AWS Snowmobile – Move Exabytes of Data to the Cloud in Weeks
aws.amazon.com
aws.amazon.com
Never underestimate the bandwidth of a station wagon full of
tapes hurtling down the highway.
http://news.bbc.co.uk/2/hi/africa/8248056.stm
It was an obvious publicity stunt trying to bring attention to the slow connection speeds but it does illustrate this issue well.
"Snowmobile is protected by 24/7 video surveillance and alarm monitoring, GPS tracking and may optionally be escorted by a security vehicle while in transit."
Actually 'in the movie' the plot line comes at the end when you find out that the government is behind the heist in order to get hold of a large amount of end user data that they need for surveillance purposes.
If humans later colonize Mars and want to have the full 50-terabytes copy of Wikipedia in the biosphere, it's faster to send some harddrives as a rocket payload on a 6 month journey rather than try to transfer it via the 32kbps uplink[2] which would take ~500 years.
[1] https://en.wikipedia.org/wiki/Sneakernet
[2] http://mars.nasa.gov/msl/mission/communicationwithearth/data...
http://www.space.com/34824-nasa-x-ray-tech-deep-space-commun...
In other words, by the time x-ray communication is a reality to the Mars colony, we'd want to copy 50-petabytes (~15 years transfer[1]) or 50-exabytes (~15000 years transfer). The 6-month rocket journey is still faster than those scenarios.
[1]http://www.wolframalpha.com/input/?i=(50*10%5E15)+%2F+((1*10...
But then I install another 60 GB game and realize not any time soon...
Assuming we get enough people on Mars, they'll have their own internet over there.
At the same time, logistics (incl insurance and security) is handled by companies that are very good at it. Fedex, DHL and the like offer physical security services for goods if you need it in addition to encryption.
Think it's a PR move only. They will probably find a few clients to somehow utilize one truck, but I don't think it's more efficient than Snowballs.
You'd have to find a thousand 1GBe ports in your data center (unless Amazon would ship an expensive switch along with Snowballs) -- that's about two server cages' (10 racks x42U) worth. You will have to find a lot of power -- while Snowmobile can bring a generator.
I doubt there will be a lot of demand for Snowmobiles, though.
I understand the size of a Snowmobile, but if you don't somehow utilize it exactly 100%, I don't think building your own solution is cheaper than renting logistics as a service from others. Basically AWS in the logistics world.
Also - think about the person having to do this - If I had to move 100 petabytes of data to AWS - I really don't want to be messing around with hundreds of appliances, tracking them, figuring what we uploaded onto what devices, figuring out how to fragment the data to properly fit on all the devices, etc....
Definitely want to go the Snowmobile route.
It seems that any datacenter that's big enough to have a PB of data is going to be able to find 25 40 GBit ports easier than 1000 1Gbit ports.
Even if you only utilize 25% of the Snowmobile, that's 250 individual Snowballs you don't have to handle.
(just bounced this off of one of our network engineers -- if we did have to handle 100 snowballs, we would buy switches, set them up in the big conference room next to the datacenter and run the 10 gig fiber drops over to that conference room where the snowballs would be. He said we're probably looking at $30K in hardware costs to set it up, so the Snowmobile might be cheaper even for 100 snowballs worth of data)
[I'm a C++ dev who likes to help design and build datacenters. It's fun.]
Still, you're right that there's a logistic issue of physically locating, networking, and powering these devices.
Corporate types, spending loads of company money, are more interested in convenience and turn key solutions to problems that rigs that cost less but give more to think about or other issues.
Then what happens in 5 years if local storage cost dropped by a factor of 10, but S3 cost did not drop?
Big risk, no?
They likely have multiple racks of high performance storage, likely in multiple data centers.
Most people have suspected that AWS bandwidth costs would go down.. but by and large, they have held steady for 5+ years.
Cloud storage is a commodity these days. The market saves you in this case -- if one cloud provider didn't use the 10x technology and pass along the 10x savings to the customer, another company would do it and steal all of their customers.
If migrating out of AWS is 100 million dollars in bandwidth, not sure you would see a lot of people jump ship to a competitor.. data lockin and all.
Many AWS services have no easy migration path off it.
If a 10x cost reduction storage technology comes along, cloud providers will necessarily adopt it and will reduce their prices by approximately 10x.
Here's why:
- If they don't, it will become more cost effective for potential customers to run their own datacenters rather than put data into the cloud, so their growth will basically stop.
- Even if potential customers don't want to run their own datacenters, both potential and existing customers will put new data to a competitor who did pass along the 10x savings to the customer. So again, their growth will basically stop.
This is the nature of a commodity product in a free market. Basically all cloud providers use an S3-compatible API, and costs and performance are in the same ballpark. There are tons of open source compatibility layers that abstract which provider you are putting to. If one of them starts costing 10x less, you just flip the switch and all new data goes there.
The ability for customers to completely cut off growth of their service if prices don't fall is a supreme motivator. The only case this wouldn't be true is if all of the cloud storage services formed a cartel to fix prices. But given that every time Google or Amazon lowers prices, the other one follows to maintain parity, we have evidence that that isn't the case.
Addendum:
I don't understand your analogy at all. IBM mainframes are a specific type of hardware that excel at highly available batch and transactional processing. Linux is an operating system. Linux is free, so it's infinitely cheaper. Also, IBM mainframes run Linux.
Here are my two counterpoints.
1) Bandwidth prices HAVE fallen 10x in the past N years. Many cloud providers (ovh, etc) DO offer this price drop. Yet how many people really left AWS or GCE for ovh? I would guess not that many.
2) As I said with mainframes.. there is a 10x cheaper option to a mainframe that has been around for oh 15 years. But people are still on them, BECAUSE they are still locked in to them. That is my point. Don't get locked into a single anything. Sending 100 PB data for a commercial entity to hold for you, with no guarantees of future pricing, is a bad move. Locking yourself in is one of the worst things you can do as a company.
I agree that the bandwidth pricing is almost certainly designed to create lock-in. What I am contending is that that is unrelated to the storage pricing. You'll notice that my arguments didn't include bandwidth pricing at all, because it is irrelevant to those arguments.
I'll give a more concrete example. Let's say I want to transfer 10PB out of S3. On their pricing sheet they actually say to contact them to get a quote, but before that the prices are dropping pretty fast as you get more data. e.g. 10TB is 9 cents per GB, but after you get past 300TB you're paying 5 cents per GB.
Let's be pessimistic and assume 5 cents per GB, even though you could probably get it for much cheaper by contacting them.
So 10PB will cost me (10PB * (5c/GB)) = $500K to export
Standard storage is about 2 cents per GB. So your 10PB sitting in S3 is costing you 200K / month just to sit there and do nothing.
Do you see the problem here? Moving onto the 20K / month provider becomes cost-positive after 3 months.
Even if they were to increase bandwidth costs 10x, it'd still become cost-positive in a relatively short amount of time (couple of years).
Furthermore, if you are paying S3 millions of dollars per year to store data, you're almost certainly in a position to get them to contractually agree to that cheaper-than-public bandwidth cost I mentioned earlier, so you don't even have to worry about the situation in which they hold you hostage by increasing bandwidth costs 1000x.
Yet furthermore, to reiterate my previous point even if they were to increase bandwidth costs 1000x on normal customers that didn't have enough clout to get contractual guarantees, that would kill their business. Nobody new would put any new data in them. Sure, they could hold the existing data hostage, but absolutely nobody is going to put any new data there.
Similarly, if they were to not decrease storage costs 10x compared to a virtually-identical competitor, nobody would put new data with them. This is the part that is totally unrelated to bandwidth costs. Everyone would start putting all of their new data into the competitor, regardless of their old data being locked in.
The fact that Amazon and Google both immediately reduce storage prices after the other one does is evidence that this is the case.
@2 I think you underestimate mainframes. People who use them aren't totally stupid. They do a specific thing very well. Right tool for the job and all that.
Considering the cost to store 100PB on site, with redundant disks, geographically distributed, including property leases, power, security, and staffing. $700k might be considerably cheaper.
PSA: consume the full content before you comment on the content.
Around this time last year BackBlaze had 200PB of customer backups. They described storing it on 54,675 hard drives across 1,215 Storage Pods
So imagine 600 storage pods or half of BackBlaze's entire operation, for just one customer. Insane.
I wish they gave more details as to what hardware was in there - are there any pictures of what the trailer looks like on the inside?
It also appears that you never have access to the inside where the racks are. You can only access the last ~4 feet for power and data connections.
Quite a dramatic illustration of the increase in data usage.
If one extrapolates next stop will be a train, and then a container ship full of hard drives.
Between that was FEDEX-ing a NAS. That's pretty much the standard data-exchange format in astronomy, so you have a fast link and don't need to bother plugging a bunch of drives in, just plug the NAS to the power and network and off you go.
So when will The Universe be too small to data considering data use is exponential.
I imagine there are lower upper limits, but this would be the upper upper limit.
Next time my client asks me how "much" a PB is, I can just say "about a shippingcontainer's worth".
I wonder what HDD capacity does it use, and when if we could see a 1EB per container.
Those photos look like they were taken with a poorly focusing cheap smartphone.
If they fill the thing in 30 days, that average to 40 Go/s, way faster than 10gbe or Fibre Channel.
I doubt you have a permanent dedicated 40+GB link between you and the nearest AWS data center.
Shipping by boat would still be way faster than a network upload.
But yeah I'd expect air-shipping as well, if only so that the container can be kept secure, I doubt Amazon is going to send security personnel for week-long trips on cargo ships.
PPS – I will build and personally deliver (in exchange for a photo op and a bloggable story) Snowmobile models to the first 5 customers.
There's no way to verify that this truck full of my corp's valuable data isn't stopped somewhere along the way and cloned and then driven to NSA or something.
> Each Snowmobile includes a network cable connected to a high-speed switch capable of supporting 1 Tb/second of data transfer spread across multiple 40 Gb/second connections. Assuming that your existing network can transfer data at that rate, you can fill a Snowmobile in about 10 days.