SSD Capacitors?

[km23]

Hi.. I read in the Intel 520 review that the SSD does not have low power capacitors so in the event of a power outage data is completely lost. Can someone explain this to me? Which SSDs do have capacitors and which don't?

I have a Mac Pro and reliability is more important than speed. Any suggestions for an SSD that has reliability? Thanks!

 

[Coup27]

I think the only SSD with capacitors in the consumer space is the Intel 320.

It's worth noting that the only data which is lost is the data which has not yet been wrote to NAND. When you write data, the controller stores the data in a buffer before deciding where to place it. It is this data which would be lost so the chances are small.

Intel do not actually store user data in its buffer but the presence of capacitors means it should be fail safe in a power cut.

 

[km23]

Thanks! Getting a hard drive is also getting complicated.

Any specific recommendations for a Mac Pro? Reliability being the most important thing. Thanks.

 

[Coup27]

I'm not a mac user so cannot comment specifically. I did read that Anand had been using a Samsung 830 in his mac without issue. Samsung has the best reliability record in the SSD industry but does store user data in its buffer and does not have capacitors.

You might be better off hitting this from the reverse angle. UPS's are cheap and small now so if you bought a small UPS you could then buy any SSD you wanted without fear of data loss.

 

[Yellowbeard]

A "super cap" as it is sometimes called is (almost???) exclusively an enterprise feature. I'm not aware of any consumer drives that include an actual super cap but then, I don't study it really. AFAIK, you'll need to buy an enterprise level drive to get this.

 

[gramboh]

How does this compare to hard drives? Wouldn't anything in the hard drives DRAM cache be lost in the event of a power outage as well? I assume hard drives don't cache writes, and only reads.

 

[Mark R]

How does this compare to hard drives? Wouldn't anything in the hard drives DRAM cache be lost in the event of a power outage as well? I assume hard drives don't cache writes, and only reads.

No. Hard drives do cache writes - but modern OSs are specially designed to work with write caches.

Let's take the example of saving a file called widget.dat. A modern OS works something like this:

1. Load directory and free space info
2. Store file data into unused areas of the drive
3. Write the following to a journal file "About to add the file widget.dat to the master directory list, the data is at sector 102345"
4. Modify the master directory to include the "widget.dat" file.
5. Write to the journal file "Master directory successfully updated with widget.dat entry"

As long as the drive guarantees that these steps occur in order, then if the comp crashes or loses power at any point - there are only 1 of 2 outcomes.
If the power is lost before step 3 completes, then the comp will reboot with no trace of widget.dat.
If the power is lost after after step 5, then widget.dat is already completely safe, and there is no corruption.
If the power is lost at step 4, then the OS during reboot sees an incomplete journal entry, and then is able to make the step 4 and 5 changes, leaving widget.dat fully savedand uncorrupted.

So, even if a drive supports write caching (which may allow the drive to reorder these steps), then the OS, will stall at critical steps (between 3 and 4, and between 4 and 5) and automatically send a "flush cache" command to the drive, forcing it to save everything from DRAM onto disk. Only once the OS gets the OK from the drive that the cache has been flushed, does it continue with the next critical step.

The write cache will accelerate "non-critical" writes, but critical writes will still require a slow flush. In order to improve performance, the most modern OSs and drives support "write barriers" which are kind of like a selective flush. The OS can select certain writes that must be completed in a critical order, and ensure that only they get flushed manually - avoiding the performance hit, of occasionally having to flush the whole cache.

The advantage of a super-cap is that the drive can safely ignore the "flush cache" command. This command is quite slow and will stall the drive and OS, especially if the cache is full, and many MB of data must be dumped to the platters or flash memory. A RAID card with capacitor, or SSD with capacitor, can use the capacitor to keep the data safe, so when a "flush" command is received, all that happens is that the drive/RAID card instantly sends back a "OK. Data saved" reponse. As a result, the OS won't stall, even during critical writes.

For certain things, like big corporate databases (where every time a record is updated in a database, this whole complex "journal, wait for cache flus, update, wait for cache flush, update journal" process is needed), the OS can spend a long time waiting for cache flushes. It is here that capacitor supported RAID cards or SSDs are a huge benefit.

 

[NP Complete]

I used to work with a guy on the Intel SSD team a few years back (in fact, I believe he may have worked on adding the capacitor to one of their models). Since the conversation was a few years ago, some details may be incorrect, but here's what I remember:

Almost all hard drives (SSD or Plater-based) use a buffer (DRAM) to store data. This almost always includes data cached to boost read & write performance, but may also include sector mapping tables.

Power failure means that any data not written from the buffer to the drive is lost. This varies from "bad" where a file may be corrupted (due to write data not being written back), to catastrophic (if the SSD doesn't have a consistent sector mapping table, the drive is bricked since often one of the mappings is where the SSD's firmware is stored on drive).

The platter based hard drive designed found a pretty clever trick to allow them to write the buffer in a power failure - they siphon energy from the spinning platter and write the buffer back before they lose all power. Additionally, mechanical hard drives were often more robust due to their design history - the firmware for the drive was often located on a EEPROM or flash chip that wouldn't wasn't ever relocated, so there was only a chance of corruption during firmware updates.

SSDs aren't as lucky - original designs didn't have a capcitor, and design decisions were made to store the firmware on the NAND chips used for the users data to save on cost. The two solutions to the problem were to either add a capacitor that would store enough power to allow the drive to write the buffer on power loss (what intel did), or to engineer firmware in such a way that the mapping table was always consistent. The second approach still results in losing the data in the buffer in a power loss, but at least you don't brick your hard drive.

That said, power failure is a somewhat rare occurence, and the chance of a power failure while critical data is in the buffer is much smaller. Most drive reliability comes down to design and QA effort, and can only realistically be determined through relying on third-party testing and reviews. To that end, both Intel and Samsung seem to be leading in reliability, according to the reviews I can find.

TL&DR: Intel and Samsung are leading brands for realiability currently. I prefer Intel, but that's because I used to work there.

 

[km23]

Interesting information, thanks. I just had a power outage, which while very rare, did happen only a few days ago which has me concerned.

To feel safe, I think a UPC is a must. It looks like either Samsung or Intel. If anyone has experience with OWC, please let me know.

 

[Red Squirrel]

I would just worry about making sure to have a decent UPS and have it setup so when the battery gets low it does a proper shut down. Even a standard spindle drive is subject to issues upon a bad shutdown. Even if the hardware takes care of it, the OS wont have a chance to because it does not know a power hit is coming, so there's always the chance of logical corruption on the file system or other data that may have been in the middle of being written.

 

[taltamir]

while initially enterprise SSDs had super caps, nowadays they use an array of regular capacitors. This is both cheaper AND more reliable.

I read in the Intel 520 review that the SSD does not have low power capacitors so in the event of a power outage data is completely lost.

The same is true for spindle HDDs. And HDDs take more power so putting enough capacitors on them is too expensive.

As people said, its only data which the drive is currently writing, and shouldn't be an issue unless you are running a bank's database or some such on it. Large transactions will be truncated anyways even with capacitors. (ex: if you are copying a file whose size is measured in megabytes rather then bytes)
The solution is to prevent sudden power loss, which is achieved via UPS battery.

The platter based hard drive designed found a pretty clever trick to allow them to write the buffer in a power failure - they siphon energy from the spinning platter and write the buffer back before they lose all power

Are you sure? How common is this (aka, only enterprise drives or all modern drives?)

 

[NP Complete]

I heard the spindle piece of information from the same guy who worked on SSD team at Intel - at the time Intel was beginning a collaboration with another Hard Drive (can't remember witch, perhaps seagate?) manufacturer, so he had heard from some of their senior engineers that this was how mechanical hard drives dealt with abrupt power loss. We used to carpool to work together and I learned all sorts of interesting SSD trivia

Definitely a bit old Xth hand information, and it may not apply to all drives. Still, makes sense - a generator is really just an electric motor in reverse.

 

[hal2kilo]

I heard the spindle piece of information from the same guy who worked on SSD team at Intel - at the time Intel was beginning a collaboration with another Hard Drive (can't remember witch, perhaps seagate?) manufacturer, so he had heard from some of their senior engineers that this was how mechanical hard drives dealt with abrupt power loss. We used to carpool to work together and I learned all sorts of interesting SSD trivia

Definitely a bit old Xth hand information, and it may not apply to all drives. Still, makes sense - a generator is really just an electric motor in reverse.

I actually wondered about the idea of using the energy stored in the spinning platter to save the data on hard drives when this post first appeared. You're saying that in fact the current hard drives actually do that. Interesting.

 

[Mark R]

I actually wondered about the idea of using the energy stored in the spinning platter to save the data on hard drives when this post first appeared. You're saying that in fact the current hard drives actually do that. Interesting.

They don't really do it to save data in the buffer, there isn't enough energy stored for that. There are very close tolerances on spin-speed, which don't allow for much deceleration before the aerodynamics of the heads, etc. cause incorrect flying height, etc. Writing demands very high precision head flight, as a failed write will leave an unreadable corrupted sector (a bad sector as far as the OS is concerned).

However, it is very important that a sector always be written completely. If only part of a sector is written, it will be unreadable, as the checksums won't match. This will therefore be reported as a bad sector to the OS, and the drive will perform time-consuming "error recovery" on the sector, trying to recover the data (which is impossible) causing the drive to drop out of RAID arrays, or cause terrible performance problems.

Because mis-writing a sector is such a big deal, drives do use this flywheel energy to guarantee that they always finish a sector before shutting down completely. Possibly, they might finish the continuous write they work working on (as long as it's not too large, a few hundre k), but the important thing is that no sector gets cut off.

The drive can't guarantee to preserve the data in the write buffer, and OSs are designed to protect critical system data (and databases and similar apps are designed to protect business data) in the event that the power goes off while saving data to a drive, even if that data was allowed to be write-back cached.

 

[NP Complete]

I actually wondered about the idea of using the energy stored in the spinning platter to save the data on hard drives when this post first appeared. You're
saying that in fact the current hard drives actually do that. Interesting.

I probably was a bit too vague in my replay - mechanical drives likely do use the stored energy from the spindle to complete what the drive would consider as critical writes. These are likely to include sector mapping(s), and possibly any writes "in progress." As Mark R states, the magnetic layout on a drive's platter is critical to guiding the heads for correct operation.

Most drives don't care about preserving all the data in their buffer - the drive is only concerned about making sure that it can still operate on the next boot up. It seems unlikely that there's enough energy stored in the spindle to write back the full buffer on modern disks (32-64 MB is a lot of data to write). The drives responsibility is to make sure it works (as in it can perform read/write operations) on the next power cycle, and counts on the OS to make sure critical data is written in a safe fashion, and also counts on the OS to detect and repair incomplete writes.

Some enterprise disks and RAID controllers take this a step further, and add large capacitors and/or batteries to guarentee all data in their buffer is written.

 

[taltamir]

They don't really do it to save data in the buffer, there isn't enough energy stored for that. There are very close tolerances on spin-speed, which don't allow for much deceleration before the aerodynamics of the heads, etc. cause incorrect flying height, etc. Writing demands very high precision head flight, as a failed write will leave an unreadable corrupted sector (a bad sector as far as the OS is concerned).

However, it is very important that a sector always be written completely. If only part of a sector is written, it will be unreadable, as the checksums won't match. This will therefore be reported as a bad sector to the OS, and the drive will perform time-consuming "error recovery" on the sector, trying to recover the data (which is impossible) causing the drive to drop out of RAID arrays, or cause terrible performance problems.

Because mis-writing a sector is such a big deal, drives do use this flywheel energy to guarantee that they always finish a sector before shutting down completely. Possibly, they might finish the continuous write they work working on (as long as it's not too large, a few hundre k), but the important thing is that no sector gets cut off.

The drive can't guarantee to preserve the data in the write buffer, and OSs are designed to protect critical system data (and databases and similar apps are designed to protect business data) in the event that the power goes off while saving data to a drive, even if that data was allowed to be write-back cached.

that makes a lot of sense.
And with 64MB caches it then stands to reason you will have dataloss on powerloss.

 

[hal2kilo]

I probably was a bit too vague in my replay - mechanical drives likely do use the stored energy from the spindle to complete what the drive would consider as critical writes. These are likely to include sector mapping(s), and possibly any writes "in progress." As Mark R states, the magnetic layout on a drive's platter is critical to guiding the heads for correct operation.

Most drives don't care about preserving all the data in their buffer - the drive is only concerned about making sure that it can still operate on the next boot up. It seems unlikely that there's enough energy stored in the spindle to write back the full buffer on modern disks (32-64 MB is a lot of data to write). The drives responsibility is to make sure it works (as in it can perform read/write operations) on the next power cycle, and counts on the OS to make sure critical data is written in a safe fashion, and also counts on the OS to detect and repair incomplete writes.

Some enterprise disks and RAID controllers take this a step further, and add large capacitors and/or batteries to guarentee all data in their buffer is written.

That makes sense. Just completing the sector being written to avoid curruption, but the data in the buffer still goes in the bit bucket.

 

[magreen]

very informative thread. My thought was also that hard drives don't protect you from power loss on the info it's currently writing.

It seems the newness of SSDs is causing people to be cautious and think up what potential problems they could have--not realizing that existing tech. often has those same problems. (I don't intend this as a criticism, OP. It's actually quite reasonable.)

 

[God Mode]

Are these capacitors connected to the main power supply or is it isolated between the controller chips?

 

[NP Complete]

Once again, Xth hand information with bits added from my own knowledge of EE:

The capacitor(s) on SSDs are connected inline with the drives power supply. There's some circuitry connected before the cap to detect power loss, and the cap provides power to the rest of the drive to allow it to go through its "emergency" power down sequence.

 

[magreen]

Once again, Xth hand information with bits added from my own knowledge of EE:

The capacitor(s) on SSDs are connected inline with the drives power supply. There's some circuitry connected before the cap to detect power loss, and the cap provides power to the rest of the drive to allow it to go through its "emergency" power down sequence.

sounds right to me.

 

[hal2kilo]

sounds right to me.

I support a system that uses a more brute force method. It's acutally a PC board in a VME chassis that's used for communicating with 4 more VME units on a fiber optic ring. The PC uses a SCSI hard drive for storage. The power supply in the VME chassis has a big LC circuit in it and as soon as the power supply detects failure of the AC input it sets an interrupt to the PC so it goes into shutdown routine saving all the registers and finishing any writing to the hardrive before the LC circuit drains down.