RAID Benchmark Results

I wouldn't worry too much about RAID-6, the main reason it's there is to address the possibility of a second drive failure during a rebuild when using very large capacity drives. You would be unlikely to use it with 15K Cheetahs or SSD drives. What would be interesting to see is the rebuild time for a failed 64 GB SSD in RAID-5 vs. a Cheetah RAID-5 rebuild.

I commented on this earlier today on the RAID card review and it's pretty frustrating. Who is in charge of the graphs? They need to start making them large enough to read again. I mean I *can* read it, but its not as easy as it could be. I'd like to be able to glance at it and see what is going on... not have to whip out a magnifying glass or squint at it to see what the results are.

-mcg

Ok... so at least the zoom feature seems to be working after clicking on the image, getting a slightly larger one, then clicking on it again. Why not just put the larger image in the review in the first place?

-mcg

agreed. The graphs used to be lareger and more visible.

Why would you use a raid card that limits your throughput so badly? Couldn't you find one that supports the theoretical 950MB/s that the SSDs were supposed to show? Instead of showing off some breakneck speeds you leveled off at a measly 560MB/s. Very disappointing. I feel like we learned very little from this article. It was always known SSDs had a lot more IOPS, we also knew about the low access time and about not losing speed when the drive is full. What did you actually show here, besides the fact that you chose the wrong RAID controller?

Some big manufacturer is starting to roll out servers with IOdrives in them (remember those from a milion years ago?), how about a look at that?

come on, the IO might be good , but it doesent matter when the storage space is PATHETIC, no enterprise in the their right mind is gonna upgrade until they actually have enough space for the millions of records regardless of how fast it is. I think after the price is about 1/3rd and the capacity is 256 megs or so then yeah, till then, no way !
or till capacity is about 2 - 250 per gig !

"I/O Performance
These results do not require a lot of comments."

They do if those results cannot even be read.

Since no one has brought it up yet, why not use ST973451SS (Savvio 15k.1, 73GB) or ST9146802SS (Savvio 10k.2, 146GB) rather than ST973402SS (Savvio 10k.2, 73GB)? The 73GB 10k rpm drive is a hard drive to recommend these days with the 146GB version only 10% more cost. This article would be better served with direct price/performance and price/capacity comparisons with more alternative arrays. An IO performance comparison of 1 64GB SSD to 8 146GB 10k.2s would be interesting as they're roughly the same cost, yet the conventional type setup would yield about 18x the capacity and likely at least 6x the throughput. In the end, I'd really be curious to know just how many extreme IO dependent DBs out there would actually need over half a TB space. Of those the percentage that actually care about the $30k/TB to build them would probably be very low.

Aren't the new OCZ and Samsung drives $479 for 128GB ? And they have SATA II 3.0GBps interface ? Why didnt TH test with those ? That's $3,832 for an 8 drive array compared to the $14,400 for the SSD brand used in this review. Besides, TH's own tests showed these drives to be very fast ( SSD power consumption review update )

I look this review and i think:"what type of factory require this products?" . In my opinion for the 80% of factory a SAS array are over dimensioned, a RAID 5 of SATA are the good chose. Bat it's really difficult to found a server with SATA raid 5 hw.

Interesting read.

For those who think the images are too small get Image Zoom for FF3:

odd... link was a no show...

https://addons.mozilla.org/en-US/firefox/addon/139

The article mentions that Tom's chose the Areca card upon request from SSDWorld... I'm guessing they couldn't really refuse since SSDWorld was generous enough to provide the samples. Besides, even if the article proved nothing and provided no useful information, who in their right mind can resist the chance to play with 8 Memoright SSDs raided together?

The article mentions that Tom's chose the Areca card upon request from SSDWorld... I'm guessing they couldn't really refuse since SSDWorld was generous enough to provide the samples. Besides, even if the article proved nothing and provided no useful information, who in their right mind can resist the chance to play with 8 Memoright SSDs raided together?

Bah,I don't really care about this article that much.I'm just sitting there in awe at 8 drives in RAID-0...*mouth-waters*

Dragunover, 8x Drives in RAID-0... yes, very nice... but the included controller puts a bad taste in my watery mouth

They still aren't great for enterprise apps. Cost way too much and don't work well with complex raid setups.

Look perfect for typical desktop, laptop, and even consoles. That is if the price would simply align with the competition.

The tests were missing what I consider the most important part of benchmarks,
which is to include tests for all of the unusually bad performance cases that
we have seen in the past. In particular, we have seen bad performance with
USB flash memory in the past, so this should be checked for.

With this in mind, I couldn't find what to me would be the most interesting information:
1. are we sure that the solid state disks use dynamic wear leveling? (dynamic
as contrasted with none and static wear leveling)
2. What is the performance for copying large trees of small files on one
array? (These tests revealed dramatic performance decreases in the past.)

All of the numbers that were given in the report just confirm what one expect
within a factor of 2 or so.

However, I have had performance problems with USB
flash memory keys where in some cases the performance is down by a factor of
300 from the best case performance, so I am particularly interested in worst case performance.
(Even if one isn't concerned with worst case performance, perhaps using 2, 4,
or 8 times as many Cheetah's at the problem would give better performance than
the solid state disks. A factor of 300 probably can't be overcome just by using more disks

I'd like to see the copy times for trees of 100000 and 1000000 files of
0, 128, and 8192 bytes in each file. The load should be varied from about
1 total to 2 times the number of drives (i.e., 16, perhaps showing for 1, 2, 4, 8,
and 16 copies running on the same array, and perhaps for 1, 2, 4, and 8
copies running between 2 arrays of 4 physical units each.)

The idea of the tests is to see if the performance decreases dramatically when reads
and writes are mixed on the same physical device and in addition logical block numbers
of the read and write operations are spread over the range of possible logical block
numbers. (My tests assumes NTFS file systems; a different set of tests might
be needed to get worst case performance out of other file systems, such as ZFS.)

Just running copy operations is not enough: something has to be done to stop
RAM buffering in the computer or controller card from affecting things and a
test needs to be performed with just one output unit and perhaps more than
1000000 files of 0 length to make sure that no buffer in the devices (Cheetah or
solid state disk is causing an artificially good performance.)

"come on, the IO might be good , but it doesent matter when the storage space is PATHETIC, no enterprise in the their right mind is gonna upgrade until they actually have enough space for the millions of records regardless of how fast it is. I think after the price is about 1/3rd and the capacity is 256 megs or so then yeah, till then, no way !"

Ridiculous. Rarely do we buy storage based on capacity. It's nearly always based on I/O, throwing a number of spindles at something without using the entire drives. This comparison is a bit disingenuous anyway, because the memoright SSD is a consumer product, not the higher performing SLC, ONFI 2.0 SSDs such as what EMC is using.

"Will SSDs Take Over The Enterprise?"

I doubt it, if the Borg can't even do it then what chance do SSDs have?