Slashdot Mirror
Four X25-E Extreme SSDs Combined In Hardware RAID
theraindog writes "Intel's X25-E Extreme SSD is easily the fastest flash drive on the market, and contrary to what one might expect, it actually delivers compelling value if you're looking at performance per dollar rather than gigabytes. That, combined with a rackmount-friendly 2.5" form factor and low power consumption make the drive particularly appealing for enterprise RAID. So just how fast are four of them in a striped array hanging off a hardware RAID controller? The Tech Report finds out, with mixed but at times staggeringly impressive results."
"So just how fast are four of them in a striped array hanging off a hardware RAID controller? The Tech Report finds out, with mixed but at times staggeringly impressive results.""
So in other words I'll get First Post much faster since slashdot switched over.
Shai Schticks:"You don't make peace with friends, you make peace with enemies"
'cause regular hard drives usually survive 5 years in an enterprise environment, yep yep.
DRM: Terminator crops for your mind!
A 1.2 GHz processor with 256 DDR2 memory? Holy crap! That's faster than my new Celeron 220! And the perennial quesion: can this thing run Linux?
Is 4 of these in a RAID-1, running a seek-heavy database. Nobody does this benchmark, unfortunately.
'cause SSD's don't cost $300-$500 more than their spindle counterparts, yep yep.
Umm.... yes they do. You need to stop shopping for drives on Ebay.
a handful of selfish greedy people are no match for millions of selfish, greedy people -u4ya
woosh!
I'll be sure to do that, and replace them every 5 years when they run out of write operations.
Winchester drives, on the other hand, use a time-honored complex system of delicate moving parts, and last virtually forever. They certainly do not start experiencing sudden failures if kept in continuous service for more than 5 years.
Make that 228 years.
Life expectancy 2 Million Hours Mean Time Before Failure (MTBF)
Hint: learn about "wear leveling"
I will be surprised.
See, in the enterprise environment that I work in the majority of our big hardware is leased. I am quite willing to use what I can to maintain performance and reliability. That being said my system is built entirely on 15K drives of various sizes. I am not worried about five years or so of read/write that SSD drives have, all I want to see is a track record. I expect to replace most of the drives I have now within five years so this "five year limit" many like to toss out is immaterial to me. Reliability over that lifetime is of more importance.
Besides, the nice benefit of SSD drives is I don't need special enclosures (read: ones that can handle the torque these puppies can put out)
* Winners compare their achievements to their goals, losers compare theirs to that of others.
I really don't get this obsession with page files these days. Say you have 4GB ram and an 4GB page file. Memory is cheap these days, so rather than using 4GB of (relatively slow) SSD, why not just get another 4GB ram?
This is a very expensive solution. What part of Redundant Array of Inexpensive Disks don't they understand?
Your enterprise environment must not be hitting its drives very hard.
Where SSDs is in disk operations that are usually lagged out by seek times; a big unwieldy database that gets a lot of writes and no downtime, for instance, is happiest when it lives on a striped SSD array.
Coincidentally, this is exactly the type of workload which is most likely to shorten a magnetic drive's life.
DRM: Terminator crops for your mind!
> 'cause SSD's don't cost $300-$500 more than their spindle counterparts, yep yep.
Hint: Enterprise storage purchasing often looks at dollars/IOPS rather than dollars/GB.
25% Funny, 25% Insightful, 25% Informative, 25% Troll
It seemed a little unfair that they only used the nice hardware RAID controller with the Intel SSDs. I would have liked to see them use it with all the other disks to get a more level playing field.
I'm no expert, but wouldn't that be a redundant statistic? if it handles normal read/writes faster than a disk drive, then could you presume paging would be faster as well?
Although it would be interesting to see a RAM-less PC try and run on SSD's only... somehow using normal data read/write, and memory read/write on the same SSD (if thats possible). Guess that's what we'll end up with eventually anyways, where your amount of MEM is the amount of free-space you have on your SSD, no longer seperated components.
in 5 years they wont
er, "where SSDs shine is..."
DRM: Terminator crops for your mind!
I think this all has to due to the program or database you are running, a pagefile maybe more usefull than just ram. I all depends on the app you are using and how it works. Some items run better with a swap others just want ram.
Doom levels????
Office tasks???
Okay folks I can only see a few groups using this kind of set up.
Not one Database test?
I mean a real database like Postgres, DB2, Oracle, or even MySQL. Doom3... yea those are some benchmarks.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
Cheap RAID controller... check
CPU-bound benchmarks... check
Couldn't be bothered reading any more... check
What's on TV?
Test it on a better system then a OLD P4 cpu.
So are SSDs ready for prime-time?
Last year I attended IBM's Sydney Technical Conference and listened in to a presentation by ?????, one of the lead designer's of IBM's 3950 chipset.
Part of the presentation was on SSD technology. Whilst viewing the graphs showing SSD closing the gap in cents / gigabyte, someone asked the pointed question: "Is there any future in spinning platter technology?". The presenter actually stopped for well over half a minue (an age in public speaking) before replying carefully "I do not speak officially for IBM now, but I can see no future at all for spinning-platter technology. Not even for bulk storage"
As others have noted that once SSD drives are available in a price-competative form at volumes of around 150GB, tradditional drives will immediately and permamently exit the notebook market.
And I can wait for that to happen
"Enterprise storage purchasing often looks at dollars/IOPS rather than dollars/GB."
Which is good news, as this Intel Slashdot advert says its "compelling value if you're looking at performance per dollar rather than gigabytes.".
Fusion-io's SSD the iodrive is faster than this by far, I would suggest looking at their speeds on fusionio.com
SSD shouldn't be for paging. That would become very expensive (even with wear leveling) if you have a minimal amount of RAM (say 256M) to run large (say 16G) operations. It would also be slow since you have the overhead of whatever bus system your hard drive/ssd is connected to.
Technically hard drives aren't supposed to be paging either, it's just a cheap and simple trick to avoid having people pay a lot for (expensive) RAM or have their programs crash when occasionally they run out of RAM. However if your system is paging heavily it's better and faster with more RAM.
Anecdote: I worked at a place once where cheap ($500) hardware was sold as dedicated SQL/IIS servers (you could fit 10 of them in 5U) and a lot of customers thought they could run whatever they wanted (Microsoft ran MSN for a whole country of one for a while) in them but they only supported a maximum of 2G RAM (4G according to BIOS but the modules back then were too expensive). Of course PHB just said: let them swap and besides the heavy slow downs they ran fairly fine. Well, those heavy users all crashed their software-RAID's in less than a year (the heavy load made Windows get the RAID system out of sync and then you had the first hard drive fail). The temperature was fine but simply swapping out was too much for the cheap hard drives (Maxtor and Seagate) and they all failed.
Custom electronics and digital signage for your business: www.evcircuits.com
Redundant Array of Very Expensive Disks?
[citation needed]
Obama's legacy: (N)othing (S)ecure (A)nywhere and (T)error (S)imulation (A)dministration
I must be doing something wrong then. Should I put my computer in the freezer when I'm not using it or something, like, to keep it fresh longer?
MABASPLOOM!
Desktop memory is indeed cheap.
However I don't think i've seen a desktop board that could go over 8 gigabytes and most top out at four. Server boards can go higher BUT
* intel xeon boards requires FBDIMMs which are expensive.
* AMD opteron boards can use ordinary DDR2 but the CPU performance sucks compared to the aforementioned xeons.
note: i'm known as plugwash most places but i screwd up registering that here somehow in the past and now can't register
http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3403&p=4
[quote]Given the 100GB per day x 5 year lifespan of Intel's MLC SSDs, there's no cause for concern from a data reliability perspective for the desktop/notebook usage case. High load transactional database servers could easily outlast the lifespan of MLC flash and that's where SLC is really aimed at. These days the MLC vs. SLC debate is more about performance, but as you'll soon see - Intel has redefined what to expect from an MLC drive.[/quote]
So.. 100GB/day/drive = 100GBx4 = 400GB/day for (5*365)= 400GB*1825(days) = 730TB of data transferred. If you seriously go through THAT much data, you are either a pirate or you actually own your own movie studio.
Obviously these numbers can change, based on case temperature, wear leveling, etc. HOWEVER, they are what Intel states the drives can handle...
All modern hard drives are Winchester drives; Winchester drives are just the first iteration, made by IBM, who figured they'd ship two 30MB platters and name the hard drive after the Winchester 30-30 rifle. Who the hell modded you insightful, especially for claiming a system of delicate moving parts lasts virtually forever...
Support my political activism on Patreon.
If you go multiprocessor (not multicore) then you get much higher memory bandwidth (NUMA). Sometimes that matters more than CPU power.
Who the hell modded you insightful, especially for claiming a system of delicate moving parts lasts virtually forever...
What about watches? /sarcasm
You are using English. Please learn the difference between loose and lose; they're, there, and their; your and you're.
Or replace them in 5 years because better drives are on the wholesale rack.
Make that 228 years.
Life expectancy 2 Million Hours Mean Time Before Failure (MTBF)
Hint: learn about "wear leveling"
Hint: you should learn what is meant by "Mean Time".
New slashdot layout sucks.
I think you're comparing against SATA drives. People that worry about IOPS are normally using FC drives which are much more closely aligned in price with SSDs. (btw, been a while since I was in the market for FC drives)
SSD shouldn't be for paging. That would become very expensive (even with wear leveling) if you have a minimal amount of RAM (say 256M) to run large (say 16G) operations. It would also be slow since you have the overhead of whatever bus system your hard drive/ssd is connected to.
You talk like you know what you're talking about; but then the reader realizes you don't understand what happens when the CPU spends 99% of its life in wait state waiting for paging operations. Swap is not a high-intensity workload; swap workload increases six orders of magnitude faster than CPU workload, meaning when you start swapping, you spend lots of time swapping.
As the hard disk is external, this number increases with CPU speed; a swap operation taking 1,000,000 cycles on a 1GHz CPU (1mS) will take 10,000,000 cycles on a 10GHz (1mS) CPU. Triggering a seek operation between 4 and 9 mS on a 2.0GHz CPU (modern AMD) is a disaster; triggering these continuously, every 10mS, halves your CPU performance and performs 50 operations a second. Write operations take more than read operations, substantially, so we're talking 20-30mS, at which point ... if you're swapping even 2-3 times a second you notice it. AND ALL THAT SEEK WILL KILL HARD DRIVES.
Support my political activism on Patreon.
Most people have this mistaken belief that SWAP is interacted with as often as RAM (hundreds or thousands of times a second, at least; RAM is interacted with sometimes hundreds of thousands of times per second). They think swap is an actual extension of RAM, not a long-term slow storage shelf.
Support my political activism on Patreon.
I suspect the performance would have been a LOT better if they'd used something like the 3Ware 9690SA. 3Ware is also a LOT more Linux friendly.
Cheers,
It can be a good ballpark figure, to differentiate between enterprise class drives and consumer drives, but should NOT be an expected number.
There are too many things to take into account: temperature surrounding the drive, how many days it's on for, how long per day it's on for, how many writes to the drive, how much voltage is supplied, etc etc..
More about it here: http://en.wikipedia.org/wiki/Mean_time_between_failures
Problems with MTBF As of 1995, the use of MTBF in the aeronautical industry (and others) has been called into question due to the inaccuracy of its application to real systems and the nature of the culture which it engenders. Many component MTBFs are given in databases, and often these values are very inaccurate. This has led to the negative exponential distribution being used much more than it should have been. Some estimates say that only 40% of components have failure rates described by this. It has also been corrupted into the notion of an "acceptable" level of failures, which removes the desire to get to the root cause of a problem and take measures to delete it. The British Royal Air Force is looking at other methods to describe reliability, such as maintenance-free operating period (MFOP). Similarly, the National Aeronautics and Space Administration (NASA) is pursuing time to failure research using scenario and condition based methods derived from the field of prognostics.
Double woosh!
They break down eventually... watch smiths fix them. A coworker just had her watch (about 10 years old) gutted and re-geared. Very nice hand-crafted piece.
Support my political activism on Patreon.
Well to be fair Doom 3 still runs like shit on my rig
Does this take into account things like, half the drive is filled with stuff that doesn't change or does it imply nothing is on the drive and you just keep writing to it for no reason? ie that would drop those MTBF factors by half? Tried to wrap my brain around wear leveling and can't seem to grasp it.
It's obvious sarcasm.
Which require repair / reconditioning from time to time - no pun intended... okay - pun intended...
I don't think anyone should be using a page file at all if you have 4 GB or more of RAM. Maybe even 2 GB. It just doesn't make sense. With that much memory what good is a 512 MB page file going to do really? And if you're swapping more than 512 MB of RAM to disk your machine is going to be thrashing like mad and unusable anyway.
It's stupid that many OS's allocate 2 times your RAM as a page file. Are you really going to swap 8 GB of RAM to disk? I mean seriously, that would be unusable.
Even when I had 2 GB of RAM I never used a swap file and now with low-end machines running 8+ GB (only about $100 of RAM), page files just don't make sense any more.
The ratio of people to cake is too big
Here is a test with 4-5 drives and Linux software RAID 0:
http://21stcenturystorage.cebis.net
No one looks at just IOPS/$ without looking at size. It's probably something like IOPS/$/GB with different weights depending on the circumstances. No business is going to pay the same price for a 3x faster drive if it has 1/10,000 of the size unless money is no object. Then you're just concerned with IOPS, not IOPS/$.
Let me know when MRAM gets to gigabyte sizes. With unlimited writes and running at SRAM speeds, it could be the future.
"What's the use of a good quotation if you can't change it?" - Doctor Who
They must not have humor where you are from. The "Winchester drive" comment was obviously sarcasm, and the "watch" comment had "/sarcasm" at the end.
Said, "It's just like dice but it's got more sides And it tells me who lives and who dies"
The article that you point out is for the MLC NAND which is what you have shown. The NAND used in the X25-E series is SLC which has anywhere from 10x to 100x the write endurance.
Intel's X25-E Extreme SSD is easily the fastest flash drive on the market, and contrary to what one might expect, it actually delivers compelling value if you're looking at performance per dollar rather than gigabytes
I hope someone got a healthy commission from Intel for writing that...
Damn_registrars has no butt-hole. Damn_registrars has no use for a butt-hole.
Other than just using one of these Flash RAIDs as a swap volume, is there a way for a machine running Linux to use them as RAM? There are lots of embedded devices that don't have expandable RAM, or for which large RAM banks are very expensive, but which have SATA. Rotating disks were too slow to simulate RAM, individual Flash drives probably too slow, but a Flash RAID could be just fast enough to substitute for real RAM. So how to configure Linux to use it that way?
--
make install -not war
I was not the original poster - and I can't cite anything - but I can quote some numbers from our testing:
I can't remember the hardware, but memory or cpu was never the bottleneck. Easily seen with iostat -x 1.
Repair of a large MySQL table:
14 SAS disks in RAID-10: ~45 minutes.
2 X25-E in RAID-1: ~30 minutes.
1 IODrive: ~2 minutes.
QPS (70% read, 30% write) for a database consisting mostly of MyISAM tables:
14 SAS disks in RAID-10: ~4.500.
2 X25-E in RAID-1: ~6.000.
1 IODrive: ~15.000.
- This is not very scientific, but our numbers confirms it.
I have a feeling that Intel's Extreme disks suffer the same problem as MTron's higher end disks, namely very bad random write speeds.
However I don't think i've seen a desktop board that could go over 8 gigabytes and most top out at four.
Maybe this was true a year ago, but not now. I recently got a PC with an Asus P5Qpro motherboard - it supports up to 16GB. http://www.asus.com/products.aspx?l1=3&l2=11&l3=709&l4=0&model=2269&modelmenu=2
Those who can make you believe absurdities can make you commit atrocities. - Voltaire
Obviously, sarcasm is +5 insightful rather than +3 funny.
Support my political activism on Patreon.
Yes.
Adherence to the truth is a form of disloyalty.
what about penises? (peni?)
We'll be focusing our attention on RAID 0 today, but the card supports a whole host of other array configurations, including RAID 1, 1E, 5, 5EE, 6, 10, 50, 60, and 36DD. Ok, so maybe not the last one.
I like Tech Report but, seriously? I found that joke to be very juvenile and I have a hard time reading the rest of the article without a prejudicial eye.
MTBF is Mean Time Between Failures, (device is repairable).
For HDDs, you should really be talking about MTTF.
http://en.wikipedia.org/wiki/MTBF
If you're really looking for high performance storage, you should go with a DRAM-based solution. This has almost no latency and can scale to any interface. Depending on your budget, you can get SAS 3GB/s 2 ports with 32GB capacity for a bargain $24,000 (URL:http://www.solidaccess.com/products.htm/> and if you need more performance or storage space, spring for the serious iron--a FC 4GB/2, 2 ports at a mere $375,000.
No need to raid this puppy. Make sure you spring for the redundant power supplies and rack-mountable UPS.
MTBF is a highly inaccurate way to show how long you should expect a drive to live. The whole Seagate Fiasco is a prime example of why NOT to believe them.
Misuse of a statistical figure is a problem with those misinterpreting it. Obviously things have changed since schools taught the difference between the mean, the mode, the median, and the minimum. If I run an ISP then MTBF is useful for me to calculate costs, both in replacements and labour costs. It's not supposed to be a measurement for consumers though that will be buying single unit quantities.
Buying a hard drive is like buying a washing machine. If I'm lucky it will go on practically for ever. On the other hand if I'm unlucky it could die tomorrow. As Piranhaa says, there are too many variables. All I can go on is that if it comes with a garauntee of 3 years then I assume the manufacturers have designed it to mostly exceed that figure otherwise they would end up losing money on the product. I still have to ensure I have a contingency plan in case it breaks down.
Phillip.
Property for sale in Nice, France
I'm Betaing Windows 7. Before going to bed I set up a swap partition for it. After getting up the next morning and checking, it was full.
I have *no idea* what W7 put in there while I was sleeping.
Democrats or Republicans. They are both taking us to the same place and they are not afraid of us anymore.
The new I7 boards from Intel are either 12GB or 24 maximum.
Democrats or Republicans. They are both taking us to the same place and they are not afraid of us anymore.
Who shall watch the watch men?
Weaselmancer
rediculous.
Hint: You should learn that people tend to compare things using some "measure". In the disk world, that's MTBF. 2 million hours MTBF is comparable or better than other enterprise drives. Hence original poster is, indeed, clueless.
I'd run my page file on DRAM chips, if I could afford it. Serious applications might want a card that interfaces through the PCIe bus. I can't even afford to ask how much one of those costs.
The cost of that cleanup, of course, will be borne by taxpayers, not industry.
Your point? It's a highly inaccurate way used on _all_ drives. The point is that these drives are as reliable or more than other enterprise drives.
Must be all that pr0n I downloaded on that win7 beta box I rooted last night.
ULTRA WHOOSH
Fool me once, shame on you. Fool me twice, watch it -- I'm huge!
No matter how much RAM you give Windows, it will still page. It's to the point where people make ramdisks to put pagefiles on.
Not that I should talk... I have 1 gig of swap on Linux, and I'm thinking I could use more. Why? Because I have 4 gigs of RAM, and if I'm actually using even half of that, I can't hibernate.
Don't thank God, thank a doctor!
Think of it like this let's say you have a 1MB log file that changes every hour, and 1 MB system file that never changed. You keep writing that log file to the same place until 1,000 times. The drive then says ok if that file is really volatile let's swap the log and system file so the log file is in a place that's been overwritten 2twice, and the system file is in a place that's over written ~1,000 times. You can then write to that log file 1,000 more times and now the average usage of the disk is ~1,000 times in both areas.
Hint: Remember to include "Hint: you should learn what is meant by "failure"." when replying to morans who quote MTBF.
Hint: "Failure" means dying or doing stupid shit unexpectedly. If you're told to replace the drive ever year, or expect capacity to dwindle to 10% over the period, or other such nonsense, "failure" has a different meaning than what most people would think.
http://www.engadget.com/2008/12/11/fusion-ios-iodrive-tested-worlds-fastest-storage-confirmed/ , etc. We do HPC stuff at work, the Fusion IO guys are cool. Their products also run 3000-15000$ so it's not exactly a fair comparison.
Flash (and buses) have a looooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooong way to go before they could ever compete with current memory speeds.
EPIC WOOOOOOOOOOOSH.
...it's just a cheap and simple trick to avoid having people pay a lot for (expensive) RAM or have their programs crash when occasionally they run out of RAM...
Don't forget it was a tactic to breed lazy-ass dime a dozen programmers that don't give a crap about RAM usage...or any other HW issue for that matter...
Why does a freaking browser need 100MB of ram on a plain jain site like /.?
Imagine a drive 99% full of unchanging data, and 1% full of constantly changing data. The wear leveler could, every X number of writes, move 1% of the unchanging data to expose a fresh section of drive for the changing data. There would be a small performance penalty that could easily be adjusted by changing various parameters, and the increase in drive lifetime would be more than worth it.
Now, I don't know specifically if the wear leveling algorithms used by Intel's SSDs or Samsung's SSDs or random Chinese OEM's SSDs actually do this, but it's certainly possible.
main(c,r){for(r=32;r;) printf(++c>31?c=!r--,"\n":c<r?" ":~c&r?" `":" #");}
The MTBF is the same or better than other *enterprise* mechanical drives, at 1.2 - 2 million hours MTBF, let alone home user drives which are considerably less. Obvious troll fails.
Oh, you can take it for granted that they use anything you can come up with in half an hour.
They probably do lots more complex stuff, too. ;)
Hardly a fair comparison. If an enterprise drive doesn't get stolen on the way out of dry dock, it surely won't be long before the thing is being attacked by Romulans.
It would make a lot more sense if the test was re-performed with different SATA controllers. Specifically, on-board Intel ICH9 or ICH10, Promise, HighPoint RocketRAID, and Areca.
Had I done the review, I would not have bothered using an Adaptec controller. I haven't used their garbage for years.
That's only because we don't have -1 Retard.
You can get a socket 771 board with an Intel 5100 chipset, and it'll be compatible with DDR2 ECC memory. It does not require FB-DIMMs. Most of those boards are 32/48GB max RAM. Most newer desktop boards for AM2 and 775 "support" 16GB RAM, but the availability of 4GB non-ECC sticks is low. So you're stuck with "only" 8GB for playing 64-bit solitaire, visiting the interwebz, and listening to itunes.
I'll be sure to do that, and replace them every 5 years when they run out of write operations.
There are things that are read-mostly but still want large IOPS that are well suited for a Flash SSD - e.g. database indices, static web content, etc.
Also many database schema are split such that recent transactions are put in a log (e.g. the full inventory + deltas based on the day's transactions) and then there is a separate batch processing step to integrate the log into the main set of indices. This is often done only once a day as there are frequently complex business logic performed in this step that would kill transactional performance. So the log (and its associated indices) could be on magnetic media or RAM SSD, and the real indices could be on Flash SSD and you still wouldn't toast the SSD prematurely.
Even saturating 4 lanes of SATA would be much slower than DRAM. 4 x 300MB/s = 1.2GB/s vs 12.8GB/s (for DDR3 1600). Yes SSDs handle random requests better than magnetic media but still offer no where near the speed required from true "random access memory". It could work if you were still using a 486 maybe. In the future it may work, but I doubt it will be with flash memory.
Why not run RAID-5 (or 50 or 15) if it is seek-heavy?
Because four drives in a RAID-10 are three times as reliable as the same four drives in a RAID-5. Arrays of large drives are more vulnerable to drive failures during reconstruction than arrays of small drives, and RAID-5 is much more vulnerable to a double drive failure than RAID-10. In RAID-5, you lose data if any two drives fail. In RAID-10, you lose data only if the drives that fail are from the same mirrored pair, and there's only a 1 out of 3 chance that two randomly selected drives will be from the same pair. This makes it that much less likely that you'll have to roll back to the tapes.
That and you don't have to read-modify-write to calculate parity on RAID-10, as XanC pointed out.
When will someone come up with a hardware or software RAID solution to enable several USB flash drives to appear as a single drive on Windows? with relatively reliable & fast (12MB/s write, 30MB/s read) 16GB flash drives as cheap as £16 each I'd love to cram as many as I could inside my Eee and have them appear as a single drive instead of many individual drives.
To do something right, you often have to roll up your sleeves and get busy.
SSD drives are not as unreliable as people seem to think, It would take a *minimum* of 5 years of *continuous* writing and thats for the cheaper SSD's like you would find in a netbook rather than the $300 Intel top of the line ones, not to mention ones being used in a RAID would have a longer life.
Besides even if some blocks go bad you can map around them, the SSD itself might even do it.
Besides, you are unlikely to be using the same drive in 5 years time and magnetic drives have a much higher chance of failure, also when magnetic drives die often the entire drive dies at once, a few blocks failing is a much better solution.
In addition to that, the following is from the info on the SandDisc 3G drives: "The G3 SSDs provide a Long-term Data Endurance (LDE) of 160 terabytes written (TBW) for the 240GB version, sufficient for over 100 years of typical user usage.(2,3)", 4GB of writing a day is typical.
cat
Are you really going to swap 8 GB of RAM to disk?
Yes. When you hibernate your workstation, it writes the entire allocated contents of RAM to the page file.
Wrong. Hibernation has it's own storage for the current state.
The others have pointed this out, so I'll just stick with a nice informative link. Up to 24GB in 6 sockets, DDR3 1600, triple channel. $234 plus shipping today. They have 23 different ones today just that support the I7.
/happens to me all the time.
Help stamp out iliturcy.
Maybe it was caching and streaming out your pr0n?
The Fastest SSD on the planet is FusionIO.
I have personally benchmarked a SINGLE FusionIO drive in Redhat at 136,000 4k read IOPs and 96,000 4k random IOPs. Setup a Software RAID of four devices and get even better numbers.
SSD drives are not as unreliable as people seem to think, It would take a *minimum* of 5 years of *continuous* writing and thats for the cheaper SSD's like you would find in a netbook rather than the $300 Intel top of the line ones, not to mention ones being used in a RAID would have a longer life.
Pretty sure that's only true for a drive that is empty. If the drive is half full, or even 75% full, the SSD has a lot less unused space to spread wear over. That can cut lifespan down to 1/4 of the original.
(Unless it's a really really good bit of firmware that moves rarely written blocks over to blocks that are frequently written during idle time.)
Wolde you bothe eate your cake, and have your cake?
'cause SSD's don't cost $300-$500 more than their spindle counterparts, yep yep.
15k RPM drives and SSDs cost about the same on a $/GB basis, and on a $/IOP basis, SSDs absolutely wipe the floor.
Sadly, drop-in SSD replacements seem to be slow making into enterprise SANs. :(
On the more expensive drives maybe... but on the cheaper SSDs? Why should the low-cost drive maker pay for that complexity?
Wolde you bothe eate your cake, and have your cake?
I'm going to go out on a limb here and say that you should eliminate page files / swap space entirely.
I know there are edge cases, and one would be where you're moving something from disk into a ramdisk or some such thing, but in my experience... there's no circumstance under common usage where running out of ram and swapping to disk is preferable to running out of ram and having the program crash. To be honest, I would rather things hard crash than run in a slow degraded state.
And these days, with 8 and 16 GB of ram being sort of standard in servers, if something fills up 8GB of ram and you weren't anticipating it to do so, it's going to fill up another 1/4/8/16GB or whatever of swap.
Swap, I think, is a legacy and we should shed it sooner rather than later. Now, having said that, I'm not exactly brave enough to do that in a production environment.
~x
sig?
Seriously folks, once you start stepping up to these kinds of things, getting a Hyperdrive actually starts to look attractive, though that does depend on the workload. Unless the FusionIO PCIe gear gets more reasonably priced, a Hyperdrive is a more realistic solution. The big problem with flash SSD's is the wear due to full block/cell writes. Arguably, a striped RAID only makes that situation worse, because you are potentially more likely to write less than a block size (depending on stripe block size, awareness at the RAID controller level, and at the OS level).
Now if somebody can bring back the concept of a RAM card using the guts of a Hyperdrive, we may have something here. PCIe 2.0 direct interface like the FusionIO stuff to skip the physical SATA interface limits, make it a full size card to give the necessary real estate for the DDR2 RAM sockets, an exterior PCI card cover opening to shove a CF card in for backup purposes, and a spot for the battery.
Drool starts when you think of putting such cards in an external PCIe chassis like the ones from Magma
http://www.magma.com/products/pciexpress/expressbox4-1u/index.html
especially if you double up the chassis capacity ( I don't see anything really stopping a 1U going from 4 to 8 hosted cards physically), and a double iPass cable to a HBA with twin outputs to give you PCIe x16 speed.
Or you could just go the ghetto route and do a poor man's Thumper with vertical Hyperdrive 5's in a 3x8 array in a 4U chassis (maybe 5U due to drive length?)
M-M-M-MONSTER WOOOOSH...
I don't suffer from insanity, I enjoy every minute of it!
FBDIMMs aren't that bad these days either. Sure, it's more than standard desktop memory, but I paid $300 for 4x2GB sticks for my Mac Pro about six months back, and that same amount has already dropped to a hundred bucks for a 2x2GB kit (about $300 for the higher-density 2x4GB kit). No, it's not the $8/GB I'm seeing at some standard DDR2 sticks at Newegg, but it's hardly crazy-expensive either.
Hell, at that price, I'm almost tempted to get that 8GB kit just because I still have two RAM slots free in the thing, not that my current 10GB isn't cutting it (almost... I'd rather get more storage, or even a nice fast SSD for a boot drive)
How are sites slashdotted when nobody reads TFAs?
As SSD drives come into the market used, how will people know how close these drives are to "used up". That is to say, we will have to worry that these cheap drives on ebay will have lots of "bad" spots that can no longer be written. We are going to be needing a program or device of some kind that can certify the state of a drive so as to set a fair value on it. I expect a lot of unhappy people when used drives get installed and start failing soon after. There will have to be some pretty sleazy warrantees to cover used SSDs.
C-C-C-COMBO-BREAKER!!!
The firmware is exactly how you suggest.
The 'wear leveling' actually shifts logical disk addresses around the physical media to account for the fact that different parts of the disk get different amounts of activity. Each physical storage element will get an equal turn holding the never-written data, given enough time.
DRM: Terminator crops for your mind!
I use a large pagefile so that I can use Suspend-to-Swap. I occasionally do swap out most of my 2GB of memory when running stress tests, but I don't need to swap all of it back in at once so my system remains usable. In any case, since 8GB of regular disk space is dirt cheap there is little need to change this behavior (although swapfiles may make more sense).The strongest argument for disabling swap on a regular harddisk would be to force the OOM killer to kill a misbehaving process that suddenly wants to allocate all memory, but there are other ways of doing this.
My wife who loves me very much bought me the 28 port 512mb version of this hard drive controller for Christmas and I have loved it, but I feel like I'm wasting it by running it in a shitty Core 2 Quad motherboard instead of a Xeon motherboard.
I have actually run this exact controller in a Pentium M motherboard during my 3ware to Adaptec transition phase, and the results are entirely different. Especially since the Adaptec drivers are poorly suited for single core systems. After all, you just don't add a $300-$1700 controller to a computer that costs more to ship than to buy on eBay.
So far as I'm concerned, the test is entirely unrealistic and invalid. It's time these guys "rebooted" and got at least a quad core PC, at least using the x58 chipset, bus performance is less of an issue and peripherals should perform at their full potential.
The original poster was correct. Ik you look at the report you will notice that with most test the adaptec card is the "bottleneck". The test is really a comparison of different ways to connect a storeage device to a underpowered cpu.
using ntfs or a optimized raid FS will not change that bottleneck. Beside that, the raid optimizations you speak about are buffers and number of seeks, things where the ssd shines anyway.
If you want a good test ot get the last byte out of a ssd you need a test that shows where the bottlenecks are on this relatively new media. Did you niotice test where write amplicication shows up? (SSD are bad when down muliple small writes), did you notice that the XM-25 performs worse just after it was completely full?
I did not care to check, because these things are not in my personal budget.
That is a good idea, but the other half of the battle is ensuring that your PC gets fresh power. You can easily make sure of this by smelling close to the power outlet. Old power has a slightly, but distinctly musty smell, like French cheese.
8 of 13 people found this answer helpful. Did you?
The Intel 5100 series chipset uses plain DDR2, and they run a lot cooler than the old the FBDIMM boards.
Have you actually loaded it up with 16G of RAM? Often times boards advertise maximum capacities which aren't really ever tested at that capacity, or have other caveats (like the memory doesn't run at full speed anymore).
Game! - Where the stick is mightier than the sword!
Testing a SSD RAID 0 configuration on a system based around Pentium 4? I personally don't consider this a relevant test.
I'd say that this post made a rather insightful point. The tone was quite sarcastic, but that doesn't in any way detract from the underlying point.
I paid $300 for 4x2GB sticks for my Mac Pro about six months back(about $300 for the higher-density 2x4GB kit). No, it's not the $8/GB I'm seeing at some standard DDR2 sticks at Newegg, but it's hardly crazy-expensive either.
So if ~5x as much isn't crazy expensive, just how many times as much is? This is classic behavior that earns Mac people scorn, as opposed to people who just use macs.
What you describe here doesn't necessarily do what the parent describes. It might exist, but I have yet to see an implementation of either that takes in-use sectors and moves them to not-in-use sectors.
Re-mapping not-in-use blocks and selecting the least-used blocks for writing is pretty easy, and can be 0-overhead in terms of main storage bandwidth. But remapping in-use blocks requires stealing at least one read and one write cycle for every block you move. Those extra cycles detract directly from the available disk bandwidth, and would drive impede performance; it hardly seems like the sort of thing a company touting the speed of the disks would do.
You're assuming the system file is re-written at some point, so that it's previous position becomes available. What happens if I have 50% of my disk full of data that won't ever be re-written over the life of the drive? Does the drive interrupt my writes, do some internally copying, and then let me access it again? It does, is there any way I can make it stop, so I don't have unpredictable delays related to garbage collection?
The "small performance penalty" is potentially tripling the amount of time it takes to write a block -- you have to read the old block, re-write the old block someplace else, and then write the new block. It doesn't happen every time, but it's quite a penalty when it does, and there's no parameter you can tune to make that penalty smaller. If you're rating drives in IOPS/$ there's no "idle" time to do housekeeping like this, and it results in a direct performance loss.
That's the porn you downloaded. Because Win7 calculates wear-leveling by categorizing data types, Microsoft simply saves your solid state disk by putting whatever you're most likely to use more in paging files. Which goes for all your porn, given your surfing habits.
You didn't read the EULA?
No, no, no good. You actually did some reading before you did some posting. Get out of here. Just go. Leave. You're not wanted here. // jk . . .
Mon chien, il n'a pas du nez. Comment scent-il? TrÃs mauvais!
I hope you are going for +1 Funny...
What would be the point in getting DRAM chips on an extra card so that you can simulate a swap file that is effectively only there cos of a lack of... DRAM cips.
It doesn't make sense (unless you have maxed out the motherboard capabilities, but then you probably have bigger issues anyways)
Tom...
The point is that these drives are as reliable or more than other enterprise drives as measured with an inaccurate metric
In case I wasn't clear enough: "Yeah, I measured those two wall with my hand, and I'm 100% sure this one is bigger than than one by a hand's breadth."
Although I questioned the use of page files in this day and age, I do recognize that there are edge case usages.
Although I use no swap on my 6GB Linux at home, one of our dev Solaris boxes at work, with 16GB physical RAM does indeed use a lot of swap. We had to enable 16GB of swap on this box. It sounds awful, but we have a case where the main workload of this box is a gazillion Weblogic JVMs running and for the most part sitting idle (but devs want them always on, typical).
This works for us as buying an extra 16GB ram would just be a waste as 99.9% of the time the JVMs sit idly in the swap file. Performance is not crucial so it's fine.
Naturally we would not do this in a production environment.
Any decent OS would dispatch the read request and switch threads to one that didn't need to swap. Of course, that only works if your "active set" of ram pages actually all fit in ram, and you don't need to swap code pages just to run.
Much like what happens on a cpu with multiple threads supported by one physical core (like "hyperthreading") and a cache miss is encountered.
Uhhh, I don't like this car, it gets less than 1000 miles per gallon!
That's effectively what the original poster said. He pointed out having to "replace them every 5 years". Considering these drives have as good or better reliability compared to traditional enterprise drives - what's his point?
Presumably it copied everything in ram to the swap allow the pages still in ram to be discarded without waiting to write them to disk if your ram usage suddenly spiked and you needed to swap.
I think it's called "pre-emptive swapping".
No, the controller will move the system file even it's just been sitting there for months.
It doesn't need to happen nearly as often as the "real" disk activity. The disk could remap as few as 100 sectors a day and it would be enough to keep all the storage elements aging at about the same rate.
DRM: Terminator crops for your mind!
Actually, according to pretty much every review I can find, FusionIO's IoDrive spanks the pants off of the Intel Extreme, even in raid-0 eight drive combinations. It can saturate PCIe x4.
It'd be great if citations were required for broad claims like this.
StoneCypher is Full of BS
That's a good theory. Consider that you have one memory controller to access an area of memory (on NUMA, there may be 1 area of memory per core; on Intel chips there is no NUMA, it's one flat memory area). Now consider that the device's DMA I/O is tying up the memory controller. Now consider you just task switched, flushed cache, and the only way to get your task underway is to access memory. These memory access operations will have to wait for the IDE or SATA bus to finish playing with memory, leaving the process stalled in a wait state waiting for code memory.
Support my political activism on Patreon.
Also, the disk has a write cache so you can keep sending data even while it's moving data around. I don't think there is a universal implementation out there, but even normal HDD do the same type of stuff when it notices a sector is going bad. (I don't know the specific's on how that works, but look up Self-Monitoring Analysis and Reporting Technology or SMART HDD.)
Anyway, lazy writes are not a problem as long as you can read from the cache so I don't know how you would notice without doing some strange things with your data.
Hopefully the disk isn't tying up the memory controller when it's not actually transferring data, so you'd get 5ms of execution of another thread before the disk seek is done and it starts reading.
Not saying it would actually work, but it would be nice if it did.
In any modern operating system, including Windows , swap isn't just used for out of physical memory conditions. It's also used to "page out" portions of the operating system and libraries, shared objects, dlls, etc., that aren't being used at the moment. This actually speeds your system up by allowing more memory to be used as disk read/write cache.
I've looked at Linux boxes with 64GB of memory in them and only using 25% of that. I usually get asked by someone, "wasn't 64GB enough? Why is there some usage in swap right now?" It's normal, I explain. The kernel just pages out sections of Linux that aren't needed, to free up more RAM for filesystem caching.
I think perhaps Windows 7 just has a more aggressive way of doing this, probably because if you need to use some obscure Windows Directmedia SuperDRM doubleplusgood Plugin X, it's just as fast to reload it out of swap into memory as it is to load the binary from disk. But 99% of home users will never load that plugin so it can stay safely swapped out, giving you more precious memory for applications and disk cache.
"When the president does it, that means it's not illegal." - Richard M. Nixon
Ah, you're right. What about context switching again? In both directions, of course... and the cost of constant TLB and page faults during that 5mS? Not very much of a performance gain anyway. And again, under swap load (i.e. swapping several times per second), the performance hit adds up really fast. And again, wouldn't the level of swapping that would damage an SSD also damage a hard disk?
Support my political activism on Patreon.
though according to intel it can "only" go go 48GB which while high compared to desktop boards is low compared to the FBDIMM based intel boards which can apparently go to 128GB and DDR2 based amd opteron boards which can go to 64GB.
note: i'm known as plugwash most places but i screwd up registering that here somehow in the past and now can't register
There is no way that a cpu doing nothing for 5ms would be giving you more performance than one that switched to doing something else. 5ms is a hell of a lot of cpu time, especially if you're swapping several times a second, it could really add up.
If it turns out that the thread it switches to needs to swap too a little later, it should just queue it and switch to yet another thread that isn't waiting on a swap.
Of course, running swap on an SSD wouldn't suffer from the same latencies, it's more like 0.1ms. Swapping at the rate that would kill an SSD would probably not damage a hard-disk (unless it's in a portable device and being moved...) but that level of swapping only happens when your current working set of memory pages doesn't fit in ram, so you have to do swap after swap after swap just to switch to the next program, so your performance would be completely gone anyway. So it's a choice between slowly killing a disk and having a pc that makes a 386 look speedy...
Or, you could just put enough ram in the box in the first place.
The biggest concern with using an SSD is making sure that writes are done in a multiple of the SSD's erase block size, so that you don't make it read the data around where you're writing, erase it all and then write it back again.
There is no way that a cpu doing nothing for 5ms would be giving you more performance than one that switched to doing something else. 5ms is a hell of a lot of cpu time, especially if you're swapping several times a second, it could really add up.
Of course, but I mean the gains are insignificant under the loads theorized here. If you have to swap several times per second, continuously, each fault requires a read and write (or else you'd reach a point where everything you need is in memory, and everything you don't is in swap). The only time this stuff really matters is when swap is part of the active working set.
Of course, running swap on an SSD wouldn't suffer from the same latencies, it's more like 0.1ms.
Throughput is much lower, but this can be addressed. The complexity of the operation is still fairly high; remember the MMU actually performs most of the virtual memory work in hardware, doing it in software much less doing it with block copy operations from slow media presents a huge performance loss.
Swapping at the rate that would kill an SSD would probably not damage a hard-disk (unless it's in a portable device and being moved...)
Seeking back and forth constantly wears on the motor. Believe it or not, hard disks aren't made for constant, continuous use, and will wear down quite quickly if you're continuously streaming data on and off them from random areas of disk.
but that level of swapping only happens when your current working set of memory pages doesn't fit in ram, so you have to do swap after swap after swap just to switch to the next program, so your performance would be completely gone anyway. So it's a choice between slowly killing a disk and having a pc that makes a 386 look speedy...
This is my point. By the time you can damage an SSD or hard disk, your computer is non-responsive anyway. It's time to buy more RAM or give up.
Support my political activism on Patreon.
Swap is a bit of an outdated concept, it's from an age when CPUs were slow, ram was really slow, and really expensive. Swap was required for a machine to be usable.
Now it's more of a hindrance, at least to those of us who don't have an OEM machine with 256MB of ram and 300MB of OEM startup crap. Especially due to that quirk of the Windows API that allows programs to insist on swap and fail to run if you have it turned off. "ABC requires at least 64MB of swap" -> "I have 4GB of ram, and 3GB of that is unused, use that" -> "No, I want 64MB of swap" grrr.
That's a quirk of programmers, as you can make the same check on Linux. It's stupidity.
Support my political activism on Patreon.
Someone once said: "You can't blame the API for the actions of the programmer who uses it" or something like that. I think they're probably right there.
I think we've reached the end of this conversation. Nice meeting you, I'll stick you in my friends so I see you around :)
bash scripting does make 'rm -rf /' perfectly legal...
Support my political activism on Patreon.
Also, bear in mind that the swapping-around is done internally by the disk, and none of that information need ever be sent down the bus. There's no reason the disk couldn't do this asynchronously, without eating up any bandwidth at all.
DRM: Terminator crops for your mind!