Slashdot Mirror
Linux On Solid State Disk
Blah writes: "A while back Slashdot made reference to The Platypus Solid State Disk. The boys down at LinuxWorld.com.au have scored themselves one and given it a look over. The article has some pictures showing just how much SDRAM this thing has on it, as well as graphs which compare its IO and transfer rate performance against that of standard SCSI disk."
- A.P.
--
* CmdrTaco is an idiot.
"Remember when the U.S. had a drug problem, and then we declared a War On Drugs, and now you can't buy drugs anymore?"
There are applications out there where $60K is a small price to pay to increase performance rather a lot, and $60K starts looking small when you start pricing Sun E10000 servers and such.
The big value to SSD comes in when you've got one of those situations of heavy database updates where eliminating latency time is a big win. If throwing on a $60K SSD allows downgrading from a $1.5M server to a $1.1M server, that was evidently a very good buy...
A CF memory card system that doesn't allow you to hit it hard with vast numbers of updates just doesn't compare. And it's still hardly cheap; there aren't $60K units, but there aren't 8GB CF memory cards, either...
If you're not part of the solution, you're part of the precipitate.
Is Samba a big enough project for you?
I don't think he's talking about systems actually required for flying the plane - things such as targeting systems (where information on which locations are planned targets could be valuable to the enemy) or map data (which would tell the enemy how much of their defense system has been located) would be more likely candidates for such treatment.
Then again, it could just be that the Air Force doesn't like sharing its sooper-sekret pr0n files with anybody else.
Yeah, 'cause those people in the non-English-speaking countries aren't very intelligent at all, are they? Look at the Japanese! Idiots - the whole lot of 'em!
I mean, think about it...ram is not _that_ pricey. There must be a lot of research dollars to compensate for. You can get 512megs of pc133 for ~US$200, so why does it cost outrageous sums for the drives? It's not like scsi itself can make up all that cost. Heck, even old edo would give faster performance than our current drives. Slap a bunch of that in a box, put on a scsi controller and I think a lot of people woud be happy. Or even better, put it in a pci card. The main problem with the pci card being physical space...
-- Who is the bigger fool? The fool or the fool who follows him? --
First, it has to be powered up to work. The external power supply is almost redundant. I won't be rebooting my production linux box more than once or twice a year, and even then, the reboots will be planned. I could easily back up the drive without the external power supply. If I were to suddenly lose power, it wouldn't matter anyways since the external power supply would most likely go down as well.
Second, for 8 gig models, having a separate PCI card holding the memory makes sense. But for less than 2 gigs, you will probably be better off just using a ramdisk. Not only will this allow you to have more control over the actual memory allocation, there shouldn't be any dramatic difference in performance. As I said before, a sudden loss of power for your server is just as likely to take out the power for the drive as well, so you're not in a much safer position the other way.
Just a few thoughts.
-Restil
restil@alignment.net
Play with my webcams and lights here
I don't see what the fuss is all about. I can put a couple of gigs of RAM in my main machine, bot up the OS, and load a gig or so RAM drive, and I've effectively got what everyone's talking about here, without any extra hardware (save the ram that is). What's the big deal? Am I just missing somehting? I'm really not trying to start a flame war here, I just don't get it...
LOAD "SIG",8,1
LOADING...
READY.
RUN
Wouldn't it have made more sense to test this against similarly priced storage systems? In the $5000 range you have fibre channel raid controllers with some storage. A good FC card will push around 200 MB/s. Why not have a measure of MB transferred per second per CPU percentage? Put that next to $ per MB storage and you get a real comparison.
This card is limited to 100 MB/s and is only 32-bit 33 Mhz so can only be grouped one per bus in order to maintain that speed. Meanwhile, most FC RAID cards are 64-bit 66 Mhz, run around 200 MB/s, support multiple cards before maxing out their target bus. For $5000, you are going to get much more storage than this thing and it will be faster. I just don't get it.
I have one technical issue with the article, too. It contains the following line: "Current PCI bus speeds are limited to 33MHz, however, 64 bit PCI bus systems are in development and have speeds of 66MHz." This isn't correct. Both 64-bit and 66Mhz PCI systems have been around for some time. I was at the Microsoft Plugfest for Windows 2000 and Millenium testing my 64-bit 66 Mhz fibre channel card in systems from various vendors. This was back in December of '99. Also, the signal rate and signal width are not automatically linked, although most 32-bit buses only support up to 33 Mhz and most 64-bit buses support up to 66 Mhz.
-- soldack
The US Air Force uses solid-state disks in at least some of its aircraft. They load the software right before takeoff. The idea is that, if the plan goes down or is captured, the pilot just has to power it down and all the software is lost, and then the plane is useless.
- - - - -
Napster-to-go says "Fill and refill your compatible MP3 player", which is a lie. It's not MP3. It's WMA with DRM.
PC133 => 133MHZ * 64bits (8 bytes) = 1GBps.
PC66 => 528MBps
Difference in price for standard SDRAM is neglegable. (pricewatch has 512Meg sticks for under $200, meaning the $5,000 price-tag is most likely NOT mainly comprised of DRAM)
But here's what happens when you underclock: The pipeline get's slowed so your latency increases.
Random read access should take a hit of about 5 clock ticks per access (actually more because of intermediate custom hardware). So random single byte reads suddenly are slowed to 105MBps, but since we were only interested in a word, we only got 13Mega accesses / second. By staying at PC133, you effectively double that minimum rate.
Now the difference in price from CAS3 to CAS2 PC133 is significant, I'll grant you.. Probably not worth the premium.
I admit that most if not all "virtual disk accesses" are going to be in 512B blocks, which comes out to 16 indepedant cache-line-bursts (8B/cycle * 4 cycles), and should thus take overhead(approx 5 cycles) + 4 cycles * 16, or under 70 clock ticks for a full sector read. That's about 1.9Million sectors per second theoretical peek (not bad). At that point, we have to contend with main-memory BW saturation and CPU over-head for disk-drivers. I believe that depending on how intelligent the drivers are, the PCI bus isn't the real bottle-neck for over-all system performance on such a ram-drive.
I am curious about the prospects of converting this PCI card into a UATA-100 hard virtual drive. You'd have higher peek bandwith, PLUS you'd be able to perfectly emulate a hard drive.. However, there is probably an advatage to putting memory on the PCI bus - namely that the OS drivers could directly access the media in little segments based on the actually requested data instead of duplicating disk-block-buffers in main memory, just to ultimately copy out to user-space.
-Michael
-Michael
Microsoft product suspected.
Linux zealot team scrambled to liquidate target.
ETA 6 minutes.
I think the real question on everyone's minds as we examine this product is obvious. Can a boost in I/O even as massive as promised by this piece of hardware actually save a poor, unwary server from the Slashdot Effect? Now *there's* a benchmark I'd like to see.
the only thing useful applications of solid state drives i can think of are:
1) doesn't require backing store, like main memory does
2) can use slower (cheaper) dram than main memory, since the bus is the bottleneck
3) if you're out of dimm slots, it could be useful to use as a swap device
4) as an fs journal device
unless they can sell these SS drives for less than the same capacity dimms, there's not going to be much of a market.
also, the article is wrong about pci. 64bit 33mhz and 32bit 66mhz pci slots have been available for quite a while.
Why didn't they make this an AGP card? It's a dedicated port designed for fast I/O. I know that AGP is the reincarnation of VESA, but does anyone know any reason why this wouldn't work?
Sort of. Its a SSD, but it connects to the SCSI2 interface on the Proliant 8500. 4GB of storage, but it isnt seen as a hard drive. What it does is as you boot and the OS loads it automatically starts caching everything to the SSD, when it gets full the controller uses the SSD as much as possible, while relying on the regular scsi drives when the info wasnt located in the SSD. Made SQL server scream though, real nice. 35,000$ though, so I doubt we will be getting alot of em.
----------------------------------
Looking for hardware (Currently need: Large Etch-a-Sketch) Have one? See my journal!
Isn't that pretty dangerous, though. I'm sure that the engineers have thought of lots of horrible scenarios and worked through them. But, what if something happens to the power in the aircraft (hit by lightning, say) and that device looses power. The plane's system can't reboot then? The pilot is dead in the air?
I suppose that the pilot just ejects and the plane is totalled on the account of a faulty power supply. Sounds about right. 1 billion dollar plane lost to a couple thousand dollar part.
that they're talking about military things, do you? "I've got the ball" indeed....;-)
Vintage computer games and RPG books available. Email me if you're interested.
That link again :-): www.dansdata.com/cfide.htm.
> A high percentage of my working day is spent waiting for compiles, as even a single change to a file requires on the order of five minutes of compiling and linking. A lot of that is file read/write time. If I could write it to memory-speed output rather than disk, I would be a happy man
Uh ? Put more RAM. Put even more RAM. And some extra RAM. Then use a ram disk for your object directory, and keep a lot of ram as the file cache. On a bsd, suppress atime update on the directory containing system include/libraries, or mount it read-only or copy it into a ram drive. Remove atime from you sources too.
> According to the task manager
Oh, I guess what your problem is. You use an OS that have a journaled meta-data filesystem (so sloow sync write for each file) and that have *very* high fragmentation (spend most of his time seeking).
Cheers,
--fred
1 reply beneath your current threshold.
Bottom line is that you need to use the right tool for the job. Sometimes it's a SSD, sometimes its real disk.
Don't forget that ram disks generate less heat and use less power and have no moving parts compared to a drive array.
Quantum has had solid state drives for almost 4 years now. They pioneered the field and their scsi SSD's blow the doors off anything out there. And with an added benefit, it's native scsi, no special drivers needed, access times in the 50ns range, as opposed to the standard 5-7ms for even Cheetah drives.
"See, we plan ahead! That way, we never have to do anything now."
Although this is quite funny, after Doug Miller's recent comments would the pot please refrain from insulting the kettle.
This isn't really a new idea, because I've had linux running on a 16Mb solid state Sandisk for a couple of months now... makes a great item to have in a router. The disk has an IDE interface and performs very well.
Skiers and Riders -- http://www.snowjournal.com
Solid-State hard drives are so expensive because they use SRAM, not the DRAM you are referring to. SRAM, or Static RAM, is an entirely different memory technology from DRAM, or dynamic RAM. It's main selling point is that it is MUCH faster than DRAM. The problem with it is it is much less dense than DRAM and uses a lot more power. These things make it much more expensive than DRAM which is why you don't use it to expand the memory of a PC. In fact, SRAM is the kind of memory used in on-chip cache in microprocessors because of its extreme speed.
If you have battery backup for the solid state cache and if it's reasonably large, then you get pretty much the same effect as if you had a disk consisting of all RAM. Sun used to sell something like that for speeding up NFS called "PrestoServe".
Of course, Linux already has non-battery-backed-up RAM caches built in. The interesting thing is that even if the kernel crashes, you can usually recover the unwritten data from RAM and (after verifying something like a checksum) write it out to disk (this fails, of course, on platforms where the BIOS messes around with memory before the OS boots).
So, the best thing to do is probably to have battery backup for your RAM and processor and use an operating system that recovers information from RAM after a crash. That way, you can use normal RAM for caching, and if your machine crashes, you can recover quickly and with no or minimal loss of data.
You do realize Australia has only 10 million people right? and what does english speaking have to do with anything?