Hybrid Storage Pools with Hitachi Flash Modules and Hard Disk Drives
by Hu Yoshida on Feb 15, 2013
The history of magnetic recording has been one of the most amazing technology achievements in the last 50 years. In 1956 IBM announced the first commercial hard disk drive (HDD) with 2000 bit/in2. Last year, Seagate announced 1 TB/in2 with their new Heat Assisted Magnetic Recording technology (HAMR). This represents a more than 109 improvement in bit density! This has also driven down the price per bit by a factor of more than 300 million since the introduction of the RAMAC!
While there have been tremendous achievements in capacity and cost of HDDs, one area where hard disks have not kept up the pace with other technologies is in performance. HDDs are mechanical, rotating devices, so there is a latency that is incurred in accessing a block on a track. There is also a delay for the read/write heads to seek to a specified track. The history of HDD development has been trying to balance performance, capacity, and cost, but performance has always come out on the short end.
Latency depends on the disk rotation speed and has been limited to 15K RPM for the past 12 years. Improvements in latency are hard to achieve and there is little hope of increasing RPM in the future. The latency time is random and half the rotation time is usually a good estimate of the average latency. Latency can be reduced to a degree by wide striping the data across many drives in order to use more disk arms to access the data.
Seek time will vary depending on where the read/write head is initially positioned and how far it needs to move to find the next targeted track. In enterprise disks this varies from less than 1 ms for a displacement of one track to 4 ms for an average seek across half the number of tracks on a disk surface. The ability to accelerate the arm from zero and stop it on a specific track in a matter of milliseconds is extremely challenging. This seek time can be minimized by short stroking or using a small band of the available tracks, and leaving the rest of the tracks unused. Because seeks are important for random reads, many performance applications use short stroking on expensive 15K RPM disk drives with wide striping across many disk drives. This is expensive but the cost is justified for the performance gains it provides with disk drives.
Solid state disks (SSD) are not dependent on mechanical access and their I/O performance is orders of magnitude better than hard disks – on average. SSDs can only write to formatted pages and so a read/modify/write is required to update pages and when they run out of formatted pages the I/O is blocked while housekeeping is done to recover and format blocks of pages. During housekeeping, the I/O performance of SSDs is about the same as hard disks. This is known as the SSD “write cliff”. Because of this overhead, SSDs have been limited in capacity to 400GB and are nearly 10 times the price of high performance hard disks.
RAID rebuild times for SSDs are about the same as for HDDs and because SSDs are less durable than HDDs, you can expect to be replacing SSDs more often than HDDs. While you may have a 3 or 5 year time period where failed SSDs are replaced under warranty, the frequency of SSD replacements and subsequent hours of RAID rebuild time will be disruptive to your operations, especially to applications that require consistent high performance and availability.
The solution developed by Hitachi is to provide a hybrid combination of purpose-built flash controller for Hitachi Accelerated Flash (HAF) modules that reside in the same pool of storage with HDDs. This way the pages of a volume that are assigned to this pool will receive flash performance when the data is “hot” and automatically move to a larger capacity HDD when it becomes less active. This enables flash to do what it does best – high performance I/O, and relieves the HDD of the burden of performance so that it can concentrate on what disks do best – reliable, high-capacity storage.
HAF is not an SSD. The controller for HAF is designed to remove the housekeeping from the I/O path; it eliminates the “write cliff” and provides consistent performance. It also provides extended ECC to reduce the refresh rate, it buffers, compresses and eliminates writes, it increases spares, and it provides global wear leveling to extend the durability of flash. The HAF controller enables an increase in capacity of a flash module to 1.6TB or 4 times the capacity of a 400GB SSD. This will reduce the price by 40% compared to similar MLC SSD capacities.
Today, HAF is roughly 4 times the cost of high performance 15K RPM HDDs, but considerably less expensive than HDDs on a usable capacity basis, when HDDs are configured for performance. A high-performance configuration of HDDs would require RAID 10 to reduce the need for parity generation. Then the LUNs would be wide striped across many spindles, and restricted to a small band of track capacity on each spindle to reduce latency and seek time. Many more spindles of HDDs would be required than HAF modules for equivalent performance, which means that the probability of an HDD failure in an HDD configuration would be higher than a failure of a HAF module in a HAF configuration.
This hybrid combination of HAF and large capacity HDDs, tiered together in the same dynamic pool of storage in Hitachi VSP, provides the benefits of higher performance, higher availability, and reliability at a lower cost than comparable high performance disk and/or SSD configurations.