Slashdot Mirror
Solid State Drives Tested With TRIM Support
Vigile writes "Despite the rising excitement over SSDs, some of it has been tempered by performance degradation issues. The promised land is supposed to be the mighty TRIM command — a way for the OS to indicate to the SSD a range of blocks that are no longer needed because of deleted files. Apparently Windows 7 will implement TRIM of some kind but for now you can use a proprietary TRIM tool on a few select SSDs using Indilinx controllers. A new article at PC Perspective evaluates performance on a pair of Indilinx drives as well as the TRIM utility and its efficacy."
I finally got the opportunity to test out SSDs this year. There may be the odd teething problem to get over, but in my mind there is no market in the future for mechanical drives except maybe as cheap low-speed devices for storing non-critical information... in much the same way as tape drives were used a few years ago.
What in the world are you talking about? The nice things about SSDs is that yes, they do fail, but they fail (or are supposed to) in a predictable, non-catastrophic way that leaves the data readable just not writable. I have had two SSDs and haven't had either fail despite heavy usage, and I don't think you could patent SSDs because the technology is everywhere because it is flash memory and even if it is patented more companies make them than just one.
Taxation is legalized theft, no more, no less.
This older Slashdot post linked in the story links to a story that covers that topic very well: http://www.pcper.com/article.php?aid=669
NILFS - http://www.linux-mag.com/id/7345/
Even the best consumer-level SSDs like the Intel x-25m/e use a volatile RAM cache to speed up the writes. In fact, with the cache disabled, random write IOPS drops to about 1200, which is only about three or four times as good as a 15k 2.5" drive. The more expensive truly-enterprise SSD drives which don't need a volatile write cache cost at LEAST $20/GB, so the $/(safe random write iop) ratio is actually still pretty close, and cheap SATA drives may actually be even on that metric as the fast enterprise SSDs. Granted, this shouldn't be the case in a year, but that's where it is right now. (Also, the performance-per-slot is a lot higher for SSDs, which can translate into different $ and power and space savings.)
Because, basically, flash drives are laid in levels.
When you delete, you simply map logical space as free.
If you go to use that free space later, you find that area, and drop shit into it. It's I dunno, a 32 KB block of memory called a page. If the page is full (to the point where you can't fit your new shit) of "deleted" files, you first need to write over those deleted files, then write your actual data.
If the logical space is full with good, fragmented (with deleted files interspersed) files, you need to read out to memory, reorder the living data and remove the deleted data, add in the full page back.
Think of it as having a notebook.
You can write to 1 page at a time, only.
Page 1 write
Page 2 write
Page 3 write
Page 2 delete
Page 2 write (still space)
Page 2 write (not enough space, write to page 4 instead)
Page 2 delete
Page 2 write (not enough space, no more blank pages, read page 2 and copy non-deleted shit to scratch paper, add new shit to scratch paper, cover page 2 in white out, copy scratch paper to whited-out page 2)
Disclaimer: I am not a SSD firmware author, although I've spoken to a few.*
As best I can understand it, the problem is that writes are scattered across the physical media by wear-leveling firmware on the disk. In order to do this, the firmware must have a "free list" of sorts that allows it to find an un-worn area for the next write. Of course, this unworn area also needs to not currently be storing any relevant data.
Now, consider a SSD in use. Initially, the whole disk is free, and writes can go anywhere at all. They do, too - you end up with meaningful (at some point) data covering the entirety of the physical memory cells pretty quickly (consider things like logfiles, pagefiles, hibernation data, temporary data, and so forth). Obviously, most of that data doesn't mean anything anymore - to the filesystem, only perhaps 20% of the SSD is actually used, after 6 months. However, the SSD's firmware things that every single part has now been used.
Obviously, the firmware needs to be able to detect when data on disk gets obsoleted, and can safely be deleted. The problems with this are that this leads to *very* complicated translation tables - logical disk blocks end up having no relation at all to physical ones, and the SSD needs to track those mappings. The other problem is that these tables get *huge* - a typical home system might have between 100K and 1M files on it after a few months of usage, but probably generates and deletes many thousands per day (consider web site cookies, for example - each time they get updated, the wear leveling will write that data to a new portion of the physical storage).
Maintaining the tables themselves is possible, and when a logical block gets overwritten to a new physical location, the old location can be freed. The problem is that this freeing comes at the same time that the SSD needs to find a new location to write to, and the only knowledge it has about physical blocks which can safely be overwritten is ones where the logical block has been overwritten already (to a different physical location). Obviously, the lookup into the table of active blocks has to be indexed by logical block, which may make it difficult to locate the oldest "free" physical blocks. This could lead to searches that, even with near-instant IO, result in noticeable slowdowns.
Enter the TRIM command, whereby an OS can tell the SSD that a given range of logical blocks (which haven't been overwritten yet) are now able to be recycled. This command allows the SSD to identify physical blocks which can safely be overwritten, and place them in its physical write queue, before the next write command comes down from the disk controller. It's unlikely to be a magic bullet, but should improve things substantially.
* As stated above, I don't personally write this stuff, so I may be mis-remembering or mis-interpreting. If anybody can explain it better, please do.
There's no place I could be, since I've found Serenity...
That's a statistic that doesn't make any sense.
20% under what conditions, and in what timeframe? Over a long enough time period everything has a 100% failure rate.
Normal hard disks also will eventually fail, due to physical wear.
Also if it lasts long enough, at some point, reliability will stop being important. Even if it still works, very few people will want to use a 100MB hard disk from 15 years ago.
That's my biggest complaint about them, actually -- these "teething problems" people mention are pretty much directly a result of OSes treating SSDs as though they were spinning magnetic disks.
No, the OS should be able to do its own wear leveling. If you need to pretend it's a hard drive, do it in the BIOS and/or the drivers, not in the silicon -- at least that way, you can upgrade it later when things like this come out.
Don't thank God, thank a doctor!
You beat me to it, but in the spirit of adding value, there's a good article here. Another benefit of nilfs2 is that you can easily snapshot and undelete files, giving it a sort of built in "time machine" technology (to use apple's terminology).
I'm just surprised that none of the linux distros are talking about it yet. You would think with the apple and ibm laptops using SSD today that there would be some option somewhere. I think everyone is distracted by btrfs.
If you go to use that free space later, you find that area, and drop shit into it.
Knock it off with all the fancy jargon!
Don't blame me, I voted for Baltar.
For a thorough (RE: long) primer on SSDs and long term performance woes, Anand's overview is a must read.
No sig for you!!
Just a small tangential nitpick: we were already more than a factor of ten past that HDD capacity fifteen years ago. The 1GB barrier was broken very early in the Nineties. I still have an HP 1GB SCSI drive from about '91 or '92, IIRC.
As far as failure rates go, I still have ALL of my disk drives (one or two outright failed) from the 15-20 years, and every single one of them still functions at least nominally. I'm still more trusting of magnetic media than I am either rewritable optical or Flash-based media.
Very interesting, I assumed the problem was similar to fragmentation and wondered why nobody compared it as such.
Now, your explanation makes things much more clearer, the global problem is amplified by the additional problem you described.
Now would implementing the logic to control the SSD entirely at the OS/FS level be much slower than implementing it in silicon in the SSD itself ?
As you said, I now understand that the OS/FS would now have to be aware of the underlying media ;-)
Everything I write is lies, read between the lines.
No way, lets have the firmware do this. The problem with your approach is that the OS wont understand the drive as well as the manufacturer does, so it will always be a sub-optimal solution. Dont tie the hands of the manufacturer to put intelligence in his drives. For instance, the best way to wipe a disk is via an ATA command, and not through multi-passes of wipes. The manufacturer knows where the heads are and how the drive writes. The SSD situation is somewhat similar.
Something as simple as deleting the wrong partition becomes an irreversible operation if you do it using a tool that supports TRIM on TRIM-enabled hardware.
This seems needlessly verbose. Let me shorten it for you:
Deleting a partition should always be considered an irreversible operation.
Hmmm, even shorter:
Don't delete a partition unless you want it to go away forever.
Even if you restore the partition table from a backup, you will likely suffer silent file system corruption, which may even not be apparent until it's too late.
If TRIM support is actually implemented on the device, the device is free to 'lose' data on TRIMmed blocks until they are written at least once.
If I understand you correctly, you are suggesting that a disk partitioning tool will use TRIM to not only wipe the partition table itself, but also nuke the partition data from orbit. And you the point out that it would not be adequate to rewrite just the sectors of the partition table.
If so, then the answer is: you don't just restore the partition table, you restore the whole partition (including data) from backup.
I for one consider much-faster write speeds to be a bigger advantage than possibly being able to reverse a partition deletion.
steveha
lf(1): it's like ls(1) but sorts filenames by extension, tersely
I can buy a terabyte hard drive for around $100. For the same hundred dollars, the best SSD I can find is 32GB. On my computer, Steam's cache folder is bigger than 32GB. My music player has a 120GB drive, my DVR has a 350GB drive, and my backup server has a 1.5TB raid. Just because expensive mobile gadgets use expensive solid-state drives does not mean hard drives are dead, dying, or even decaying.
Just another "DOJ fascist authoritarian totalitarian bootlicker" -- Zeio
Once upon a time, a technical subject on /. gave insightful and informative responses that were modded up. Time changes, I guess.
The "fragmentation" that SSD drive have don't really come from wear leveling, or from having to find some place to write things, but from the following properties:
* Filesystems read and write 4KiB pages.
* SSD can read many time 4KiB pages FAST, can write ONCE 4KiB pages FAST, but can only erase a whole 512KiB blocks SLOWLY.
When the drive is mostly empty, the SSD have no trouble finding blanks area to store the 4KiB write from the OS (he can even cheat with wear leveling to re-locate 4K pages to blank spaces when the OS re-write the same block). After some usage, ALL THE DRIVE HAVE BEEN WRITTEN TO ONCE. From the point of view of the SSD all the disk is full. From the point of view of the filesystem, there is unallocated space (for instance, space occupied for files that have been deleted).
At this point, when the OS send a write command to a specific page, the SSD is forced to to the following:
* read the 512KiB block that contain the page
* erase the block (SLOW)
* modify the page
* write back the 512KiB block
Of course, various kludges/caches are used to limit the issue, but the end result is here: writes are getting slow, and small writes are getting very slow.
The TRIM command is a command that tell the SSD drive that some 4KiB page can be safely erased (because it contains data from a delete file, for instance), and the SSD stores a map of the TRIM status of each page.
Then the SSD can do one of the following two things:
* If all the pages of a block are TRIMed, it can asynchronously erase the block. So, the next 4KiB write can be relocated to that block with free space, and also the 127 next 4KiB writes.
* If a write request come and there is no space to write data to, the drive can READ/ERASE/MODIFY/WRITE the block with most TRIMed space, which will speed up the next few writes.
(of course, you can have more complex algorithms to pre-erase at the cost of additional wear)