“Spinning rust?” you may say? Yes, rust, or another iron oxide, is the ferromagnetic coating on the platters spinning inside every hard disk drive. To read and write your data an actuator arm, driven by a separate motor moves a tiny recording head over each platter surface, flying on a microscopic layer of air to change the magnetic polarization one bit at a time. Advances in disk technology have been impressive and the rate of capacity growth has been amazing. As you might expect though, the rate of performance improvement in this 50-year old technology has been slower than any other component in your computer, laptop, workstation or server.
Measurements presented by Intel at Intel Developer Forum 2011 suggest that since January 1996, processing capability has improved 175x. In that time, RAM speed and bus design have enabled a 40x improvement and that is further enhanced by the significant advances in layer 1,2, and 3 caching on the processor. I/O bus speeds have improved from PCI at 133MB/s to a standard server PCI Express 2.0 x8 link that runs at 4GB/s, a 24x improvement and the PCIe bandwidth doesn’t get shared between slots as it was with PCI. Even disk interface speeds have improved with current SATA interfaces running 60x the speed of early ATA connections. But disk performance? Intel suggests a 1.3x improvement since 1996. At the system level, let’s say there has been a 2x improvement since 1996 based on the doubling of rotational speed, actuator speed improving and areal density doubling several times.
Certainly, disk capacity has grown more than 500x since 1996. And transfer rates for long sequential access have improved significantly, perhaps as much as 10x. IOPS is the metric that has just not improved. IOPS, Input/output Operations Per Second, is the measure of how many distinct reads or writes a drive or storage system can do every second. When it comes to random Input/Output operations, disk drive progress has been almost un-measurable compared to everything else in the system because of the mechanical limitations inherent in disk drive technology.
Are modern RAID controllers ready to handle the performance boost offered by SSD drives, or will that become the next bottleneck?
Mike,
Most modern RAID controllers have a lot of power, but most of the optimization that has been done over the last ten or fifteen years has been aimed at minimizing or hiding the mechanical latency in arrays of spinning disks. The LSI and Intel-branded MegaRAID controllers that ION’s SR-71mach4 SpeedServer™ and SR-71mach5 SpeedServer™ are designed around have been optimized for arrays of SSDs. RAID 5 SSD performance on ION’s SR-71mach4 Speedserver has been documented at up to 1.4 million random read IOPS for 4kB transactions. Doing a little arithmetic, 1.445M IOPS from (3) arrays of (8) SSDs each, or 1.445M divided by (24) SSDs is a net performance of about 60,000 IOPS per SSD. That is just about all of the small random IOPS that one current SSD has to offer.
Note that this performance is achieved with (3) RAID controllers, each of which is managing one 8-SSD RAID 5 array. With (24) SSDs on a single controller, the performance of the controller would limit the total that can be achieved, but the latest controllers are able to deliver about half of that, which is still quite impressive.
Thanks for your input! Keep on reading and feel free to share more thoughts and questions.
Keith