Flash Drives Replace Disks at Amazon, Facebook, Dropbox
wired.com
wired.com
Basically for many workloads, it's now (edited: original said "not" -- oops) possible to fit entire data sets in RAM, backed by cheaper traditional disks that essentially just stream write logs.
And for workloads where storage costs actually dominate (I'm thinking youtube/dropbox type stuff) it seems like access is mostly streaming, where HD performance is actually quite good.
Obviously SSDs make a ton of sense in the consumer space, where you only get one drive. But if you could pick your RAM/Flash/Disk mix yourself and tailor it to your app, I guess I don't see a lot of advantage to putting your money into the "middle" of that stack.
Since they have so much lower power demands I wonder if they are going to move further up the pipe? Instead of the ISP routing a request to an index server which then decides which is the nearest DC or akaimai host with the actual data - would the first line of dropbox/amazon/google/netflix actually move into the ISP's local rack.
SSD and DMRD win when the query volume is high compared to the data volume since with HDD you'll need to build multiple copies of the system (or equivalent) to handle the load, whereas SSD and DMRD serve more queries with less hardware.
On the other hand, in cases where query volume is low, HDD is cheaper. But if time-to-market matters, you really do want to reduce the turnaround time for your batch jobs and exploratory queries because in the development of KB's with a subjective element you'll probably need to throw 10-100 things at the wall to get the one that sticks.
Also, I'm really surprised there's a datacenter that has a meaningful amount of Facebook, Mozilla, Dropbox and Google servers in the same DC. Looks like someone might be trying to bolster their DC image a bit.
Also: for app sever storage (and temporary database tables), SSD makes a lot of sense, and contemporary needs vs. available size (100-200 GB is plenty for many system installs, if not a factor-of-ten overkill) makes SSD vs. rotational a no-brainer.
This will make flash more competitive as it cuts price per GB in half or by a third over that same period.... but I think it is a temporary effect.
Innovation in spinning rust has not slowed down in the labs, I presume, so I think that when the manufacturing sector is back up to speed they'll be pushing out the same innovations in 2013 they would have otherwise, and possibly skip a generation.
Which is not to say that the writing isn't still on the wall for spinning rust... eventually Flash will catch up in price/GB, though volumes will explode even when its still twice as expensive as the demand will ramp up dramatically.
The biggest effect I've seen is that a bunch of companies (generally big companies, but not HDD creators) have made a shit-load of profit from reselling stock they had before the floods because they realised it would make them more money than using the drives for the purpose they were bought for.
https://lwn.net/Articles/497024/ http://bcache.evilpiepirate.org/
AnandTech had a pretty good article about it a few months ago. I'm sure there are others as well.
http://www.anandtech.com/show/5518/a-look-at-enterprise-perf...
http://www.tomshardware.com/reviews/ssd-520-sandforce-review...
Briefly, under heavy use, a drive can last at least five (5) years.
Edit: Corrected a typo.
It's under $1 per GB on SSDs now, which is definitely affordable for average consumers (around $90 for 128GB).
I bought a 256GB SSD in Feb this year for around $360. That exact same drive is now $260.
2 weeks ago I bought a Samsung 830 (faster) 256GB (twice as big) for $220.
The price drop is pretty amazing.
> Due to a typo, this article originally said that a 300 terabyte hard drive sells for about $350.
300 terabytes for $350? In one drive? Sign me up. =D
I've seen both drive controllers get themselves wedged to the point where they just shut down and go dead. This happens because they aren't doing a straight 1-1 mapping and along with wear leveling (which is necessary) they're trying to compress (sandforce) and optimize (intel) the way data is stored but their algorithms are too complex to be fully debugged (as of last year, anyway.)
I've gone back to spinning rust as a result. I'd love to move my servers to flash, but not until they go back to really straightforward controllers (I hear the samsung controllers are very reliable, and this is why Apple chose them for their laptops) and don't do anything more than wear leveling.
For performance in servers, I'm going with a lot of RAM.
Right now the drive failure rates of SSDs are way too high in my experience, and reading between the lines and user reviews on retailer sites it seems that my experience is not unusual.
Servers should basically be created to cope with drive failure in such a way that % of failure isn't an impact on "will this cause problems" but merely on working out the cost of using them.