Slashdot Mirror
Ask Slashdot: How Do You Test Storage Media?
First time accepted submitter g7a writes "I've been given the task of testing new hardware for the use in our servers. For memory, I can run it through things such as memtest for a few days to ascertain if there are any issues with the new memory. However, I've hit a bit of a brick wall when it comes to testing hard disks; there seems to be no definitive method for doing so. Aside from the obvious S.M.A.R.T tests ( i.e. long offline ) are there any systems out there for testing hard disks to a similar level to that of memtest? Or any tried and tested methods for testing storage media?"
All I usually do is:
1. smartctl -AH
Get an initial baseline report.
2. mke2fs -c -c
Perform a read/write test on the drive.
3. smartctl -AH
Get a final report to compare to the initial report.
If the drive remains healthy, and error counters aren't incrementing between the smartctl reports, it's good to go.
Can You Say Linux? I Knew That You Could.
S.M.A.R.T is a joke, but not in implementation. It's a joke because most HDD failures occur on the logic board. It's a known fix in data recovery services to simply swap out the PCB for another of the same vintage make/model/firmware rev. Though I have ran tools such as HD Tune to view out-of-spec metrics and benchmarks. For example, I once had a user that reported that her workstation was running extremely slow. I suspected the drive was at fault and the graphs proved it, but technically it wasn't a failure. S.M.A.R.T would have flagged it if it was mechanical, but it wouldn't have if it was a controller issue. Now that may have changed with newer drives, but that's been my overall experience.
Life is not for the lazy.
Not completely related to how to test, but...
In 2007 Google reported that for a sample of 100k drives, only 60% of their drives with failures had ever encountered any SMART errors. Also, NetApp has reported a significant amount of drives with temporary failures, such that they can be placed back into a pool after being taken offline for a period of time and wiped. Google also had a lot of other interesting things to say (such as heat has no noticeable effect on hard drive life under 45C, that load is unrelated to failure rates, and that if a drive doesn't fail after 3 months, it's very unlikely to fail until the 2-3 year timeframe.
You can find the google paper here: http://static.googleusercontent.com/external_content/untrusted_dlcp/research.google.com/en/us/archive/disk_failures.pdf
A few other notes that you can find from storage vendor tech notes if you own their arrays:
* Enterprise-level SAS drives aren't any more reliable than consumer SATA drives
- But they do have considerably different firmwares that assume they will be placed in an array, and thus have a completely different self-healing scheme than consumer-level drives (generally resulting in higher performance in failure scenarios)
* RAID 5 is a really bad idea - correlated failures are much more likely than the math would indicate, especially with the rebuild times involved with today's huge drives
* You have a lot more filesystem options that might not even make sense to use with a RAID system, like ZFS, as well as other mechanisms for distributing your data at a layer higher than the filesystem
Ultimately the reality is that regardless of the testing you put them under, hard drives will fail, and you need to design your production system around this fact. You *should* burn them in with constant read/write cycles for a couple days in order to identify those drives which are essentially DOA, but you shouldn't assume any drive that passes that process won't die tomorrow.
Spinrite is not that meaningful thees days since drives don't give you that low level access to the media like the days of old. Since you can't low level format drives which was one of spinrites strong points, save money and use badblocks. Use badblocks in read/write mode with random test pattern or worst case test pattern a few times. Then do a SMART long self test. Keep an eye on the pending sector count and the reallocated sector count. A drive can remap a difficult sector without you ever knowing unless you look there. Also keep an eye on drive temperature, even a new drive can act flaky if it gets too hot.
>Still works 100% as HDD tech is still the same
Not entirely true. Back in the days of MFM/RLL drives, SpinRite could perform a "low level" format on each track. This ensured every last magnetic 1 and 0 was re-written to the disk. Back in the day I witnessed many times when SpinRite would completely recover bad sectors, presumably damaged by electrical/controller issues rather than physical surface issues, and a full pattern test would prove the space was safe to use.
Modern IDE drives don't allow low-level formatting, and as far as I know, even re-writing the user content of the drive does not re-write sector header data. Modern IDE drives also have hidden reserved space for "spare" tracks and space where they store their firmware, which likewise never gets tested or re-written.
Additionally, on MFM/RLL drives SpinRite could use low-level formatting to optimize the sector interleave for the specific system. (You would be surprised, moving some disks and their ISA controllers to a faster system would actually require a higher interleave, slowing them down incredibly until SpinRite was run)
Still, SpinRite is the only program that I know of that can do a controlled read/write pattern test and modify the underlying file system when needed.
Spinrite may do an OK job of exercising disks, but 90% of what it claims to do is BS.
An easy test to prove that Spinrite is BS is run it against a USB key. Not a SATA SSD, but a USB flash drive. Make the USB key bootable with DOS, put Spinrite on and boot a PC with no other drives. Run its "tests" against the USB key. All the "low level" tests Spinrite claims to do will appear to work, but are impossible on a USB device.
Infact, they are impossible on a modern mechanical HD as well. As yacc143 pointed out, modern drives are not the same as MFM/RLL drives of the past. The low level tests that Spinrite claims to do are simply impossible.
It's also a terrible data recovery program, since it can only write recovered data back to the same disk. That's a data recovery 101 no-no, and Spinrite fails.
Not until the hardware fails and you need the data that was on there but not on the backup (or realized the backup failed a long time ago...).
For performance, yes, hardware is fastest. For reliability though, software RAID is better (hardware RAID can have interesting firmware version issues).
Linux running an md RAID array? If the server goes down, pop the drives in another server, a couple of mdadm commands later and the array is up and running. Hell, even Windows' software RAID ought to be able to work to recover an array where the server hardware died.
So if you're using RAID not for performance reasons, but for protection against hard drive failure, soft-RAID works very well. Hell, one of my NAS appliances died, and all I did was take the drive out, attach 4 USB adapters to them, and plug them into my Linux box. Instant access to the data,
There's nothing like the panic that happens when an array goes down due to non-drive hardware failure.
Speaking as somebody that has done hardware qualifications and burn-in development at very large scale for companies you ahve heard of let me tell you the tools I use:
fio: The _BEST_ tool for raw drive performance and burnin testing. A couple of hours of random access will ensure the drive head can dance, then a full block by block walk through with checksum verification will ensure that all blocks are readable and writable.. I usually do 2 or 3 passes here. You can tell fio to reject drives that do not perform to a minimum standard. Very useful for finding functional yet not quite up to speed drives. The statistics produced here are awesome as well.. Something like 70 stats per device per test.
stressapptest: This is google's burn in tool and virtually the only one I have ever found that supports NUMA on modern dual socket machines. This is IMPORTANT as its easy to ignore issues that come up with the link between the CPUs. The various testing modes give you the ability to tear the machine to pieces which is awesome. Stressapptest also is the most power hungry test I have ever seen, including the intel Power testing suite that you have to jump through hoops to get.
Pair this with a pass of memtest and you get a really, really nice burn in system that can burtalize the hardware and give you scriptable systems for detecting failure.
Except when it happens a lot... then your drive is F***.............I.....N...............G slow for now apparent reason.
We "test" our drives by filling them with whatever data we have laying around. We do this 5 to 10 times (depending on how soon we need the drives). What eventually happens with a bad drive is that the SMART counter ticks over to some magical number and starts reporting health issues (A requirement for some RMA processes). We also time each fill cycle. We expect the first two or three runs to take longer (EVERY drive these days will have relocations going on for the first few runs). For later runs we expect to see a more consistent fill time and the relocated sector count stop climbing so alarmingly fast.
There are bad sectors on your brand new drive. You can count on it. You have to make the drive find them and map around them because it won't happen in the factory. Write to every byte several times. Don't wait for it to happen naturally... you'll just hit performance problems and put yourself closer to warranty cutoff time. They're counting on you not finding a problem soon enough. You must burn them in or suffer later.
I want this account deleted.
To a degree, you can rule with certainty that everything is working.
New equipment does tend to have ghosts. Given enough systems, with homogeneous roles, it doesn't matter: if it starts to fail, you pull it and put another one in.
If you've got an environment with only a few servers with dedicated roles, having a new 'production server' go tits up is a very bad thing. For a system like this, you really do want to do a 'burn in' period, IMO for at least a couple weeks, where the system is not being depended upon. Your 4-year-old system doing the same thing at relatively diminished capability is not nearly as bad as doing a cut-over and having things go south, then.
You do, however, want to do a "burn in" on that new equipment. My preference is to stress a new piece of equipment with something like building kernels (which will stress every significant subsystem to some degree) while doing file operations (eg. something like bonnie+ if you're not copying files to the machine) for a period of at least a week without any stability or significant performance problems. This is due to the following subjective observations:
* getting a system with a defective disk is not uncommon these days. It's not common, so it's not a serious concern.
* Short of initial failure of the disk/DOA status, the disks will likely run a number of months before your first failure (depending on how many you've got, of course)
* Instability, inconsistent behavior, flaky RAM, or odd behavior from RAID or NIC controllers, and 'ghosts' can almost invariably be traced back to the PDU or PSU. These seem to die within about two weeks to a month if they're defective/poorly designed. With a server, troubleshooting this can be a huge bitch due to how loud they are and the multiple-dependence issue on the PDU. This is kind of an end game for me, and I have a hard time trusting any of the equipment after I've had a PSU fail.
* if you plan on taxing the system at all, you'll probably have a driver related performance problem somewhere down the line. Better to find it before you need the performance.
* Every once in a while, you've got a bad solid state device (RAM, CPU, SSD). These seem to either work, or not work, if they pass initial "does it work?"
~/ssh slashdot.org ssh: connect to host slashdot.org port 22: too many beers