5 ms·

And to use 5TB in a month you have to sustain over 15 megabit/sec 24 hours a day. DO is betting on you not doing that (and in aggregate nobody does), because in

And to use 5TB in a month you have to sustain over 15 megabit/sec 24 hours a day. DO is betting on you not doing that (and in aggregate nobody does), because in any other context sustaining 15 megabit is about as expensive as Amazon. Quote a 100 megabit drop some time if you don't believe me. If every VPS on Digital Ocean and Linode actually used their quota, they wouldn't have uplink capacity to support it and the network would fail. That's overselling. Numbers that high are extensive overselling. Linode has 40 Gbit links (at least they used to), and give 2 TB (6 Mbps) to each small Linode, meaning about 7,000 Linodes actually using the quota would saturate the link. They have a few more than that. Do the math. The bandwidth quotas are sales stuff so that you will say exactly this in threads like these, and it's amazing how well it works.

jamiesonbecker 11y ago
The oversubscription model actually works, just like insurance does. We consume massive amounts of bandwidth on a continuous basis due to thousands of servers checking in. (We're not spikey in our bandwidth or CPU -- we're surprisingly rigid at 87% to 90% continuous.) We're anomalous, which kinda proves the point. What is ironic about your statement is that t2 CPU allocations are massively oversubscribed in the same way. (but far more aggressively!)
- jsmthrowaway 11y ago
  A former Linode employee, actually (why did you edit out the ad hominem accusing me of working for Amazon?), but my employment is irrelevant to pointing out that over subscription means you are not actually getting what you pay for. If everybody used it you wouldn't have it any more. It's not yours. You are deciding entirely on perceived instead of actual value. And yes, AWS's CPU over subscription on these instances is important if your workload is CPU intensive. They're very quickly outgrown due to that but useful in a pinch.
  - icebraining 11y ago
    over subscription means you are not actually getting what you pay for. If everybody used it you wouldn't have it any more. It's not yours. You are deciding entirely on perceived instead of actual value. I don't agree; as long as you can always use it when you needed, you are getting what you paid for - that's actual. What you're talking about is the value in an hypothetical situation. Now, if you are actually impacted by the oversubscription, then the actual value is below what was promised, but that doesn't seem to be the case for jamiesonbecker's use of DO.
  - derefr 11y ago
    Like the parent said, it's just like insurance. You can get what you pay for, because almost everybody else is paying for something they don't use, just in case they need it—and there's no situation where everybody's "just in case" will happen at the same time. Aggregated traffic is predictable in a way that individual flows aren't; you can have enough surety to stake your business on giving some of your customers a throughput/latency SLA, without reserving any circuits for them—because you know that even the peak of your traffic's froth is predictably below your uplink's capacity, and that there's no way for this fact to change on a timescale small-enough that your staff can't cope with it.
martinald 11y ago
To answer your question, 100mbit isn't much these days. This article says you can get transit at $0.63/mbit. http://drpeering.net/white-papers/Internet-Transit-Pricing-Historical-And-Projected.php http://drpeering.net/white-papers/Internet-Transit-Pricing-H... Explain why despite wholesale internet prices collapsing, AWS continues to charge the same with virtually no yoy decline? Azure et al also do this fwiw. It's just gouging.
- ranman 11y ago
  What? They decrease prices all the time... There have been something like 3 price decreases just this year.
  - virtuallynathan 11y ago
    AWS has not decreased bandwidth pricing in a long time
- virtuallynathan 11y ago
  You can go even lower depending on the commit. I've seen $0.35/Mbit from the low-end carriers, $0.50 from mid-range.
skuhn 11y ago
Sure, 100Mbit/s circuits aren't very cost effective -- also, no one buys them. If your needs are a little larger, you can do a 1 gig commit on a 10 gig circuit in a well connected datacenter in a first tier region (e.g. Silicon Valley), and figure on paying between $1.50 - $2.00 per megabit at 95th percentile. Much better prices are available with some elbow grease applied. Add $350 per month for the cross connect and $500 for the installation. That comes to $22,700 for a year of service if you don't exceed 1 gigabit/s at the 95th percentile. In Amazon's bandwidth terms, that's 331 terabytes transferred, or $17,000 per month if all of that bandwidth is outbound. You would have to limit your average transfer to 110mbit/s in order to achieve cost parity. So yes, bandwidth is incredibly expensive in AWS and you can easily do better in datacenter-land (or with cloud providers who bill differently), provided that you are operating at sufficient scale.
- discodave 11y ago
  "provided that you are operating at sufficient scale" Where "sufficient scale" probably means you're spending 100k+ a year on infrastructure. Let alone staff costs to support it. Also in your example you have no datacenter redundancy so it's 44k if you want to be in 2 DCs, which means "sufficient scale" is probably more like 200k+. "if you don't exceed 1 gigabit/s at the 95th percentile" That's a big if, the internet, startups and business in general are uncertain and it's often very difficult to know what your 95th percentile will look like in the future. Also, it's unlikely that your network utilization is constant. A utilization rate of 50% would be ambitious. So a better comparison is 44k for 1gig in 2 datacenters vs 102k for AWS (17/2*12), assuming you can predict your network utilization pretty well.
  - amazon_not 11y ago
    > "provided that you are operating at sufficient scale" Nothing says you have to do it all on your own. Let the DC operator do the heavy lifting, all you have to do is rent a dedicated server, VPS or colocate. Total spend to sufficient scale: a few dollars - a few grand per month. > That's a big if, the internet, startups and business in > general are uncertain and it's often very difficult to > know what your 95th percentile will look like in the > future. It's really not that hard. Look at historical usage and assume it continues as before. If you are worried have a spare dedicated server or two with 1G/10G interfaces. > Also, it's unlikely that your network utilization is > constant. A utilization rate of 50% would be ambitious. This really is a moot point as you effectively only pay for peak usage with 95th percentile billing. Alternatively you can pay per terabyte of traffic, which is effectively average usage. Average usage also correlates strongly with peak usage, so in the end it's all the same, give or take a constant. > So a better comparison is 44k for 1gig in 2 datacenters > vs 102k for AWS (17/2*12), assuming you can predict your > network utilization pretty well. Nope. A proper comparison is about $1000 per month for two dedicated servers in two DCs versus $102k per year for AWS.
miles932 11y ago
You think if every ec2 instance used their available upload capacity, simultaneously, that AWS regions wouldn't fail?
dorfsmay 11y ago
Isn't that the whole point of VPS pay a fraction because you you'll only use a fraction. Even internal clouds, people migrate their data centres to VMware, xen, kvm etc... because most servers only use a fraction of CPU and network allocated to them. When you do need full cpu or bandwidth, you know it and you build for it.

jamiesonbecker 11y ago

The oversubscription model actually works, just like insurance does. We consume massive amounts of bandwidth on a continuous basis due to thousands of servers checking in. (We're not spikey in our bandwidth or CPU -- we're surprisingly rigid at 87% to 90% continuous.) We're anomalous, which kinda proves the point. What is ironic about your statement is that t2 CPU allocations are massively oversubscribed in the same way. (but far more aggressively!)

jsmthrowaway 11y ago
A former Linode employee, actually (why did you edit out the ad hominem accusing me of working for Amazon?), but my employment is irrelevant to pointing out that over subscription means you are not actually getting what you pay for. If everybody used it you wouldn't have it any more. It's not yours. You are deciding entirely on perceived instead of actual value. And yes, AWS's CPU over subscription on these instances is important if your workload is CPU intensive. They're very quickly outgrown due to that but useful in a pinch.
- icebraining 11y ago
  over subscription means you are not actually getting what you pay for. If everybody used it you wouldn't have it any more. It's not yours. You are deciding entirely on perceived instead of actual value. I don't agree; as long as you can always use it when you needed, you are getting what you paid for - that's actual. What you're talking about is the value in an hypothetical situation. Now, if you are actually impacted by the oversubscription, then the actual value is below what was promised, but that doesn't seem to be the case for jamiesonbecker's use of DO.
- derefr 11y ago
  Like the parent said, it's just like insurance. You can get what you pay for, because almost everybody else is paying for something they don't use, just in case they need it—and there's no situation where everybody's "just in case" will happen at the same time. Aggregated traffic is predictable in a way that individual flows aren't; you can have enough surety to stake your business on giving some of your customers a throughput/latency SLA, without reserving any circuits for them—because you know that even the peak of your traffic's froth is predictably below your uplink's capacity, and that there's no way for this fact to change on a timescale small-enough that your staff can't cope with it.

martinald 11y ago

To answer your question, 100mbit isn't much these days. This article says you can get transit at $0.63/mbit. http://drpeering.net/white-papers/Internet-Transit-Pricing-Historical-And-Projected.php http://drpeering.net/white-papers/Internet-Transit-Pricing-H... Explain why despite wholesale internet prices collapsing, AWS continues to charge the same with virtually no yoy decline? Azure et al also do this fwiw. It's just gouging.

ranman 11y ago
What? They decrease prices all the time... There have been something like 3 price decreases just this year.
- virtuallynathan 11y ago
  AWS has not decreased bandwidth pricing in a long time
virtuallynathan 11y ago
You can go even lower depending on the commit. I've seen $0.35/Mbit from the low-end carriers, $0.50 from mid-range.

skuhn 11y ago

Sure, 100Mbit/s circuits aren't very cost effective -- also, no one buys them. If your needs are a little larger, you can do a 1 gig commit on a 10 gig circuit in a well connected datacenter in a first tier region (e.g. Silicon Valley), and figure on paying between $1.50 - $2.00 per megabit at 95th percentile. Much better prices are available with some elbow grease applied. Add $350 per month for the cross connect and $500 for the installation. That comes to $22,700 for a year of service if you don't exceed 1 gigabit/s at the 95th percentile. In Amazon's bandwidth terms, that's 331 terabytes transferred, or $17,000 per month if all of that bandwidth is outbound. You would have to limit your average transfer to 110mbit/s in order to achieve cost parity. So yes, bandwidth is incredibly expensive in AWS and you can easily do better in datacenter-land (or with cloud providers who bill differently), provided that you are operating at sufficient scale.

discodave 11y ago
"provided that you are operating at sufficient scale" Where "sufficient scale" probably means you're spending 100k+ a year on infrastructure. Let alone staff costs to support it. Also in your example you have no datacenter redundancy so it's 44k if you want to be in 2 DCs, which means "sufficient scale" is probably more like 200k+. "if you don't exceed 1 gigabit/s at the 95th percentile" That's a big if, the internet, startups and business in general are uncertain and it's often very difficult to know what your 95th percentile will look like in the future. Also, it's unlikely that your network utilization is constant. A utilization rate of 50% would be ambitious. So a better comparison is 44k for 1gig in 2 datacenters vs 102k for AWS (17/2*12), assuming you can predict your network utilization pretty well.
- amazon_not 11y ago
  > "provided that you are operating at sufficient scale" Nothing says you have to do it all on your own. Let the DC operator do the heavy lifting, all you have to do is rent a dedicated server, VPS or colocate. Total spend to sufficient scale: a few dollars - a few grand per month. > That's a big if, the internet, startups and business in > general are uncertain and it's often very difficult to > know what your 95th percentile will look like in the > future. It's really not that hard. Look at historical usage and assume it continues as before. If you are worried have a spare dedicated server or two with 1G/10G interfaces. > Also, it's unlikely that your network utilization is > constant. A utilization rate of 50% would be ambitious. This really is a moot point as you effectively only pay for peak usage with 95th percentile billing. Alternatively you can pay per terabyte of traffic, which is effectively average usage. Average usage also correlates strongly with peak usage, so in the end it's all the same, give or take a constant. > So a better comparison is 44k for 1gig in 2 datacenters > vs 102k for AWS (17/2*12), assuming you can predict your > network utilization pretty well. Nope. A proper comparison is about $1000 per month for two dedicated servers in two DCs versus $102k per year for AWS.

miles932 11y ago

You think if every ec2 instance used their available upload capacity, simultaneously, that AWS regions wouldn't fail?

dorfsmay 11y ago

Isn't that the whole point of VPS pay a fraction because you you'll only use a fraction. Even internal clouds, people migrate their data centres to VMware, xen, kvm etc... because most servers only use a fraction of CPU and network allocated to them. When you do need full cpu or bandwidth, you know it and you build for it.