tents very quickly, useful for scrubbing, as well as answering queries.

This is all encouraging, but once it became possible to study the behavior of disk storage at a large scale, it became clear that system-level reliability fell far short of the media specifications. This should make us cautious about predicting a revolution from flash or any other new storage technology.

economics

Ever since Clayton Christensen published The Innovator’s Dilemma10 it has been common knowledge that disk-drive cost per byte halves every two years. So you might argue that you don’t need to know how many copies you need to keep your data safe for the long term, you just need to know how many you need to keep it safe for the next few years. After that, you can keep more copies.

In fact, what has happened is the capacity at constant cost has been doubling every two years, which is not quite the same thing. As long as this exponential grows faster than you generate new data, adding copies through time is a feasible strategy.

Alas, exponential curves can be deceiving. Moore’s Law has continued to deliver smaller and smaller transistors. A few years ago, however, it effectively ceased delivering faster and faster CPU clock rates. It turned out that, from a business perspective, there were more important things to spend the extra transistors on than making a single CPU faster. Like putting multiple CPUs on a chip.

At a recent Library of Congress meeting, Dave Anderson of Seagate warned4 that something similar is about to happen to hard disks. Technologies—HAMR (heat-assisted magnetic recording) and BPM (bit pattern media)—are in place to deliver the

table 1. the time to read an entire disk of various generations.

1990

2000

2006

2009

2013

240

720

6450 8000 12800

Disks have been
getting bigger but
they have not been
getting equivalently
faster. this is to be
expected; the data
rate depends on
the inverse of the
diameter of a bit,
but the capacity
depends on the
inverse of the area
of a bit.

2013 disk generation (that is, a consumer 3.5-inch drive holding 8TB). But the business case for building it is weak. The cost of the transition to BPM in particular is daunting. 24 Laptops, netbooks, and now tablets are destroying the market for the desktop boxes that 3.5-inch drives go into. And very few consumers fill up the 2009 2TB disk generation, so what value does having an 8TB drive add? Let alone the problem of how to back up an 8TB drive on your desk!

What is likely to happen—indeed, is already happening—is that the consumer market will transition rather quickly to 2.5-inch drives. This will eliminate the high-capacity $100 3.5- inch drive, since it will no longer be produced in consumer quantities. Consumers will still buy $100 drives, but they will be 2. 5 inches and have perhaps one-third the capacity. For a while the $/byte curve will at best flatten, and more likely go up. The problem this poses is that large-scale disk farms are currently built from consumer 3.5-inch drives. The existing players in the market have bet heavily on the exponential cost decrease continuing; if they’re wrong, it will be disruptive.

the Bigger Picture

Our inability to compute how many backup copies we need to achieve a reliability target is something we are just going to have to live with. In practice, we are not going to have enough backup copies, and stuff will get lost or damaged. This should not be a surprise, but somehow it is. The fact that bits can be copied correctly leads to an expectation that they always will be copied correctly, and then to an expectation that digital storage will be reliable. There is an odd cognitive dissonance between this and people’s actual experience of digital storage, which is that loss and damage are routine occurrences. 22

The fact that storage is not reliable enough to allow us to ignore the problem of failures is just one aspect of a much bigger problem looming over computing as it continues to scale up. Current long-running petascale high-performance computer applications require complex and expensive checkpoint and restart schemes, because the probability of a failure during execution is so high that restarting from