Slashdot Mirror
Google Proposes New Hard Drive Format For Data Centers (thestack.com)
An anonymous reader writes: In a new research paper the VP of Infrastructure at Google argues for hard drive manufacturers and data center provisioners to consider revisions to the current 3.5" form-factor in favour of taller, multi-platter form factors — with the possibility of combining the new format with HDDs of smaller circumference which hold less data but have better seek times. Eric Brewer, also a professor at UC Berkeley, writes "The current 3.5" HDD geometry was adopted for historic reasons – its size inherited from the PC floppy disk. An alternative form factor should yield a better TCO overall. Changing the form factor is a long term process that requires a broad discussion, but we believe it should be considered."
Nope. SSD's are dead.
A Mechanical drive will last 7 years, a SSD used in the same machine will last 2 years, 3 tops. You pick SSD's for low-write high-bandwidth applications, eg streaming video and images. Mechanical drives are for low-bandwidth high-write applications like logging. So things like Google where Youtube is going to file a lot of content into cold-storage, SSD's are a waste, and in fact it may be more practical to just burn discs for content that nobody is watching.
In the case of what Google is proposing, what makes practical sense is to instead of having 5.25" and 3.25" platters, mount 4 2.5" drives into the same space that a single 5.25" drive space takes, perhaps make a new transverse 5.25" mount, so you you fit a 2.5" height with the platters to the depth of the 5.25" bay. So you'd get something that resembles "blade" installations but one PCB instead of 4 separate PCB's.
However the one omission in this proposal of making the drives taller to have more platters in them is that yes theoretically you get twice the capacity if you double the platters (eg 2 to 4) and twice the data bandwidth IOPS, but you also are still pushing this over the same SAS/SATA connection, which means you approach a SSD's lower-end performance.
As for SSD's, they are terrible in a data center environment. Try copying a large file from one SSD to a mechanical drive or to another SSD of weaker performance, the system will, ALWAYS lockup, regardless of the drive because the OS's disk cache will exhaust the ram on the system in doing so. Or at least this is what always happens on Linux. I don't have any SSD's on BSD machines for comparison, but I have a dozen Linux systems running SSD's, and they are the worst choice for servers. You are better off maxing out the RAM in the system and then running various memory caches (eg memcache, varnish) with a mechanical drive than you are trying to use a SSD, because all the performance is at the memory cache. Instead of trying to save money with conservative amounts of RAM, and expensive SSD's, buy cheaper SATA mechanical drives and push 1/4 of the RAM on the system into software caches (not the disk cache itself.)
Don't get me started on how awful cloud services are that push SSD's. The "virtual machine" server is the worst thing to ever take root in a data center. Now instead of buying 10 fast systems with the right CPU, RAM and Disk configuration, you end up provisioning those 10 servers with 80 VM's, one per CPU core and 1GB of ram, and instead under-size 80 virtual machines with the naive assumption that unused performance/ram will be given to the machine that needs it. The reality is, even at Amazon Web Services, they provision the machine so you have 100% of that division of the machine, and any unused performance is just gone and wasted instead. Over provisioning results in poor performance, and that is what the bean counters do not understand. Once you throw SSD's into the mix, you end up wearing out the drives much faster as machines are spun up and down on the same physical storage. So now instead of having SSD's that last 3 years, you get SSD's that last 90 days.