Tuesday, September 09, 2008

Now we're beginning to talk

A while ago I wrote that I didn't see SSDs replacing common HDDs for a while. I think this drive, however, shows that the while has elapsed. With a bit more refinement and larger capacities, it seems to me that now there is a more proper replacement for the hard drive.

The next observation I'd make on Intel's design, which uses 10 parallel read/write channels with a recombination buffer, is that with the tiniest of efforts one could make a failure tolerant hardware RAID version of the drive in the same enclosure.

Or, seen from a different point of view, Intel's approach is analogous to using 10 way RAID striping within a single drive. Therefore, perhaps it's just a matter of time before 10 way RAID striping / mirroring is also available... or, maybe by combining something like AMD's Fusion (or an equivalent) with a sufficiently large memory buffer, one could have n-way RAID 5/6 in a (perhaps largish but yet compact) single enclosure.

If one could, in addition, physically separate the board holding the flash memory from the board having the controller, then replacing fried controllers while keeping the data intact would also be possible with ease.

What I have not seen yet is a comparison of reliability under long term heavy load between SSDs and HDDs. In particular, how much data can be written to an HDD before it fails? How does that compare to today's SSDs?

Assume an HDD with 10^6 hours of MTBF. How much data can it write at a conservative average of 50mb/sec? It's a staggering figure, really. At ~176gb an hour, the number is ~176 petabytes (or pebibytes, if you prefer). Note that this is independent of the capacity of the drive, as it is only bound by the MTBF.

Then, assume further that Intel has not yet improved the 10k cycles per memory cell reliability of flash memory. Therefore, a 160gb SSD drive would be able to write no more than about 1.6 petabytes before it fails. Note that in this case this number is bound by the cell reliability and the capacity of the drive, not time per se.

So... one or two orders of magnitude improvement in the reliability of flash memory, and the hard drive becomes completely obsolete. Easier said than done, I am sure. Hopefully soon.

No comments: