Garth Gibson, a professor at Carnegie Mellon University also known for co-authoring the seminal “Berkeley Raid paper,” was the keynote speaker at the Storage Systems, Hard Disk and Solid State Technologies Summit that was held in Singapore on November 1-2.
The name of this conference is rather lengthy but it was chosen to call attention to the need for both hard disk and solid state technologies to address the future requirements for storage. This conference was sponsored by the Data Storage Institute of Singapore and was targeted at storage developers. Hitachi researchers from our Yokohama research lab presented on virtualization and data ingestion, and Hitachi Data Systems was there to demonstrate a Big Data solution for media editing.
Garth opened his keynote with a segment called “Storage Systems Issues for Shingled Magnetic Recording.” In this session he stated that solid state technologies will replace high performance hard disks, and that hard disks will now be able to concentrate on what they do best: provide low cost, high capacity storage.
Although it looks like the hard disk industry will be moving on to Heat Assisted Magnetic Recording (HAMR), hard disk storage bit densities are no longer doubling every two years as they have in the past. Storage vendors like Seagate and Western Digital are looking for ways to increase densities by changing the way they record magnetic bits on hard disks using current magnetic head and disk recording technologies. Shingled Magnetic Recording (SMR) has the potential of increasing track densities by overwriting tracks like shingles or tiles on a roof. Writing a track on a disk takes a lot of energy so the tracks are wide, but reading a track requires less energy so only a portion of the original track is required for reading.
Garth is a proponent of Shingled Magnetic Recording, and he has spoken at many events to educate the industry on the issues with SMR, so that we can be prepared to use this new technology when it becomes available. Following is a summary of the notes I took from his keynote.
SMR tracks are written in bands with guard bands in between the bands. Writing in this manner is projected to increase track densities by 1.5 to 2.5 times. The last track on the band would not be over-written so it could be updated as usual; however, all the over-written bands would require a Read/Modify/Write which could require seconds that would result in timeouts. SMR disks will have the “write once read many” characteristics of tape with the exception that it can be randomly read. Determining the right track requires three revolutions and signal processing in two dimensions.
Garth points out that NAND flash storage uses Read/Modify/Write so this technology already exists and should be transparent to system software. Read/Modify/Writes could also be avoided through the use of log structured files.
The analogy to flash was very interesting and could lead to the same rationale for designing a controller for SMR to optimize performance and density as Hitachi Accelerated Flash storage has done for flash storage.
SMR may well be in our future and we need to start understanding what differences and benefits it could bring to our operations. For more information on SMR, the best place to start is with the Carnegie Mellon technical report on the Principles of Operations for Shingled Disk Devices.