID arrays, so only use HDTune for HDDs not for RAIDs. It's a raw I/O benchmark that does not simulate real access. CDM and AS SSD work on the filesystem so can provide meaningful scores.

@Red Falcon: how would you get 150MB/s+ performance when PCI is capable of 133MB/s bandwidth (not throughput!) max? RAID and PCI are eachothers enemies; PCI is not designed for parallel access.

 

[Red Falcon]

@Red Falcon: how would you get 150MB/s+ performance when PCI is capable of 133MB/s bandwidth (not throughput!) max? RAID and PCI are eachothers enemies; PCI is not designed for parallel access.

I have it in a 32-bit 66MHz slot, effectively doubling it's bandwidth to 266MB/s.

When it was in a normal 32-bit 33MHz slot, it was incredibly limited to the bus speed of 133MB/s and it only got around 124MB/s read and 80MB/s write speed.

Now it gets 158MB/s read speed and around 103MB/s write speed which is more appropriate for drives of this size and age (80GB and about 8 years old).

EDIT: My controller: http://www.siliconimage.com/products/product.aspx?pid=28

It's backwards compatible with either 33 or 66MHz, not all controller cards can do that as most are limited to 33MHz.

 

[sub.mesa]

I see. But RAID or any other parallel access would scale much better on dedicated point-to-point interfaces instead, like PCI-express. This should mean that doing I/O on disk A should not affect any speed/latency on the I/O done on disk B, for PCI this is the case however and may be why it didn't scale properly for you.

RAID0s above 4-disks are quite rare, though if you look at SSDs you can see the 8-disk or 10-disk internal RAID0 scales well in benchmarks. You can see my own RAID0 benchmarks (2 or 4 disks) over here:
http://submesa.com/data/raid/geom_stripe

RAID0 can scale perfectly, but that means killing all potential bottlenecks, and the interface is one of them!

 

[netboi]

I see no benchmarks i can comment on.

Try benchmarking with CrystalDiskMark or AS SSD and post the screenshot here. HDTune may yield incorrect results when used on RAID arrays, so only use HDTune for HDDs not for RAIDs. It's a raw I/O benchmark that does not simulate real access. CDM and AS SSD work on the filesystem so can provide meaningful scores.

@Red Falcon: how would you get 150MB/s+ performance when PCI is capable of 133MB/s bandwidth (not throughput!) max? RAID and PCI are eachothers enemies; PCI is not designed for parallel access.

thanks! turns out hdtune wasn't the tool for the job. I tried CDM and the scores are somewhat believable. Here's for the 2 wd 320gb:

btw, is it true that throughput for this card (SAS3082E-R, LSI1068E based) can provide only 150-175MB/s/port of concurrent I/O, which is good enough for HDDs (but not SSDs)?

link: http://blog.zorinaq.com/?e=10

 

[skiddy]

Will using 2 SSD allow me to reach around 400Mbps using RAID 0?

Here's my crystaldisk benchmark for my 2 Intel 80gb G2's in RAID0 to give you an idea. Using the crappy onboard Intel controller.

 

[notarat]

thanks! turns out hdtune wasn't the tool for the job. I tried CDM and the scores are somewhat believable. Here's for the 2 wd 320gb:

btw, is it true that throughput for this card (SAS3082E-R, LSI1068E based) can provide only 150-175MB/s/port of concurrent I/O, which is good enough for HDDs (but not SSDs)?

link: http://blog.zorinaq.com/?e=10

Can you try picking up some 1TB drives? they're cheap and the increased platter density makes for some solid improvements in throughput.

(just posting 1TB scores for comparitive purposes...not making a "mine is better" post)