While capacities, fill times, and bandwidth have changed, latencies have, however, remained relatively static. Hard disks still use rotating platters with magnetic pits, and spin latencies can only decrease with increases in rotational speed. Spindle speeds have risen, from the 3500RPM Maxtor of yesteryear, through 4200, 5400, and today’s 7200RPM spindle speeds, with the transition to 10,000 and 15,000 RPM speeds accomplished in some sectors of the server space and burgeoning in the desktop space. But this reliance on physical moving parts has made their development hard; this is only a doubling of spindle and a latency cut from about 30 milliseconds to about 10 milliseconds. While main memory latencies have fallen by a factor of ten, and CPU speeds have risen by factors of thousands, hard disk latencies have remained comparatively static in the milliseconds for those same 17 years.
The result of all of these trends is that it’s easier and easier to store huge amounts of data at lower and lower cost, but the increasing data bandwidth and low latency that modern storage needs demand is harder to come by. I’ve heard from a source that a copy of the text of the entire Internet, stored by Google and its competitors and searched for text in web search queries, is in the vicinity of 20TB. In hard disks, this would now cost a mere $4,000 in disks (much more in servers) but would be completely unsuited to this kind of storage due to the long read time of the disks; a total read for a search would take hours. Instead, Google hosts its web servers from huge clusters of thousands of servers, storing complete copies of the text of the internet in RAM, at about half a million dollars in DDR2 per cluster. (ArsTechnica)
1 Response to “SSD ready to usher old hard drives into the sunset”