Super-Fast Hard Drives

codders writes: "An Australian startup company, Platypus Technology, has launched a range of RAM-based solid state drives. These QikDRIVEs can offer sustained data throughput rates in excess of 110MBps and can be up to 8GB in size."

Comments

Why?

[HeghmoH]

I don't quite understand why you'd use one of these things. It seems to be that you'd get much better speed by simply putting all 8GB on your motherboard. Surely there are motherboards that can take that much memory, in addition to a 64-bit processor to address it all.

I guess the part about an independent power supply is useful. If the power goes out, a UPS is going to be able to power a dinky little card a lot longer than an entire server. However, if your server is under enough load that you need one of these things anyway, you probably have multiple safeguards in place should the power die. You could always keep a hard drive on standby and write your ramdisk to it when the UPS notifies the computer that the power's dead.

So, am I missing something? Is it less practical to cram that much memory onto a motherboard than I thought?

Required Microsoft Bashing Post

[Stoutlimb]

Microsoft should buy up this company, and then make these drives as cheap as possible for everyone to use. Reboot time would would go down significantly! They would save billions in not having to make their code more efficient.

"The hardest thing to understand is the income tax." - Albert Einstein

Before anyone gets carried away...

[orpheus]

Before anyone gets carried away, please remember that these are volatile RAM drives.

The Platypus overcomes the HDD's primary liability (read/write latency) at a serious cost to the HDDs primary function: reliable storage. Note that it doesn't even have an on-board battery. It simply has a separate external Power supply and (optional) UPS

While a UPS is wonderful for keeping my system running, it's much less reliable than it needs to be if an outage (or office idiot kicking the plug out) means I lose *all* my data (sales for the day, etc.) In a sense, the platypus drive is not much stabler than having 8GB of system RAM and *no* HDD ["not much" is relative. The MTBF of a UPS is orders of magnitude less than a good HDD)

I doubt the usual high reliability filesystems could maintain a RAID/HA type redundant backup to disk precisely because the RAM HD is so much faster than the disk. It would be like having a scribe backing up your HD to quill-and-scroll -- the more you utilize the tremendous speed of the RAM HD, the farther behind the disk will fall (and thrash).

It's a nice product (though hardly a new idea), but I see it having limited application (e.g. as a HD accelerator in some server applications)

Perhaps someone can do a hardware workaround using an intermediate NVRAM between the SDRAM HD and the hard disk, using principles borrowed from both cache technology and High reliability file systems. But it'll take a bit of work.

Is there already a solution out there? Or is this essentially just a giant unidirectional HDD cache, good for serving up data faster than an HDD, but not good for critical rewritten data?
_____________

I've used a prototype (not from Platapus)

[darial]

Unfortunatly, I'm under NDA, so I can't tell you everything about it, but I have a fairly good grasp on the performance of this class of device.

The biggest problem with 32 bit machines is not the 32 bit int, which is really sufficient for most things, it's the inability to address more than 4 gig of memory. This provides a relatively clean solution to that problem by using this device as swap. The burning question, of course, is performance... how much worse is it than on board memory?

The answer is, of corse, it depends: If you are in a single process enviroment, the time it takes to swap pages is somewhat killer, because the machine justs sits on it's ass while the DMA moves the block. Now, don't get me wrong, it's a lot better than disk, but it's not like real meory.

However, on multi-process machines like servers, it's great. There is a delay for the page swap, but the other processes keep the cpu busy and the DMA keeps the bus busy. Since throughput is more important than response time, this is almost as good as onboard ram. But, you say, this is MORE expensive than real RAM? not really... for an app like this, it will be an smp machine anyways, and the difference in cost between comodity x86 parts and a 64bit+8gig-uberboard setup from a proprietary vendor is so great that you could buy one of these things with the spare change. This could easily save many tens of thousands on certain types of server projects.

Some more info

[Chairboy]

I read about these in March. Here's some info from then that might have changed in the interim:

$1,538 for a QikDrive1 with up to 1GB storage.
$9,840 for a QikDrive8 with up to 8GB storage.

The QikDrive is on a PCI card, but has its own internal power supply for data security. Presumably to keep the drives from being wiped by a system power failure.

They support between 15,000 and 20,000 I/O transactions per second (versus 200-300 for Winchester-style drives)

Eliminating Latency is the big win

[billstewart]

Devices like this have been around for a decade or more, but as memory keeps getting radically cheaper they become more interesting. The big performance win isn't usually how many MB/s throughput you get, though that's nice, but it's the elimination of rotational and seek latency which can be 5-10ms - a long time for a database transaction, and even with journaling filesystems or datastructures you've got to deal with some of it.

Be sure to use a good UPS with the things, and make sure your powerfail shutdown procedures work well.