Slashdot Mirror
Costs Associated with the Storage of Terabytes?
NetworkAttached asks: "I know of a company that has large online storage requirements - on the order of 50TB - for a new data-warehousing oriented application they are developing. I was astonished to hear that the pricing for this storage (disk, frames, management software, etc...) was nearly $20 million dollars. I've tried to research the actual costs myself, but that information seems all but impossible to find online. For those of you out there with real world experience in this area, is $20 million really accurate? What are the set of viable alternatives out there for storage requirements of this size?"
Why is it that 90% of "Ask Slashdot" pieces seem to boil down to "I have no real world experience, and I'm just wondering how I can solve problem X for Y dollars when twenty different vendors all sell solutions for 100 * Y dollars?"?
--
Twoflower
the new 320 gigabyte harddrives previously mentioned. And you divide 50000 (50TB) gigs by 320. you get an approximate cost of having 50TB by multiplying that by 350$ the appoximate cost of the drive. However, with that much data a RAID is certaintly in order. So multiply the number of drives by 1.5 or 1.75 to get the number of drives needed for a RAID. Then multiply that by 350. This comes out to a little over 80000 dollars. The only cost left is the cost of all the raid controllers (expensive) and networking all the drives together. So for the raw storage of 50 terabytes it costs about $80,000. If you were to buy ultrafast scsi drives instead of the 320GB drives the price will be multiplied by about 3 since a 100MB super fast scsi drive is also about 300$ with 1/3 of the space. So that brings it to $240,000. Add to that the cost of labor and all the other hardware and I don't see how it could come out to more than 1 million dollars. I'm not an expert, but just doing the math it seems that more than that is too much.
The GeekNights podcast is going strong. Listen!
Add to that switches, routers, T1s, building, cooling, etc. Now if you need that to be robust you would build RAID units and that could double the cost.
Imagine how long FORMAT C: would take...
It's more involved that how many bytes you need to store, of course. How fast do they come in and go out? How often do the bits turn over? How reliable does the data need to be, and how fresh the reliability (do you need to mirror it real-time at a remote, hardened site, or back it up once a month)? What systems does the data need to feed and be fed from? What are your labor costs (tape changers, administrators, etc.)? How much wood do you need to buy for office furniture ?
Because it makes a nice change from the "I'm stuck, somebody tell me what to do!" pieces.
sorry for sounding a bit trollish, but the current replies here seem to follow the formula of checking the biggest ide drive on pricewatch and multipying that out to give you a number.
:)
forget all that.
if all you wanted was a pile of ide hard drives, maybe this would be ok, but anybody looking for 50TB of storage is not just looking for some disk to hold the pr0n they downloaded last week. large scale storage systems need to manage multiple host access to high speed (15krpm U3SCSI) drives in flexible raid configurations with maximum redundancy, high speed caching (with GBs of RAM to do it), fiber channel switching, cross platform capability, high end management and monitoring, HSM backup and data migration, offsite vaulting of disaster recovery data, power and air conditioning, and a fat service contract from the vendor. none of the above are going to be found at pricewatch.com.
your best bet is to talk to multiple storage vendors about your needs. call up EMC, Hitachi, IBM, and Fujitsu to start, them let them see each other's numbers. With the amount of money that you are going to spend (and it almost certainly will exceed $10 mil - but maybe not $20), each of these vendors will do backflips to get your business (and EMC is particularly good at junkets - take them for all they're worth
I am not an expert in this field, but Google was willing to tell me lots.
RaidWeb sells rack mountable RAID units that take IDE drives and have SCSI or fibre connectivity. A 12-bay 4U SCSI (with 12x 120Gb IDE drives) system comes in at just under $8000, giving over 1Tb fault tolerant storage. There are several other companies that have units like this.
Rackmount Solutions sells rackmount cabinets. A 44U cabinet with fans, doors, etc. will come in at around $3000.
In theory, a single cabinet could house 11Tb of data, and cost around $91000. This still doesn't consider cabling, cooling, power distribution, networking, a proper server room (air con, false floor for cables, access control), and in all likelihood one or more controlling servers.
More practically, depending on how they are going to make this data accessible, you could be looking at 9 raid units per cabinet plus 3 2U servers and a switch in the remaining space. Each server can support multiple SCSI cards and gigabyte networking. Such rackmount computers will set you back in the region of $6000 (incl. network and SCSI adapters, excl. software).
So you can call it $100,000 for 9 Tb storage ... $600,000 for 54Tb. That doesn't answer the management software question, and may not be a suitable solution. But it sure is a lot cheaper than $20 mil ;)
i-name =twylite [http://public.xdi.org/=twylite], see idcommons.net
Search terms: IDE Raid Chassis
Sponsered link: raidzone.com
Their 4U 2T system goes for $25K, so 50T would be about $750K and fit in 2 1/2 racks. They claim that they will be doing iSCSI soon, but right now it's just NAS. Still, this is a far cry from $20M. If budget is a concern, you can figure out how to use an array of NAS in place of a SAN.
If you are hell-bent on SCSI or FC, you are going to be into serious dough as SCSI drives are almost 10X the price of IDE at this time, and don't come with as large of capacity (which means that you will need more rackspace, chassis, power, etc.) $20M is probably not too far off. Modern IDE drives with dedicated smart controllers are really not too bad. Just keep a pile of them to swap out bad ones as you are going to be going through drives pretty quick.
With the size of your drive array, backup is going to be a serious issue. You are going to need a multi-drive robotic array of good size. Those are not cheap either.
Get a clue man.
Where is your failover?
How are you going to connect this disks together? NFS? Samba? That kind of speed (or lack of) is not an enterprise storage solution.
How do you replace disks as they fail without taking stuff offline?
Conformity is the jailer of freedom and enemy of growth. -JFK
i don't know nearly enough to put such a thing together, but i do know enough to know that every real-world project probably costs 50x what a geek-fantasy basement equivalent would cost.
Because in this case, it's pretty obvious that the prices are overly inflated. He's paying almost a thousand times what the raw drives go for.
I think it's pretty reasonable to feel that you could put something like this together for under $100K.
May we never see th
In raw disk storage, maybe. But you're forgetting actually putting those drives into a useable state with disaster recovery plans.
In other words, someone dealing with 50TB and who wants backups of that data will be spending many, many times the amount it would cost to just purchase enough hard drives to get the bragging rights of 50TB. And a backup located in the same room/floor/rackspace/whatever as the source data will be pointless in the event of fire, floods, nuclear fallout, etc. So, they would also need a way to transfer all that data to offsite backups in a timely manner (waiting five weeks for a full backup to transfer over a 100Mb/s pipe would probably not be acceptable).
Aside from backups, how would the drives be accessible? Even as JBOD, you're talking 40 IDE/ATA controllers (assuming 320GB drives and 4 ports per controller), or 20 SCSI channels (assuming 160GB per drive and 15 non-host devices per channel) to support that many disks. You could also use Fibre Channel and get away with only a couple arbitrated loops. Physically, you're talking about hundreds of disks that need to be mounted somewhere, so you would also need dozens of chassis to hold the drives.
But, hundreds of disks in a JBOD configuration means you'll have hundreds of partitions, each separate from the others. Hell, if the clients are Windows machines, they won't even be able to access more than a couple dozen at a time. And even for operating systems with better partition/mount-point addressing, it would be unmanageable.
So, now you get in to needing a RAID solution that can tie hundreds of disks together. If you're talking about hooking these up to standard servers through PCI RAID cards, you'll need several of those machines to be able to host all the controllers necessary (especially if all the disks are not 160GB or larger each).
The only realistic solution for this much storage, at least until we have 5TB hard drives, is a SAN-like setup. Specialized hardware designed to house hundreds of disks in stand-alone cabinets and provide advanced RAID and partitioning features. SANs don't come cheap.
Add to the SAN the various service plans, installation, freight, configuration, management and the occasional drive swapping as individual disks fail and you've already multiplied that $50K several times, as a bare minimum (and you still haven't priced out the backup solution).
There's a lot more to it than just having a pile of hard drives on the floor. I wouldn't even be surprised if the drives are the cheapest component.
From experience (with EMC - Sun) your price tag sounds a bit on the high side, but not by very much. Considering that EMC storage (after all mission critical data should be stored on EMC/Hitachi/StorageTek, NOT on consumer IDE) costs much more than consumer IDE/SCSI (25 - 75x) and that's only the disks.
If you're going with EMC, you'll need to put those disks in something, like a frame (cabinet), and for your size, more like 5 cabinets. With that many cabinets, you'll need some sort of SAN switch and associated fibre cables (not cheap). That gets your disks into cabinets and all hooked together.
You wanted to access the data? Then you'll need EMC fibre channel cards ($15k a pop for the Sun 64bit PCI high end jobs). But you'll more than likely be serving data from a cluster of machines, so count on buying three ($45k) per machine (so each card is on a different I/O board hitting the SAN switch, redundancy)
Who's going to set this up? For that kind of coin, EMC (or whomever you go with) will more than likely set the thing up and burn it in for you on site. The price probably also includes some kind of maintenance contract with turn around time fitting the criticality of the system.
Yes, my 'big ass storage' experience may be limited , but I think that 20Million for 50TB installed/supported/tested by a big storage vendor is in the ballpark.
Good luck.
Figure you get two IBM Sharks with two expansion frames, maximum cache and 36GB disk eighpacks.
... about $500k
That's like $6MM for most customers.
Fibre channel directors and switches
Tape robot... $1 MM
Storage Mgmt software like TSM... $400,000
The extra $10MM is probaly for full-time consultants, a more expensive solution like EMC or a more fault-tolerant solution.
Conformity is the jailer of freedom and enemy of growth. -JFK
What they've prob got is a massive SAN (storage Area Network) running over 2 or more sites. If one site goes down you can run on the other and at 30 miles apart.
Also accessing this amount of data at reasonable high rates is expensive, think Storagetek silos, HDS SAN's etc etc. All this is highend very very fast stuff.
If you've got 50 TB of data running in an OLAP cube you've got to have massive IO capability to properly load and spin the cube around. Ie the cost ain't in the actual storage media, but the IO (esp if you've got a split system requiring multi-site system).
There should be plenty of examples of this sort of data storage now - telcos to web logs. Pricing, well depends on the deal you can get at the time...
When you get a Symmetrix frame from EMC, you also get a support contract. EMC will send multiple people to your installation for maintenance. EMC will remtoely monitor your Symm via modem. They will help you plan your storage needs (including what kind of backup and reliability you need). EMC will provide 24x7 support for everything you need. Then there's management software, etc.
Don't forget that the hardware isn't cheap: Frame, multiple redundant hot swappable power supplies (requires specialty power connection), dozens of scsi drives, dozens of scsi controllers, 10-20 fibre channel connections, an interconnection network between FC and SCSI controllers that includes fiber and copper ethernet, hubs, etc., and a management x86 laptop integrated into the frame.
$20 mil for this is a fair price in my opinion. Anyone who rolls their own is just insane. There are hundreds of engineers behind each of these boxes, and it shows.
No, I don't work for EMC.
Using your sig line to advertise for friends is lame.
Floppies. Lots and lots of floppies. They are so cheap right now! And the come in pretty colors too.
I know, IDE? Who the hell is using IDE/PC hardware for production data warehousing?
It's 10 PM. Do you know if you're un-American?
IDE-RAID with 3ware 7500-12 controllers and 3U 14-bay cases (available from rackmountpro, and probably others) could be one possibility, but I don't think you would get a 'flat' storage-space from it, probably have to be segmented instead. As others have pointed out NFS/Samba aren't really manageable ways to handle a filesystem spread amongst multiple machines. People who do this, like archive.org and google, have custom software to access the data stored on their machines. But it doesn't have to be that way forever...
I think iSCSI could give very interesting possibilities for open-source SANs using this type of hardware...maybe front-end servers which map requests as necessary to back-end servers holding the storage, you could have a rather nice fully-resilient highly-scalable system that way, which would just appear as another drive to a client machine, no NFS/SMB etc...
Ok, Lets see. 50 Terabytes divided by 600 megs per CD means you will need 83334 CDs (rounded up.) At about 20 cents each (retail) that should only set you back about $17k. Add in $100 for some of those heavy duty shelving units from Home Depot and a wintel box to read and write them, and you are looking at well under 20k for total hardware cost. At this point, just go hire someone away from their McJob for a reasonable amount to swap the CDs and you are in business.
Searching eBay for EMC provided some interesting results (these are mostly "buy it now" prices):
EMC Symmetrix 3930 w/ 12 TeraBytes = $57K
(With the proper drive configuration, this unit should be able to deliver up to 70TB in a single system).
This one comes with 12TB of storage (256x50GB HD's). If you throw out all 256 of those 50GB HD's (or just give them to me as a consulting fee for saving your company over $19.5 million) and buy 256X181 GB HD's, you're just short of you 50 TB mark (~46,336 GB).
On Pricewatch those drives come out at $999 ea x 256=$255,744.00 add the initial $57K and you've got a machine that meets your specification significantly less than $20mil
Here are some other EMC machines for sale on eBay:
EMC Symmetrix 3830-36 With 3 TB No Reserve! = $59K
EMC Symmetrix 3700 6TB w/Install & 1YR Mnt! = $48K
EMC Symmetrix 5700 3TB Storage System = $9K
This is what I found by doing minimal research. I'm not 100% sure that the Symmetrix 3930 can handle that configuration (its not my money) so before you go down this road -- do your research (better than I did).
-Turkey
You could buy IBM if you want to loose your data all over the floor. Why do people always reduce these conversations to price...data is priceless...would you send fine china in a paper bag across country and no insurance(IBM Shark, Hitachi), or in a double wall cardboard box with bubble wrap(EMC Symmetrix/Clariion)...if that data gets lost its gone history...hasta la bye bye...its not all about cost people, don't get burned buy the right tool for the right Job...
Power Corrupts,Absolute Power Corrupts Absolutely, leaving one person(group)in charge is absolutely corrupt.
I'm not sure what the prices are running these days, but back in 1999 I put together a 6TB system running RAID 5 on an all fibre-channel system using (at the time FC hubs -- switch fabric was too immature) StorageTek (aka Clariion) arrays for right around $2.5M.
Keep in mind, that's just for the disks, array controllers/cabinets, hubs, and Sun FC cards. No servers are included in that price.
There are so many variables that you didn't go into that it's hard to give you an educated answer to your question, but it seems feasible to get to around 50TB today for that kind of money taking into account the increased storage density that we've gotten in the last couple of years.
Does this stuff have to be online for immediate access, or would ti tbe acceptable for it to be online in a very slow filesystem and be available within 1 minute?
I built a system using spectralogic Bullfrog AIT changers, and LSCI's SamFS system. It sticks metadata for the files on your actual disk, but when you request one of the files, it goes to tape and gets it for you. For 50TB (uncompressed), you would be able to get by for under $500,000. However, that's without mirroring tapes. Trust me, you want to mirror your tapes. I've had them fail before. Figure double the price if you are going to mirror. Also, I'm not sure if the new AIT drives are out yet that will hold 100G uncompressed. If so, this will bring the cost down.
I know, the system sounds sketchy, but it works quite well. Seek time is definitely slow, but once it finds it on tape, the actual transfer is quite fast.
Need Free Juniper/NetScreen Support? JuniperForum
I've read a lot of books in my day, but quite frankly, most of what little knowledge I have comes from the kindness of people who have helped me to learn.
I don't think there's any excuse for asking a question without first doing a little basic research, but here we have somebody who has legitimately never had any experience with terabyte storage asking if there's a cheaper way. It's a legitimate question, and one that probably could not be answered by looking in a book. So the person here is right to ask, and has already gotten some very good answers.
I have a somewhat similar problem: how do I make sure that on the order of a terabyte of audio and video data survives the next hundred years? This given that the disk on which the first 80 gig of this data were delivered to me has two errors that have corrupted two of the files already, and the data isn't even a year old.
What I've been doing is asking other people how they've solved the problem, and also thinking about it on my own. It's how problems get solved. I've gotten some very good and thoughtful answers to my questions already.
The big factors in storage cost, breifly:
r) Reliability
s) Speed
c) Cost
In rough terms, c=s*r, meaning the cost will rise dramatically for high speed reliable storage versus low speed crap storage.
In addition, how the storage is designed (and how much more it can cost) depends a lot on data access patterns as well (read-mostly vs write-mostly, oltp vs dss vs datawarehouse vs
Maxtor has 0.3TB IDE for $1/GB. If you built a huge array of IDE controllers for these, your disk cost for 50TB would be around $50k. If some vendor actually built a beast with the requisute number of IDE busses and whatnot, the chassis might run you another $100k. All in all, real cheap storage. But it would suck on performance and reliability, put out too much heat and noise probably, etc, etc.
Highly available disk arrays with extreme disk platter performance and large amounts of caching can easily run $20 million for 50TB, if not more. There are middle of the road solutions though, it doesn't have to be that expensive unless you're going all out for huge concurrency and speed in an OLTP environment that requires 99.999999% uptime.
11*43+456^2
The trend is to use iSCSI on the network side and IDE on the hardware side. Since a network file
server only has FS daemons doing I/O, and the drives
are always hot, there is no SCSI advantage as there
is in a multitasking workstation environment.
-I like my women like I like my tea: green-
Oh, and I might also add the SerialATA, with it's
tagged command queueing, is very shortly about to
render the 300% SCSI price premium obsolete in all
but a few narrow verticals.
-I like my women like I like my tea: green-
320 GB = 156 / 4 IDE drives per box = 40 Boxes
I recently picked up a Promise Supertrak 6000. I droped an email to Promise and found out that one machine should be able to run two supertrack (hardware raid) cards. With 6 IDE buses (12 drives) and two cards per machine. You could setup each machine with linear raid and 24 320 meg hard drives. If you did that you would only need seven machines!
Realistically though you would need redundancy in this kind of system which means some lost storage. With raid 5 it would be more like 14 machines for 40+ terrabytes.
Now figure about 1000 bucks for a MB case processor gig of ram etc, and 300 for each card and $350 for each drive * 12. That comes to $5800 a piece. $3500*14=$81,200. If you want hot swap it would be +130 for the card Hot swap package, plus 9 more drives at $80 a pop is $720 per machine, multiply that by 14 and hot swap capibility will cost you an aditional $10080 dollars. That is a hot swap, raid 5 total of 91,280 dollars.
With a few top notch megabit switches with channel bonding, you're looking at $100,000. Can you hire me for the additional $19,900,000? I would like to continue to work in NY.
Novel theory: Modern Man evolved from psychopath
I have access to the capital. What do you need?
Explain your customer needs and how you are going to satisfy them and why you need the money now. If it all adds up, I can find you the money.
Capital is never a problem, it is an excuse.
Joe
Joe Batt Solid Design
Is HSM a solution? We rarely access very old data but we still like it to be easily available. With HSM we can move data to tape or some other cheaper storage while it still appears to be on the local filesystem. Applications don't know the difference other than they have to wait about 45 seconds as the data is fetched to local storage. In the end it depends on how you access your data. http://searchstorage.techtarget.com/sDefinition/0, ,sid5_gci214001,00.html
I gladly take the $170,000 pay cut for my foolish exuberance to 19,730,000. :)
I love Linux. :)
Novel theory: Modern Man evolved from psychopath
Now, those 28 drives will need to be attached to something. Maybe an Adaptec SCSI RAID 5400S, which is a four channel card that can accept up to 60 drives and is priced at about $900. Add to that a machine to put the RAID card in with at least GB ethernet, at around $6000, 3 40U racks at $2000 each and a UPS for each rack at $2500 each.
All told, that's $67,200 each for drives and arrays, $900 for the SCSI RAID, $6000 for a single box, $6000 for racks, $7500 for UPS's, at a sum total of $154,800 for a single 50TB array. Primary point of failure is the single box running it. For a backup system, running a full second array as redundancy would cost a net $309,600. All of this is not inclusive of labor, which for setup might run easily $100k. Thus, a redundant reliable RAID solution would run you $400,000. All that's once the 320GB IDE drive is released by maxtor.
Does that answer your question?
Please note, this won't be the best array money can buy, just a large array on the cheap. (what RAID was intended for)
- Disk Arrays
- Fiber Channel infrastructure (i.e. switches, HBA's, etc.)
- Tape Libraries
- Tapes
- Storage Management Software
- Backup/Recovery Software
- Disaster Recovery
- Ethernet and Fiber network management tools
- Raised Floor space, power, air
- People costs, including consultants
IF you use Tivoli Storage Manager for your backup/recovery solution (it uses the least tapes per GB backed up of any solution) you will need about 500 LTO tapes, at an average cost per tape of $110. That is $55,000. A tape library that can handle that many tapes online will cost you about $400,000. The software will cost you over $100,000. You see how the numbers start adding up? Throw in consultants at $300/hour (this isn't a skill set you pick up over night). 16 port fiber switches with GBIC's will cost you $25,000 each, how many of those will you need? Or do you need Director Class switches (likely), better quadruple that price for the switches. HBA's are $1500 a pop, you need two in every server, minimum, for redundancy. Your disk arrays have to extremely fast to keep up with the demand for data from the servers, or you will be I/O bound. We aren't talking about MaxAttach NAS here.You get the point I hope. $20 million is probably reasonable actually.
In my universe I'm perfectly normal, it's not my fault you don't live in my universe.
...I realize that accepted pricing is well above the price I mentioned. And yes, obviously I left out the maintenance.
The problem is that I find that corporate spending on IT purchases has gotten ridiculous. Let's buy a TEMPEST array! Let's buy something with a Sun nametag because the name sounds good! Let's buy a $2k piece of software for each workstation even though there's a free alternative!
I'm not saying that anyone *provides* something in the price range I was talking about. No one is crazy enough to do so, if companies are willing to pay much, much more. I'm saying that, if you're asking whether it's possible to *build* something like this for the price range I mentioned, off the cuff it doesn't sound so unreasonable.
Yes, a seasoned IT person who works with high-end systems like this will laugh. Why? Because they're used to paying huge amounts of money. Because it's an accepted part of the culture to throw down this much cash. What I want to know is -- how often do people question these basics? How often has someone said "Wait a minute...this is wrong."
Are you telling me that if you were in a third world country without the exorbant amount of funding that we USians enjoy, and someone asked you to put together a 50TB storage system for under $1M, you'd simply say "It can't be done"? No consideration, nothing?
I mean, when I look at the fact that the *case* on, say, a Sun high end system costs more than a whole cluster of workstations, I start to wonder just how much excess is going on here.
Say we take the bare-metal, dirt cheap approach. Grab a bunch of Linux boxes. Throw RAID on them configured so that 1/3 of your data is overhead for reliability, and a 100Mbps Ethernet card in each. The figure used earlier was $1 per gig. Put 6 200 GB drives in each. Throw down $250 for the non-drive cost of each system. You have 800GB of data on each system, 400GB of overhead. That's 63 systems. $16K for the systems, $75K for the drives, and we come in to $91K. I left out switches -- you'd need a couple, but certainly not $9K worth.
You'd need some software work done -- an efficient, hierarchical distributed filesystem. I didn't factor this in, which you could consider not fair, but there may be something like this already, and if not, it's a one time cost for the whole world.
Maybe another few systems up near the head of the array to do caching and speed things up, and you still aren't even up to $150K, and you have failover (at least for each one-drive-in-three) group.
I haven't looked at this -- it might be smarter, since you'd want to do this hierarchically, to have caches existing within the hierarchy, or maybe Gbit Ethernet at the top level of the hierarchy. And obviously, this may not meet your needs. But as for whether it's possible to build something like this for that much money? Sure, I'd say so.
Finally, existing SANS or any sort of network-attached storage are overpriced, no two ways about it. Very, very healthy profit margins there. Sooner or later, someone is going to start underselling the big IT "corporate solution providers" and is going to kill them unless they trim margins by quite a bit.
May we never see th
3ware Escalade 7500-12.
Twelve drives per card, multiple cards allowed.(I've heard of 5 controllers in a machine being used, however, PCI bus speed is an issue long before even the second controller goes in)
The Promise 6000 is a 6 drive controller. You can't do that doubled up of master and slave on each controller that you planned on doing.
So, would IDE really be that bad? Wouldn't it be better to put together a Beowolf cluster of smaller databases, each tasked with a portion of a search? Intelligent distributed processing is a much faster way to do a query of a database. If you have some large (but not unmanagable) number of notes, lets say 50 (one per terabyte), with backup nodes extending it to 64, any few failures would be correctable at full load.
I know that I don't have the skillset to put together the 50 Terabyte database right now, but I really believe that I could do it in less than 1 year, with half the budget, assuming free telecom to backup sites.
--Mike--
I included maintenance contracts in my off the cuff tape library price, assuming the selection of the IBM 3584 library and a 5 year life cycle. The 500 tapes is assuming a 2:1 compression ration, not necessarily what is actually achievable.
The other thing I didn't point out is that SAN solutions that are over 5 TB are generally custom solutions architected for a very specific environment. That raises the price because you are now talking about bringing in consultants (like me) who cost you $250 to $350 an hour, and needing hundreds of their hours. You could certainly implement a bunch of NAS boxes with big ATA drives in them for a few hundred thousand. But the system would be so I/O bound that it would be of no use to anyone.
In my universe I'm perfectly normal, it's not my fault you don't live in my universe.
and don't forget the big honkin' UPS to keep these things going long enough to shut down if power goes out. UPS' this size don't come cheap.
"For I am a Bear of Very Little Brain, and Long Words Bother Me"
The documentation clearly states as many as 12 drives. I could have sworn there were cables with two connectors with the card. I'll check tonight.
Same basic idea. What experiance do you have with the Escalade, I heard it was undiscontinued due to demand.
Novel theory: Modern Man evolved from psychopath
It seems that this question is extremely dependent upon the kind of application.
Are you mostly reading, or also frequently writing this data? Are you searching or doing indexed lookups? Is this a nasty bandwidth hog or a trickle? Is this a zillion parallel transactions or only a few users? What kind of latencies are expected? What reliability is required? What access is needed to historical data?
Consider some concrete examples that are *very* different from each other yet could each total 50TB and would have very different solutions:
- Video-on-demand system for a Hollywood studio deciding that peer-to-peer pirate systems can only be beaten by a legitimate system that is better.
- Online credit card transaction system for, say, Visa.
- SETI data that needs to be collected and searched for messages from extraterrestrials.
- Particle accelerator data that needs to be collected at truly horrendous rates.
- Lexis/Nexis database.
- Google database.
- Echelon data.
- IRS data.
- "Dictionary attack" database for a lone cryto-analyst.
The possibilities go on and on. At the minimum a 50 TB database might be a small number of equipment racks with a single computer attached to them, all totaling maybe $100,000.
And on the other end, I can easily imagine a system where $200,000 of a much larger total might be spent for, say, a terabyte of DRAM.
I can easily imagine a system with less than $5,000 of battery backed up power supplies, and I can imagine a system with hundreds of throusands in generators.
This question has enormous dynamic range.
-kb, the Kent who would enjoy working out solutions for specific instances of this question.
As long as you're here, I've got a question: Why do people buy systems like this?
I design and build software for a living, including stuff for banks, and I've been trying to imagine a system where I really need 50 TB in one place. Email for 10 million? Customer records for 50 million? A search engine for the entire web? For all of these, my designs would end up like Google: an array of cheap, commodity boxes that each are responsible for a portion of the data.
So is it that there are applications that really require this? Is it that some architects are used to drawing the one single "storage" icon and a $20 million bill isn't enough to make them say, "Gosh, is there a better way to do it?"
Or is it that the sysadmin costs and pain associated with maintaining 25 racks of gear make it worth coughing up for the centralized system in the long run?
I think you forgot read / write specs. Allow the data to be pullable slowly enough and you could do this real cheap.
I know that, being on slashdot a lot, you see a lot of people making cheap, shoddy, unreliable, and definitely not enterprise-class "solutions" out of string and tinfoil, but for a data warehouse application, that kind of cost is not unreasonable at all.
We have a somewhat-smaller situation at work, with a single Hitachi Lightning SAN providing our data warehouse nodes (two IBM p-series 680 servers) with a terabyte or so of fully-redundant fiber-connected disk. A single terabyte cost us nearly $750,000, and Hitachi bid competitively.
Enterprise-class solutions call for enterprise-sized wallets. Do not expect to slap together a few IDE drives and call it a day, unless you enjoy being fired.
- A.P.
"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?"
The problem is that I find that corporate spending on IT purchases has gotten ridiculous.
I find it's gotten too spendthrift, myself. We just got a Blade 2000 (anniversary edition) with only a single system disk. Why the OS of a 30,000 dollar machine is not mirrored is beyond comprehension, to me.
Let's buy a TEMPEST array! Let's buy something with a Sun nametag because the name sounds good!
No, let's buy those things because, if something in them breaks, the production payroll machine doesn't go offline. Or let's buy those things because, if something does break, I can have a tech on-site in 4 hours with a hot-swappable replacement part. Let's buy them because my customers (my users) won't notice the downtime while I pull a CPU module, PCI card, or disk and replace it without powering the server down.
Let's buy a $2k piece of software for each workstation even though there's a free alternative!
No, let's buy a $90,000 piece of software because it allows us to precision-machine aerospace parts more efficiently than hand-drawing the same models in two dimensions on a drafting board, or because we can run simulation testing on our airframe to see how much stress it can take before it destroys itself. Let's spend our money smartly to produce more revenue and profit.
Say we take the bare-metal, dirt cheap approach. Grab a bunch of Linux boxes.
I've seen horror novels with better beginnings...
Throw RAID on them
Apparently "throwing RAID" on something is good enough for enterprise-level.
and a 100Mbps Ethernet card in each.
This will work great on a network where every client is connected at 100/full, and the normal servers have fiber or gigabit uplinks. You may have gotten away with this in 1995, but it's 2002.
The figure used earlier was $1 per gig. Put 6 200 GB drives in each. Throw down $250 for the non-drive cost of each system.
$250 for the rest of the system? Motherboard, RAM, CPU, power supply (dual? Hah!), and case? Our AIX NFS servers have RAIDed MEMORY, not to mention at least triple the amount they'll ever need of that, CPUs, local disk, power supplies, and PCI expansion chassis.
You have 800GB of data on each system, 400GB of overhead. That's 63 systems. $16K for the systems, $75K for the drives, and we come in to $91K. I left out switches -- you'd need a couple, but certainly not $9K worth.
Yeah, you could just go down to CompUSA and pick up a few Netgear 8-ports. Nobody will ever need a VLAN. (The modules in our 6509s cost more than $9k.)
You'd need some software work done -- an efficient, hierarchical distributed filesystem. I didn't factor this in, which you could consider not fair, but there may be something like this already, and if not, it's a one time cost for the whole world.
Yeah, you could hack something together. Let us know how that goes.
Meanwhile, I'll be enjoying another day of outage-free administration, at least on the machines we built the right way.
- A.P.
"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?"
Regardless of quality, cheaper is always better, though. What part of that don't you undersand?
"Is". It's missing an apostrophe and two consonants.
- A.P.
"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?"
Bullshit. Expensive Sun servers crash all the time due to memory and CPU failures. The more CPUs and the more memory, the more chances for failure. These boxes do not have redundant CPUs and memory until you get into absolutely isnane price levels. If you care about reliability, it is better to have truly independent machines, and let the software handle the redundancy. Sure, mirror the storage, because you know hard drives fail, and have redundant network interfaces to protect against a witch failure. But don't forget that a "high availability" E6500 is 22 times more likely to crash than a "workstation class" ultra 1.
Why the OS of a 30,000 dollar machine is not mirrored is beyond comprehension to me
This is part of what I'm complaining about. Hardware vendors have sold users on expensive, heavily hot-swappable systems where they make huge profit margins. They work very hard to steer clients away from consumer-level stuff, where their profit margins are nearly nonexistent. If you're willing to make a system the fundamental unit of failure here, you can easily buy a $3K system with a second failover $3K system. Why pay five times as much so that you can swap out a CPU instead of just swapping out a whole system?
The whole measure-system-capabilities-by-dollar-value thing is what I'm objecting to -- your first response was "This is a $30K system".
No, let's buy those things because, if something in them breaks, the production payrool machine doesn't go offline.
I severely doubt that more than 10% of the people with TEMPEST systems actually need them. I was looking at one cluster of very overpriced and very underused set of TEMPEST workstations at a company a while ago. They would have been better off with some stock x86 machines.
hot-swappable replacement part
See above. It's much cheaper at this point to buy two consumer-level systems and let failover take over for one system than to buy a single high-end system.
No, let's buy a $90,000 piece of software because it allows us to precisions-machine aerospace parts more efficiently...
The price I quoted was $2k. You're listing $90K, which is well into the vertical application market. There -- yes, you don't have much of an option. You need an airfoil simulator that does foo, baz, and bar, and there's only one vendor with it -- you pay for it.
I'm talking about buying horizontal market things like commercial variants of CVS, compilers, or other systems where there are very good free alternatives, yet companies persist on evaluating things based on price.
Apparently "throwing RAID" on something is good enough for enterprise-level
Who's to say that this approach is fundamentally flawed? Sun? IBM? Of course they're going to scoff -- they've got machines and service contracts to sell. A high-level IT person? They've been suffused in the "spend lots more to get decent quality" propoganda from said companies for so long that it'd be hard to get an objective viewpoint.
and a 100Mbps Ethernet card in each
This will work great on a network where every client is connected at 100/full, and the normal servers have fiber or gigabit uplinks
Notice that I mentioned having the front-end systems, the ones doing caching, have faster interfaces.
$250 for the rest of the system
For a file server, very little is needed in terms of CPU juice, or RAM (before you start screaming about caching, as mentioned above, I want a systemwide cache sitting at the front of this). Make the cache able to cache anything on the SAN, so that you're efficiently using your resources. Why would I need PCI expansions chassis or RAIDed memory? I've already listed everything every box needs, and I'm willing to bet that the number of RAM chips you've had suddenly and unexpectedly fail (for God's sake, this is solid state storage) is right up there with numbers of servers hit by lightning.
Nobody will ever need a VLAN. The modules in our 6509s cost more than $9k.
Why would I want a VLAN within my storage system? To the outside world, this is a single entity. For that matter, Cisco systems definitely fall into my "overpriced because IT will buy it because it sounds sexy" category unless you really need the few systems that they do that *no one else* can duplicate in functionality. You can run VLANs off a Linux box.
Meanwhile, I'll be enjoying another day of outage-free administration, at least on the machines we built the right way
As I said earlier, I never claimed that this is available out of box right now -- just that you can build something like this. And neither did I say that your systems are outage-prone. I do think that name brand systems are oversold on vague reliability promises. Is my RAM going to suddenly fail? No.
I've found that the primary reason that purchases will spend their employer's money is the ASHF (Avoid Shit if it Hits Fans) syndrom. IT personnel are willing to make suboptimal purchasing decisions so that they have someone *else* to point to if something goes wrong. "Sun's supposed to fix that, not us." "This is a best-of-class component that failed."
Now to some extent, the corporate culture fosters this, but I just want to point out that every time I hear people bragging about the cost of the systems they administer, I wince and think about this.
My guess is that this is going to die over the next five years or so. At the moment, there's a glut of secondhand networking and serving systems available from dying dot-coms. Once that's over, though, you have companies in India and Eastern Asia that can't afford to waste the kind of money that US companies do on systems. So you get manufacturers (probably non-US) springing up to create low-cost systems that fill their needs, without the exorbant profit margins. Eventually, as reputations become established, they'll start selling to US corporations trying to bring down costs and compete with those foreign competitors, and overpriced IT purchases will be a thing of the past.
Linux is part of the advance front of this -- it's cheap to set up, runs on cheap commodity hardware (who's manufacturers make very little profit per unit), and you can build fancy things on top of it. As a matter of fact, that's most of the reason Linux has been propelled into the business market at all -- not because a bunch of geeks think it's sexy to use (though it sure would be neat if that *were* the reason), but because the profit margins are in a more sane range.
Almost all products follow a process of starting out very expensive, becoming more common and understood, commoditization, and eventual drop of profits to near zero. And once a product has reached the end of this process, bringing the price back up is very, very hard.
May we never see th
Expensive Sun servers crash all the time due to memory and CPU failures.
Really? Hadn't noticed. If by "all the time", you mean our E420 with an ecache parity problem (yes, this is a known issue with that series of CPU) which used to go down once a week until I took the faulty CPU offline from the command line, then, yeah, it crashed all the time.
The more CPUs and the more memory, the more chances for failure.
You obviously haven't heard about things like ChipKill and ELIZA fault-tolerance initiatives.
You can run a machine with a bad CPU for months without worrying about it, and bad memory modules can now be cycled out of use without even causing memory access violations.
don't forget that a "high availability" E6500 is 22 times more likely to crash than a "workstation class" ultra 1.
What are you smoking, and where can I get some?
- A.P.
"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?"
It's generally accepted practice AND documentation.
If one of your drives dies on the chain, sometimes it'll take the other drive on the chain offline with it. As well, you can't hotswap in a drive if there's another one on the chain. AND, it makes cabling more difficult when you have to put the two drives near enough to reach for the cable.
And the escalade is generally accepted as one of the best cards out there. They were discontinued, yes. But now they're back, they have 4,8, and 12 port models. As well, they have Serial ATA cards.
You know that I am talking about commonly available multi-CPU systems, and not exotic (and insanely expensive) systems with redundant CPUs and memory.
What are you smoking, and where can I get some?
Do you seriously believe that an E6500 or similar system will not crash if there is a faulty CPU? Despite your impressively low slashdot UID, if you believe this, you have virtually no experience with such systems.
What were they doing with them?
Basic numerical analysis.
voltage spike that makes it past the UPS
I've yet to see a spike damage even a system on a cheap surge protector, much less a nice UPS. I *have* seen surges over POTS lines damage equipment, though. Come to think of it, my neighbor's house was hit by lightning at one point, knocking out her modem...yet leaving her computer intact.
Side note: You would use g++ as a compiler for your product? The code it produces is about as efficient as a fully-loaded Excursion full of fat chicks
Not anymore. Take a look at the code that a gcc-3.2 build puts out...It's light years beyond the 2.7 and earlier era, the time that built gcc such a bad rep. It's competitive with the better compilers out there now (at least in generated code...Sun's C++ compiler compiles more quickly). Oh, and the good code generation is on the x86 -- never tried comparing recent builds on SPARC or PPC.
Case in point: we broke IBM AIX 5.1 a few months ago
So you're asking me both to believe that this had to be fixed immediately (as in, whatever you were doing before you broke AIX 5.1 was no longer an option) and that Linux wouldn't have been fixed quickly (and while there probably are issues that have taken a while to fix, I tend to see patch times that beat competing OSes).
They'll meet the ethernet card full-on and be very disappointed at what they see
You're talking raw streaming of a huge sequential series of reads, which may or may not be an issue here -- but that's besides the point. You're leaving out the possibility that data could be interleaved across different machines to avoid exactly this issue. Do it in software, I say -- it's cheaper.
simply because it can do one or two of the things a real piece of networking gear can do
Okay, I'll bite. Short of sheer mass bandwidth that you absolutely require custom hardware for, like a backbone provider, what specific features are you complaining about the lack of?
I've found they spend more in the short term to save more in the long term. If you think doing something the right way is expensive, try doing it the wrong way.
I agree that doing something the wrong way can be more expensive -- I'm just not sure that saving money necessitates "doing it the wrong way".
May we never see th
Don't you think people don't upgrade their systems or implement new ones?
/proc is inefficient. However, there's also a silly perception that unless something costs excessive amounts of money, it must not be up to par with the competition.
:-) However, I suspect what you meant was "I don't think you've ever been in IT handling thousands of users", which you would be correct about -- I'm an engineer, not an IT person.
We aren't talking about SANs any more? There isn't a lot of reason to be dropping new OSes or new servers on components of the SAN.
Just as I agreed that there *is* a justification for vertical-market applications, I'm not saying that every copy of AIX should be purged. I just think that items like these are frequently sold in situations where they are not needed. That doesn't mean that they're never needed. I don't claim that Linux is the best alternative if you're using, say, oh, a system that needs to dump process info from the kernel very frequently -- Linux's
I am beginning to wonder if you have ever worked in a company with more than 100 employees.
Well, you're definitely wrong in the literal sense.
Which would explain some of the different focus here -- you're complaining that given a list of options from different providers, no one currently gives you what I'm talking about. My interest is in adding another option to that list -- whether it's possible to create a new option for the prices being talked about.
May we never see th