Slashdot Mirror
Getting Rid of the Disks
Kneht writes "Dan's Data has an interesting article on what it would cost to get rid of your HDDs and replace them with SSDs because hard drives suck. Several aspects are examined, such as required UPS, compact flash, etc. Read the article and you may get a new appreciation for your lowly 7200rpm drive."
Funny, I was just thinking that I should start using 120GB disks as my removable media.
There's a lot of research going on in this area. In particular, there's a newly completed Ph.D. thesis studying a persistent memory/disk hybrid filesystem for linux, named conquest. The performance is quite impressive, although the reports are that it's nowhere near ready for use - the term 'researchware' gets tossed around a lot.
Basically, by storing metadata and files smaller than 1mb in memory, the typically accessed information is much more convenient, and the larger files left on the disk are typically in their 'best case' (it's much more common to read large files than to write them, and typically they're read in some near-linear order: if you watch a moving, you may skip once or twice, but then it's sequential reads). The combination seems to work quite well: We compare Conquest's performance to ext2, reiserfs, SGI XFS, and ramfs, using popular benchmarks. Our measurements show that Conquest incurs little overhead compared to ramfs. Compared to disk-based file systems, Conquest achieves 24% to 1900% faster performance for working sets that fit in memory, and 43% to 96% faster performance with working sets larger than the memory size. .
Video for Online Dating Profiles
http://ask.slashdot.org/article.pl?sid=0
-- -- --
Help my mini cause: My journal
When you look at something like this, it makes $5000 for 20GB seem like a conservative estimate.
Well, I haven't read the article ... but ... have you ever tried booting an OS from a RAMDisk? I hear people liked doing that with Windows 3.1 - sort of an instant-on feature. While that sort of blazing speed is unnecessary for most storage needs, it would make operating systems less susceptible to hard drive induced interface lag. And stuff. STUFF!!!
They could use FRAM (Ferromagnetic Random Access Memory)
It is as fast as RAM, but is non-volatile. Oh, and its endurance is unlimited. Right now they aren't big enough, but a the technology improves...
It might be that tape drives aren't really hugely cheaper than hard-drives. Lets go for the 20Gb Internal Travan from seagate. $180 for the physical drive and one tape. Compared, Western Digital 20 GB, $63.
So lets assume that the cost portion wasn't stacks 3 to 1 in the favour of the hard-drive. We also have the performance factor. I've supported these beasties. They are slow, especially if you even think about using them like a hard-drive for random access storage (which regretably HP did at one point)... the benefit comes in easily storable and removable media. It might be cheaper to buy 5 hard drives to do your rotation on, but its much more bulky and more labor intensive to do. Thus the 3 to 1 price tradeoff for using the slower tape for archival purposes outweighs the cost problems for some people.
Now, lets assume that this solid state is meant to do exactly the same as a hard-drive (which by the description of the article, it is.) We're looking at a 100 to 1 price tradeoff. The only way that kind of increase in price becomes worth is if your doing some highly critical things which absolutely must be done faster. The average game of Quake doesn't need it.
Thus, hard drives, could they be better? Yes. But if the next alternative is that much more
pricey, chances are they are good enough.
Last I checked, you would only have to spare one chip worth of space to store the parity information, regardless of how many chips are being used to begin with.
-Lucas
The current basic and EXTREMELY old computer architecture (which is CPU, memory, storage device, and IO devices) already solves this !
You store everything you need on the STORAGE DEVICE, and access stuff by copying it to MEMORY. If you what you need to access is big (like a database) - add more memory.
Actually, the only difference between MEMORY and STORAGE DEVICE is speed. If they were the same speed, we wouldn't have needed one of them. Shoving the memory away from the processor is like saying 'let's put a hard drive instead of memory - that way we'll have hundreds of GBs of memory !'.
To be fair, I'll add that it might help on 2 occasions:
1. Systems which are memory-limited - like my PC which is limited to 4GB. But I'm guessing that computer manufacturers will continue to expand this as needed (both for PCs and servers).
2. Loading-up such a system - reading those GBs from an HDD to the memory will take longer than loading it from memory.
But other than that, I think that stuffing the memory in the storage device and saying that you have a fast storage device might be true, but it's plain stupid !
However, there is a need in 2G-4G drives. Mainly in the corporate market. I've got this nice new Dell workstation at work for development and to my suprise I noticed it had a 8Gig harddisk. We're supposed to store our stuff on the network so, basically it only has apps and the OS on it. My question was: where the *cencored* did they get these small disks. Well, I just launched the disk administration and fair enough: 40Gig disk, with 8Gig partition. A complete waste.
Ahhh...the great dumpster continuum. Many a free computer will be found there. -- sowth (748135)
What if someone started making SSDs for the consumer market, though? How cheap could they be?
Produced at modest volumes in the USA (not made by the boat-load in China), we've been looking at somewhere in the $250 to $300 (usd) range for the bare board with 16 or 20 DIMM sockets, IDE interface, and power management circuitry with aux power inputs.
The unit is planned to fit into the form factor of a cdrom drive, which allows just enough room for 20 sockets and a couple inches to pack in all the circuitry, IDE and power connectors. There just isn't room for a battery, so the plan is to have 2 or 3 "aux power" connectors that accept 9 to 12 volts. We'd make a battery pack that fits into a 5 or 3 inch drive bay and recharges itself from PC power, so you could connect 1, 2, or maybe even 3 battery packs, or maybe a battery pack and 12 volts from some external source like a "wall-wart" power adaptor plugged into a cheap UPS, or maybe something a bit more "reliable". I'm not sure what the battery pack will cost, but it's hard to imagine it'll be over $50-60 even if we splurge a bit for a fancy microcontroller-based rapid charger and advanced battery monitor.
Today, 512 meg DIMMs are the most affordable, and today's pricewatch says about $40 for PC100-SDRAM and $46 for PC2100-DDR. Prices fluctuate quite a bit... a few months ago the 512 meg PC100-SDRAM was $30. But assuming you pay $40 each for 20, plus $280 for the bare drive and $60 for a battery pack, that puts you at $1140 for a 10 gig ultra-ultra-fast drive. Ouch. Even if the prices drop back to $30, which puts you under four digits, it's still quite expensive.
But not as expensive as the article claims.
Anyway, at this point the project is pure vapor. The earliest you might see it would be about one year from now, but 18 months is more likely. Even though DDR is more expensive today, the design will almost certainly use DDR because it is expected to become cheaper and remain more easily available for the years to come. It's also quite likely I'll do serial ATA only, as S-ATA is going to become the mainstream down the road, and it's already gaining acceptance now. My hope is that 1 and 2 gig DIMMs will become more common and their price/byte will come in line with the 128/256/512M sizes.... 'cause there's no way we're going to get more than 20 DIMM sockets into the 5.25 inch drive bay form factor.
The project also has a number of technical challenges... including the difficulty of connecting that many unbuffered DIMMs (the design will need 4 or 5 separate memory channels and a lot of buffers & PLLs that there aren't really room for on the board).
Well, enough vapor for one day.
PJRC: Electronic Projects, 8051 Microcontroller Tools
I use a belkin usb 2.0 removable storage case. works very well, even has an external fan and can easily remove hd's on the fly.
usb 2.0 is very fast
Does anybody know how big disk tracks are these days? If 2MB was enough on a 20GB disk, does a 200GB disk need 20GB, because the tracks are 10 times as large, or does the disk have 10 times as many tracks of the same size, or somewhere in between? The price of memory hasn't come down as fast as the price of disks, but it has come down a lot, and 10MB of RAM costs about $1 - even though the price of disks is really competitive, drives might as well have as much as makes sense for current geometries and speeds. The sizes are still likely to be on the order of 10MB, not 256MB, and since there's got to be _some_ chip there, it's cheaper as well as more reliable to just make the chip big enough rather than adding sockets for plug-ins.
Large quantities of write-cache on a disk drive are bad, though, because they're not backed up by battery - the system needs to know that when it's written something to disk, it's really written in some form that can be read back later. Read cache is harmless, because losing it just loses a bit of repeatable fetch work - you need enough to cache a couple of tracks of data, but more than that doesn't usually accomplish much, unless there's a big mismatch between your disk speeds and the bus that transmits to your system memory.
Caching cards are usually silly, unless they either provide battery backed-up RAM or are part of RAID controllers where they can help in the assembly/disassembly process. Their main purpose is to make up for limitations in operating system caching design (i.e. they help Windows a lot more than Unix) or making up for other hardware limitations (e.g. CPU RAM limitations, or bus speed differences, or letting you run server disks off the otherwise-unused AGP port instead of the PCI bus.) Their other main purpose is to take advantage of memory speed / price differences - disk caching works just fine with cheap PCI-100 memory, while system RAM needs to be the fastest Quadruple-Data-Rate Gigahertz-RAMBUS Quadruple-Price memory you can buy to keep the CPU running at maximum speed, so if you're buying large quantities of the stuff, it's sometimes worth spending an extra $50-100 for a card that can hold lots of cheap memory.
Battery-backed RAM cards are actively useful for applications that need secure writes, such as database commits or NFS writes. A decade or so ago, the Legato Prestoserve NFS accelerator cards had a meg of battery-backed RAM, which was enough to commit writes to while waiting for the disk drive to spin. This meant that you could respond to NFS requests in sub-millisecond time rather than waiting 10ms or more for a disk to seek and spin (seek time was still slower than rotational latency back then, plus your request might be queued with other disk requests), so you could handle one or two orders of magnitude more requests per second, and a megabyte was more than enough to buffer traffic from a 10mbps ethernet. Database transactions might be generated much faster than NFS requests, but it was still enough to handle caching for a lot of disk space.
Bill Stewart
New Fast-Compression-only CPR http://preview.tinyurl.com/dy575ks
The operational costs of using that tape were generally higher than the disk drive - a good tape robot and automation software can reduce them, and to make up for the lack of random access on the tape, you need to make the tape from backup/mirror disk rather than from working disks (while queuing up journalling request for the drive you're making the tape from), but restoration was the hard part.
It's possible that tapes are still cheaper for offline archival backup - the weekly stuff you dump in the safe deposit box so you can defend yourself against that patent lawsuit three years from now, or get that copy of version 0.6.1.8 to support that customer who's got a problem with the embedded device he built last year that uses your code. I've never found tapes to be highly reliable; you certainly don't want to trust them for restoring last night's build when the developers scrambled things this morning, and one reason you can trust them for your weekly archives is that if you lose Week N-48, usually Week N-47 or N-49 will have a copy of what you need.
But you've also got to keep a spare tape drive in your archive vault, and a spare copy of your backup software, because it's likely that any Really Bad Event that wipes out your onsite disk storage will also trash your tape drive, and tape drive formats seem to change at least annually, or at least the interfaces and backup software do, and the Non-Random-Access performance of tape and the popularity of compression means it's highly likely that the data format on the tape is much different than a disk file system; you can't simply find an old ATA controller or USB1.0 bus and expect to plug the thing in and read it. You don't necessarily have to store a backup tape robot - a single drive of the appropriate type is probably adequate - but you at least need the drive and all the software.
Bill Stewart
New Fast-Compression-only CPR http://preview.tinyurl.com/dy575ks
Unfortunately, LVM is on its way out. 2.5 kernels no longer have it, and it appears that the jury's still out regarding its replacement.
Six sick
Hmmm... that came across sounding harsher than intended. Sorry about that.
Still, even though I agree with the conclusion of the parent post, I also think that there's something to my point. 20GB isn't top-of-the-line capacity for either hard drives or tape drives, and older hardware often goes up in price before going down (compare SDRAM vs. DDR at Crucial, for example). So right away I'm inclined to think that there's something odd about this comparison.
The funny thing is, if anything, the grandparent post is too nice to tape drives. I can't find that tape drive for just $120 anywhere, and hard drives can definitely be had for a lower cost per GB.
The special cases are things like /tmp, which look like disk drives but mostly contain files that are created, used, and destroyed, and never really need to be saved on disk if there's enough cache space to keep them. The tmpfs file system type was designed to optimise these - it stores files in RAM and uses the virtual memory mechanisms to handle its data rather than a separate disk partition, and can really speed up applications like compiles because there's no need to wait for disk latencies or to even bother the disk bus with writes in most cases.
Bill Stewart
New Fast-Compression-only CPR http://preview.tinyurl.com/dy575ks
With any type of real market, these prices should come down very nicely.
-- -- --
Help my mini cause: My journal
I don't know where you're shopping, but at one online retailer I use 10-packs of Memorex DVD media are priced right now as follows:
That's right: DVD+R is actually cheaper than DVD-R. DVD-RW is almost 50% more expensive than DVD+RW while also lacking some DVD+RW features like defect management and lossless linking, and being very slightly less compatible with consumer A/V gear (dvdrhelp.com).
You can get DVD-R media for $0.70 apiece if you buy a hundred or so of some off-brand that won't work in most drives. You can do the same with DVD+R. The claim that DVD+R is "several times" more expensive was never true, and is not even close nowadays.
Slashdot - News for Herds. Stuff that Splatters.