Disk drive vendors have been telling us for years that more than half the drives that get returned to them for warranty repairs fall into the category they call NPF--no problem found. As an IT professional, I assumed that the true cause of this phenomenon, like many of the problems that beset the helpdesk, was located between the keyboard and chair. A recent blog post by LSI's Rob Ober shines new light on the subject of false drive failures and has me wondering why this problem persists.
Ober notes that false failures are a major problem not just for hobbyists that buy bare drives from Fry's and Newegg but also for major data center operators. Data center operators, like thee and me, have substantial costs when a drive fails. For instance:
• System performance drops off, often for days, as the RAID system rebuilds as much as 4TB of data onto a hot spare drive. In distributed environments using scale-out storage, this also affects network traffic as the rebuild data has to be consolidated across multiple storage nodes.
• Someone has to go change the drive.
• Because the drive has sensitive corporate data, it has to be sanitized or destroyed. If you're not big enough to have an agreement with your storage vendor to replace failed drives on your say-so, it may mean you also have to pay for the replacement as you can't return drives.
The problem is that today's disk drives are run by internal microcontrollers that have firmware. Just like your PC or Mac, that software occasionally gets confused and the processor freezes. The drive hits a series of requests and states that weren't completely debugged in the development process and its processor stops responding to commands from the host or RAID controller.
If a host or RAID controller reports such a drive as failed, that drive will work just fine when removed from the host and tested elsewhere. (We all know turning off and turning on the power solves a lot of computing problems.) In fact studies have shown that drives that have suffered this kind of false failure are just as reliable, after they get a reset, as new drives fresh from the factory.
Mr. Ober actually got a large data center operator, who remains nameless, to share its drive failure statistics with him. This datacenter, while small by Google or Facebook standards, is pretty huge with over 200,000 servers.
• 30+% of their SAS drive failures are false, adding up to 10-15 a day or a 1/1000 annual false failure rate.
• SATA drives, directly connected to server motherboards, have an even higher false failure rate, approaching the 50% number that drive vendors have long reported, and a frightening 1% annual false failure rate.
A few vendors have tried to address the problem. Five years ago Xiotech and Atrato were talking up "self-healing" disk arrays that would perform repair tasks rather than starting a RAID rebuild immediately when a drive stopped responding to commands. Xiotech, working closely with Seagate, could even keep running a drive with a damaged surface or failed head by mapping accesses around it. Of course the first step in this recovery process was to perform a hard reset on the drive.
[ Join us at Interop Las Vegas for access to 125+ IT sessions and 300+ exhibiting companies. Register today! ]
With the industry turning its fickle attention to flash, self-healing arrays aren't cool anymore. Atrato has gone the way of all flesh and Xiotech, now re-named X-IO has faded in relevance as its last independent competitors Compellent, 3Par and even Nexsan were acquired.
Because the disk drive market is essentially a duopoly selling high-volume, low-margin products, I don't expect Seagate or Western Digital to build a highly redundant circuit board into drives that could detect false failures and reset itself. However there are a few things industry players, including LSI, could do.
SAS controller vendors such as LSI could build false failure detection and reset into the controller. When a drive fails to respond, the controller could give it quick kick before starting a RAID rebuild. This is harder on SATA drives as they lack some of the connections needed, but the folks that control the SATA spec could add a hard reset capability in the 6-12Gbps upgrade that's coming in the next few years. Short of that, array vendors could add the ability to cut the power to individual drives to force the reset.
No matter how you cut it, a 1% AFR is unacceptable. The industry should be working on real solutions, not just faster rebuilds.