You get what you pay for, and you DO pay!
A minute with my phone's calculator shows that a DRAM array of the same size as the disks costs over $110 million, almost exactly thirty two times as much.
Solid state storage supplier Kaminario has grabbed the SPC-1 storage benchmark with the first million-plus IOPS score, a full 134 per cent faster than previous king-of-the-heap IBM. The Storage Performance Council 1 (SPC-1) benchmark aims to provide comparison performance numbers for a storage subsystem while carrying out …
YOur missing the point Alan. You calculate $/TB, which makes the DRAM stuff monstrously expensive. How many databases need 282TB of spinning disc?
If you find the right problem, e.g. speeding up Oracle, then its a lot cheaper than moving to Exadata.
BTW Kove claim 5.2Miilion IOPs, at 8useconds latency, and it only costs $180K/TB
What about when they make 282TB DRAM-based storage clusters for under $3m?
I think that it is correct to have the differing technologies on the same benchmark, although maybe the benchmark could weight cost per TB more than it does now which could level the playing field somewhat.
I know kaminari (雷) is Japanese for "thunder"(*), but what's with the 'o' at the end? Is it actually deriving from some Spanish word?
* Jim Breen's JDIC also lists kaminarioyaji (雷親父) as an irascible old man (a Victor Meldrew type, no doubt) but that's surely not what the name is hinting at...
If you have a need for that many iops and that little of storage - in many cases you're probably better off buying a couple servers and slapping some Fusion I/O cards in them directly. It is an amazing amount of equipment for only 3.4TB of "storage". I assume if they go beyond 96GB of ram/server performance degrades.
There's no point in the spinning rust array makers to try to compete with this since it's such a niche product.
Kaminario themselves are pushing the result as a measure of $/IOP. Which there is no doubt it achieves, how many will be sold with all DRAM is another matter, I'm sure someone out there needs 1.5TB of mental quick storage. They sell the all DRAM (K2-D) model with up to 12TB, which will certainly hurt the $/IOP. Looking at the latency graph I think that the box will top out at about 1.3M IOPs with a wedge more latency. A 12TB system is, by a very quick back of the hand, going to be about $3/IOP. Which is about half the SVC result ($6.92/IOP), but still a long long way off on the $/GB.
They also sell an all flash (K2-F) and mixed system (K2-H). I think the mixed system is the more likely to be seen. But they have't tested one of these so I have no idea how fast their flash is?
Latency? I don't like looking at the peak values TBH, you don't want to run a box like that, the graph says more; if it's still linear the box has more headroom, if it goes exponential the box is usually at its limit.
The SVC also uses DRAM, there's 192GB (raw) in the SVC and 256GB (also raw) in the V7000 back end. A chunk of this is used as a cache in front of the spinning disks. With about 1/10th the usable DRAM the SVC gets about the same latency at about 1/10th the IOPS. There's a thing.
So... horses for courses anyone?
The problem with the result isn't that DRAM systems shouldn't be allowed to compete with HDD systems, and at $423/GB I wouldn't really say this system was competing with HDD systems, the problem is that SPC clearly needs some sort of minimum capacity to OP count requirement similar to TPC. Without that, in the limit all you need to do is take a really powerful server with a few GB of DRAM, add a UPS and you could post up totally insane $/OP numbers, but they would be as meaningless as these Kaminario numbers are.
Take a Intel E7 box off the shelf (IBM x3850 X5 is nice), load it with DRAM and RAMdisk software (e.g. SuperSpeed) and voila, you have up to 3TB of DRAM-speed disk...from a real vendor...with real support...for 1/3rd the price.
FYI, this is what SAP uses for their HANA in-memory database.
I was wondering how it Kaminario could have possibly come up with an 'architecture' to make DRAM go thousands of times slower than it should. 3.4 milliseconds? That's ridiculous.
I looked at the benchmark docs...this thing is nothing but a rack full of forty-seven Dell server blades (cheapest available) connected to a FC SAN, with RAMdisk software, software RAID and an APC UPS to make it "non-volatile". Oh yeah, and a 1,900% markup piled on for good measure.
No, really. I'm not kidding.
One of our primary intentions in going through the SPC process was to test the SPEAR architecture’s scale-out capabilities. The SPC-1 put the K2 through a heavy workload over a 24-hour period. It came through with flying colors showing that Kaminario allows you to achieve high-end performance without suffering from bottlenecks of any type.
Some comments have focused on the fact that the K2 configuration included DRAM. So of course, the IOPS were going to be super fast. From our vantage point, the media type matters less than the architecture. We happened to certify the K2-D in this test result, but SPEAR will also enable high-level, scalable storage performance for Flash. That is the beauty of having a powerful and flexible architecture. It provides you with more choices and the ability to adapt as newer memory types emerge -- All in an environment with advanced data protection.