DSI Delivers up to 3GB/s with Solid State Disk

olivesaregross writes "'Running at what the company says is 250 times the speed of conventional hard drives, it won't come cheap, but it will be fast. It uses DRAM memory to store data instead of spinning platter hard drives, giving an access time of just 20 microseconds.' It still does use platter-based drives but it's a cool idea anyway. Techworld has another story on it."

Comments

Pricing

[Faust7]

Traders link to the system over FDDI, T3 or ATM links, and the Eurex back office servers connect via 2Gbit/s Fibre Channel links and switches to the SSDs. The system uses DSI's 3200 solid state disks, with two to eight Fibre Channel ports that can push out 250,000 IOPS - up to 3Gbit/s - and contain 16-64 GB of capacity. There are two hot-swappable power supplies and three hot-swappable drives per DSI 3200. Uptime is five-nines - 99.999 percent.

Well! A consumer-level version ought to be cheap in about... ten years.

Re:Pricing

[b0r0din]

From the article:

"Trying to get the general population to think beyond the big RAID systems is our biggest impediment to solid state disk acceptance..."

That and the hefty pricetag, I'm sure. Obviously the better demand, the better a chance the price will go down, but seeing as RAID stands for Redundant Array of Inexpensive Disks (for some people) may explain why it's so popular. Same reason most people don't drive Ferrari's - it costs like 1000 dollars for an oil change.

I don't see this having much useful business application for the time being outside of very high enterprise computing, although it's about time. I'm still waiting for optical storage drives.

Re:Pricing

[ktulu1115]

The I can mean either Inexpensive or Independent and the D can be either Drives, Disks, or Devices.

I never understood that... RAID has historically been set up with SCSI drives, (ATA RAID has been more recently developed/implemented) I would hardly consider them cheap compared to ATA (altough recently I have noticed the price differiental somewhat lessened)

Regardless, I still use almost all SCSI in my boxes.

Interesting Uses Possible

[Pavan_Gupta]

I can imagine this kind of technology being really applicable in situations where large databases are in use -- but potentially, slightly cheaper then just keeping the entire database in ram. I think it would be interesting to use, but a bit more interesting to play with.

Re:Interesting Uses Possible

[Neon Spiral Injector]

But the entire DB is in RAM. But it isn't local to the processor, it is on the other side of a FC bus.

When this makes sense, is when you do want your entire DB in RAM, but don't want to put the required amount of RAM in each node of a cluster. So you just attach each node to this SAN.

26.6666666 times faster

[Anonymous Coward]

Wasn't there already a solid state transfer rate of 80GB/s reported from SGI/Cray a year or so ago?

Slashvertisement?

[stratjakt]

It still does use platter-based drives but it's a cool idea anyway.

From the article, I gather these are merely SAN boxes with up to 64GB of DRAM, fiber channel output, and 3 hot-swappable hard drives that act as backup.

Has a record been broken? Has anything special happened? Sure this is high-end stuff, but it doesnt seem new or particularly exciting.

Caching?

[amalcon]

Is this sort of thing more or less expensive than plain ol' RAM? If it costs more, then just caching 3gb of data from disk into memory at bootup is more cost-effective. If it's cheaper, then perhaps people will start using this technology for swap space, etc. In any case, I've been waiting to see an HDD using solid-state RAM for quite some time now. If we're lucky, it'll be cost-effective before too long.

Consumer edition

[swb]

How many 1G sticks of RAM could you put into your standard Firewire/USB2 enclosure? Why couldn't someone make a USB2/Firewire/SCSI enclosure that the host system saw as a mass storage device but was actually just a smaller version of the above? It might be really useful for some DB applications, video editing, etc.

I can't imagine that an enclosure of that type would run more than $500, plus the cost of the RAM that went into it. It might not be consumer cost effective, but it could be worthwhile at the prosumer or low end, where the RAM disks shown on /. are almost never affordable but by the richest organizations.

Re:Consumer edition

[Anonymous Coward]

It ought to be significantly less than $500, closer to $300. Consider that you all you need is:

1) A nimh battery for backup
2) Sockets for 16 dimms
3) Physical port for firewire
4) FPGA to talk firewire out one end and ddr sdram out the other and battery recharge/discharge control out the third end.
5) Board and case.

Economy of scale could push something like this down to well under $50 without RAM.

Solid State Discs have been around for a while

[Burdell]

I've got a mail server that uses an SSD for the mail queue filesystem. It is great for that because the random I/O transactions per second rating is 10,000 (vs. a typical hard drive thrashing hard at 150 tps).

The SSD we have is a Nitro!Xe from Curtis, Inc.. It looks like a standard 3.5" wide 1" high Ultra2 SCSI drive with an 80 pin SCA connector. We have a 2G model with a 2.5" notebook drive for backup (it has a battery to dump RAM to disk on power off) and it greatly improved the performance of our mail server (high performance mail queue is all about I/O TPS).

Wafer-scale integration?

[steveha]

Whatever happened to wafer-scale integration?

I read an article about this years ago. The idea was something like this:

Memory chips are made on wafers. They are made side-by-side, then sliced apart, then each tested and mounted in a package. (Then eventually mounted on a little circuit board, and thence into our hands to install into our computers.)

The idea was to make a wafer of memory chips, but not to just have them side-by-side; actually have traces connecting one chip to another. Then use the whole wafer as a RAM unit. You would need to test and find any defective RAM chips in the wafer, then cut a trace (or burn out a fuse, or whatever) to disconnect them from the rest of the wafer. (Not too different from bad-block management on a hard drive, really.) Finally you could make a stack of these wafers in a box, and sell it as a disk drive.

This should be much cheaper than current RAM-based disk drives. It would be slower (the traces connecting the chips would be slower than a direct memory bus to each chip) but still way faster than a drive with moving parts.

My understanding is that wafer-scale integration isn't very interesting for most applications, but for the specific niche of RAM-based storage units it seemed promising. Clearly I'm wrong since it didn't happen. Anyone know why?

steveha

The RAM itself is one thing...

[Kjella]

...that you can actually do in software (get a PowerMac or AMD-64, load it up with 8-12GB of ram). What usually cost money, is to have some sort of flush-to-disk feature.

I'd love a SSD at least big enough to boot from, to combine with some other fanless stuff to create a 100% fanless, no moving parts PC (except from burner, which is silent when not in use). That + GbLan (to copy everything in from fanless machine, no damn spinning CD/DVD) using a direct crossover cable to a file server, preferably in a sound-isolated galaxy far, far away.

That is my dream for my next setup. I've looked at doing the same simply dragging DVI + USB cables + external burner at machine, but it's not that great. A network cable can go so much longer...

Kjella