RocketRaid 3120 -- with onboard IOP using 2xRaptorX

[keithlm]

SOME (Personal) HISTORY AND DEFINITION OF HARDWARE RAID

A long time ago on an old dual CPU (pentium III 1Ghz) I played with two 30Gb drives using the onboard RAID-0. It was fun and you could really see a difference. At the time I thought I was using "hardware" raid. I had no idea what the difference was. (Most Windows users never do.) However, being an avid FreeBSD and sometimes LINUX user I quickly discovered that even though you set it up in BIOS it was not "really" a hardware raid system. It is "software" raid in that you system's CPU must do all the work via drivers. A true hardware raid is always on an external card and has an onboard CPU called an IOP. (NOTE: Some really expensive server motherboards DO have a "true" raid system built into the motherboard. But I have never seen a user/enthusiast motherboard with a "real" raid system.) The reason you see the difference with FreeBSD and LINUX is that the developers do not want to support software RAID. They have an aversion to it. (I think they do support the popular ones nowadays; they didn't back in 2001.)

Anyway fast forward. At one point in time I bought 4x34Gb raptors and used them on a REVO64 PCI card in RAID-0. VERY nice. But after awhile 1 drive died and then I was at RAID-3 with the 3 remaining drives. Nice to have redundancy; but it is not needed on my gaming machine.

When I upgraded to Vista and a 8800GTS, I went ahead and bought a RaptorX (150Gb). It was almost as fast as 2x34Gb of the older raptors in a RAID-0 (or 3x in RAID-3) In addition the Netcell company went out of business... so there would never be a REVO64 on a PCIe card. So I stopped using RAID-0 for a year.

Recently I decided to play with RAID-0 again. So I bought another raptorX. The 2x150Gb Raptors bottlenecked BADLY on the REVO64 card. In addition that card doesn't play well with NCQ. Actually for some unknown reason the two 150Gb raptors ended up being slower than the 2x34Gb "old" raptors. I didn't bother investigating since it was bottlenecked about 108Mb/s even though PCI should be able to do 133Mb/s (or 150Mb/s??). I ripped that card out and went with the onboard NVidia software RAID (i.e. also known as "fakeraid").

I got bad "breathing patterns during benchmarking due to the CPU and waiting.

So I searched and decided on a Highpoint 3120 "True" hardware raid card with an onboard IOP and 2xSata connections. These are 1/3 of the price of a 4xSata card from Areca. This card also supports port multipliers if I ever decide to use an external cage of 5xSATA in the future. The card also knows about NCQ and other advanced features.

I hooked up the 2x150Gb Raptors and tested it out.

Here are some benchmarks I created. I thought I'd present them here for anybody that is interested in this card OR anybody interested in Hardware VERSUS Software RAID-0 (i.e. fakeraid) Please note this is under Vista32 so the CPU utilization numbers are very questionable: (64%? ick. There is another thread about that subject.)

HDTach both with 64K Stripes: Notice the obvious "breathing" pattern due to using the system's CPU instead of an onboard IOP... it's VERY obvious.

HD Tune: (64k Stripes) and the "breathing" gets really heavy here....

and a few using ATTO. I forgot to save the NVidia RAID for this test, but I did both a 64K and 16K stripe on the RocketRaid 3120; I see many people telling others to use 16k stripes. On a true hardware card... that was actually slower:

HOW does it work in "subjective" terms? It's great. There is a noticeable improvement over using the NVIDIA software RAID-0. (I knew there would be. After you use a "true" hardware raid system... it is hard to go back to software raid... because you "know" better.)

I didn't do a software raid-0 test that I want to try out just for giggles: Run a 2x copies of Prime95. One for each of my cores. Then run the disk drive benchmark... see what speed you get. I may actually rip out the card and <sigh> reload again and try that after I do that with the hardware card. (Unless someone with a NVidia software RAID-0 wants to run it so I don't have to...)

But anyway... mostly... the RocketRaid 3120 does exactly what it is supposed to do: be INVISIBLE.

(Also works in FreeBSD and LINUX but I've not had time yet to load either.)

Would I rather have SAS or a new FLASH drive? YOU BET I WOULD. How about 2xflash drives RAID-0 on this card... er... it would probably bottleneck the 250Mb PCIe x1 bus... darn it. I'd be in the same boat that made me buy this card... and I'd have to get the Areca 4 port card and pay a lot of money.

 

[nukemaster]

just gonna toss an intel bench on, Seems to not play as much as NV. Was the cpu usage that high or was the computer doing something else?

For the old drives its not too bad
Dont ask why the burst is so high...does that with write back cache on...

Once again, ignore the burst since it must be buffering in ram or something...

 

[keithlm]

Was the cpu usage that high or was the computer doing something else?

It was running nothing other than the benchmark.

NOTE: It is running Vista. That could explain it. I ran these benchmarks twice for each one... and the results were the same both times.

I may throw a couple of older 2x34Gb raptors into the machine and benchmark on the old REVO64 card and also on the Nvidia RAID-0 while running 2 copies of Prime95 like I mentioned... just for giggles. I want to know what happens to NVidia when the CPU cores are pegged at 100% versus a true hardware card. (I'd use the old REVO64 card because I don't want to reload the entire system... yet again. I did that like 6 times last week.)

 

[nukemaster]

ahhh good idea...loading 2 x f@h and testing again

Folding 2 x SMP

 

[keithlm]

I also tried the "pegging the cpu" benchmark. (Only using Prime95 to peg the CPU.)

It didn't really change the results very much when you compare 100% CPU pegged usage and benchmark ONLY running.

As you can see HDTune:

and also here HDTach: (Why is it at 0% CPU? The CPU is pegged.)

SO THEN: EPIPHANY
I decided... okay no big CPU change caused by starving the RAID of CPU. What about if I run a VIRUS SCANNER on the SAME DRIVE while we benchmark? I used CLAMWIN and started it on a LARGE file. The 700meg Vista64 patch file was available. I ran the same tests.

HDTune showing ClamWIN running and then without:

HDTach showing ClamWIN running and then without:

WOW... I was impressed. It lost a little but not a bunch. Lost a little transfer speed and a bit of latency.

I decided to run the same two tests on my WD2500 "backup" drive that is a single non-RAID disk.

HDTach with ClamWin running and then without:

WOW... look at the yellow dots on the ClamWin run... the latency is ALL OVER the place. UGLY!

HDTune with Clamwin running and then without:

WOW... What sad results on a single SATA2 disk! Transfer is VERY bad and latency went way up. (Although I might re-run the ClamWin on the WD2500 run again... I don't know why it shows about 60Mb/s at the start and then loses it... anomoly or not?)

I'm now even more impressed with my 2xRaptorX in RAID-0. (And/or the Highpoint 3120 card.)

BAZOOKA: Just for FUN I ran ClamWin and then ALSO ran 2 copies of Prime95 to peg the CPU. Just to make it even more interesting while those ran I ran the NVidia graphic demo "Cascades" which uses all the GPU power. Then I ran the HDTune.

The EVERYTHING AND THE KITCHEN sink run:

BTW: Here is a partial screen shot showing everything running. (I have dual monitors. The NVidia demo was fullscreen on the main monitor. Yes I realize I forgot to have ClamWin do only the BIG file... but this run wasn't compared to anything else so I'm not going to re-run it.)

CONCLUSION:

My test might not have shown a definitive difference between hardware and software RAID-0 with the 2xRaptorX as I had originally set out to do. Pegging the CPU at 100% didn't really affect the speed of the RAID system on either the hardware or software RAID. (Basing my software RAID opinion on the post by nukemaster. I'm going to assume his software raid results are similar to what I would get with NVidia. Thanks nuke!)

However running the Virus scanner (ClamWin) and then running the test revealed that the RAID-0 runs RINGS around the single disk drive... not just in speed which is expected... but in the fact that the RAID is not brought to it's knees by the Virus scanner. You can see a drop in performance... but it's very acceptable. The single drive goes to 4.5Mb/s average transfer. Ugly.

So the jury is out on "Software" versus "Hardware" when it comes to benchmarks. Personally I can't live without a dedicated hardware RAID card. I know others don't agree.

BUT I BELIEVE that I "accidently" showed a definitive benchmark that reveals WHY a RAID-0 system actually IS better than a single drive. (Although I can't easily run the RaptorX as a single drive at the moment so that I could run the tests on a single RaptorX.) Anyone with a single RaptorX 150Gb want to run the benchmarks just to show us how a single raptor performs in this situation? (I.e., will it do dramatically better than the single WD2500?)

 

[nukemaster]

Wow. you really wanted to beat the system up did you.

With the virus scanner going on the test drive the head(s) have to move ALLLLLOT(think of jumping from scanning to reading how fast you are at a certain point every few MS) more causing a massive drop in performance. Now do that with a SSD and that bottleneck would be gone.

If you want me to run any tests just ask, but old drives are not gonna ever match those raptors....

I did try a hardware vs software benchmark once with a Silicon Image Steelvine hardware raid 0 / 1 raid chip(built into the P5W DH board) and the results ended up being almost the same. But in some more real world benchmarks(reading and writing actual files, i forget what bench it was, but i will look) the Intel was ahead. But that Silicon Image one had no buffer or anything...

 

[keithlm]

Wow. you really wanted to beat the system up did you.

....Now do that with a SSD and that bottleneck would be gone....

Drool... yes... SSD.

But that'll be in a year or two. (or three or four. Unless you're rich.) Besides they're 32Gb now? Too small. But it would handle just the OS nicely...

And yes... I wanted to beat things up on this machine. I dont' like how people benchmark things in "perfect" situations and then accept the results without questioning them. Hey... there is a reason we have MULTI-CORE systems now... let's beat them up...) I may be wrong in some cases on what I believe the outcome will be... but if you don't run the test... you'll never know the outcome.

I call it "Extreme Benchmarking". I plan on doing some more ridiculous things when I upgrade my CPU/MB/RAM in a month or so. If a benchmark runs on 1 cpu... why can't I run 3 copies of Prime95 on the other cores if it is a QUAD core? (Or 2 if the benchmark will use 2...) How much will having the Prime95 running affect the benchmark? The answer SHOULD be "not much" just as pegging the CPU at 100% on this test didn't really affect the drive benchmarks. But I've not DONE the test so I won't believe it until I see it. If the Prime95 applications eat up all the cache while they run.. then the benchmark will be affected.

 

[nukemaster]

if you get the latest prime95 one copy will run as many threads as cores....that way you are always at 100%

 

[Scaj]

Hmm, never considered the option of "Hardware" Raid when onboard "Software" Raid is more convenient, less expensive and still effective enough for improving HD performance. Incidenty I use WD1500AHFD's in all my systems and some benchmarks are as follows w/ HD Tune v2.55:

Single Raptor Transfer Rates
Minimum - 2.7 MB/s
Maximum - 83.4 MB/s
Average - 55.9 MB/s
Access Time - 9ms
Burst Rate - 107.7 MB/s
CPU Usage - 7.8%

"Software" Two Raptors Raid 0
Minimum - 77.2 MB/s
Maximum 139.4 MB/s
Average 118.5 MB/s
Access Time - 8ms
Burst Rate - 104.9 MB/s
CPU Usage - 3.1%

 

[jkflipflop98]

I guess people just inherently turn their noses up at some things. Weird.

 

[keithlm]

Hmm, never considered the option of "Hardware" Raid when onboard "Software" Raid is more convenient, less expensive and still effective enough for improving HD performance. Incidenty I use WD1500AHFD's in all my systems and some benchmarks are as follows w/ HD Tune v2.55:

Single Raptor Transfer Rates
Minimum - 2.7 MB/s
Maximum - 83.4 MB/s
Average - 55.9 MB/s
Access Time - 9ms
Burst Rate - 107.7 MB/s
CPU Usage - 7.8%

"Software" Two Raptors Raid 0
Minimum - 77.2 MB/s
Maximum 139.4 MB/s
Average 118.5 MB/s
Access Time - 8ms
Burst Rate - 104.9 MB/s
CPU Usage - 3.1%

Those figures are with the Virus Scanner running? The single raptor seems to work fairly well... it only slows down to about the same speed as a single WD2500 drive. The software raid doesn't slow down too badly either.

What are your figures for those drives when you don't have the Virus scanner running? I'd guess that your normal average speeds would be about 75Mb/s for the single raptor and 140Mb/s for the raid-0.

 

[cpt1nsano]

Sorry, but where are the Raid 5 tests? That's where a hardware only card shines. (and needed)

 

[cjl]

Agreed - for RAID 0, the Intel (and I would assume Nvidia) works fine, but RAID 5 or 6 really requires a card for best performance.