6 ms·

How Google Compute Engine stacks up to Amazon EC2

134 points by namityadav 14y ago

gingerlime 14y ago
"Multi-region images We’re not quite sure why AWS doesn’t support this, but images on GCE are multi-region (“multi-zone” in their terms)" Whilst images are not multi-region, you can now copy them[1] across regions with AWS. [1]http://aws.amazon.com/about-aws/whats-new/2013/03/12/announcing-ami-copy-for-amazon-ec2/ http://aws.amazon.com/about-aws/whats-new/2013/03/12/announc...
- count 14y ago
  Keeping the image in one region allows the customer to guarantee data provenance - EU folks don't want their data accidentally getting stuck on US servers (and the other way around).
codesuela 14y ago
Looks impressive but would I be wrong to assume that disk write performance will go down as GCE adoption spreads? As far as I know providing virtualized storage is not something that you can consistently limit like RAM for example. Still competition is always a good thing and especially if the giants are fighting it out (Amazon, Google, Microsoft)
RyanZAG 14y ago
The biggest problem GCE has is not performance or anything technical - it is the terrible, terrible support that Google offers (or fails to offer). I'm pretty sure that most people who have used any of Google's pay for services knows just how bad that support can be. What happens if something goes wrong with your billing on EC2? No problem, you call them up and sort it out. What happens if something goes wrong with your billing on any Google product? You'll be down for days while attempting to get Google to even acknowledge a problem. Until Google fixes their support issues, anybody using a Google service for anything business critical needs to be very careful in weighing up pros and cons.
- cnlwsu 14y ago
  Funny considering the main issue we had with EC2 was amazons horrible support. Even spending a lot (lot) per month our only option was to post on the forums and wait. We ended up at Rackspace after a few years of frustrating issues.
  - MBCook 14y ago
    You didn't sign up for the paid support?
    - elq 14y ago
      From my limited experience - I'm at best on the periphery of dealing with AWS support; I can characterize aws support as such - Free tier - occasionally get answers. answers are generally vague. questions to forums. Paid tier - always get answers. answers are generally vague. you get a support portal. Top tier paid - always get answers. answers are generally vague. amazon employees will fly out to your site to give vague answers.
      - makeshifthoop 14y ago
        So true. I ask a question on the paid support plan and I got an answer copied from a blog I already googled. After several back and forth and escalations, I keep getting the same unhelpful answers and gave up.
      - bborud 14y ago
        Welcome to 2013. To be successful you need to address an immensely huge audience. It is just not physically possible to provide everyone with the kind of customer support you got back when you bought meat from the person who had killed the pig.
    - cnlwsu 14y ago
      We did. And we were paying enough (6 digit monthly bill, not huge but significant I would think) that we did get some nice treatment, although we were definitely not a big enough customer apparently. We adopted a lot of things right away when it came out (ie windows instances, elb). I think that was a mistake and lead to a lot of issues that resulted in responses acting like its not a problem until 6 months later when a fix came out. EC2 was a lot better then my experiences with Terremark though, but we were definitely pennies to them.
rdw 14y ago
It seems like the inter-region bandwidth test would be especially likely to be negatively affected by increased adoption of GCE. They lay a fixed amount of capacity, and if there's relatively few other customers, you get a proportionately larger slice of it. Hard to prove either way, but it seems that hitching your wagon to GCE based off of these performance numbers might not pan out for the long term.
- crucifiction 14y ago
  one part not accounted for is that the wide-open internet, while slower and with more latency, is going to be pretty much 100% reliable compared to their dark fiber link between regions.
count 14y ago
So, uh, could they show what they used to get those numbers? A EC2 micro vs. the GCE big version (I'm not familiar enough with GCE). I'll note a HUGE factor is missing: the word 'cost' isn't even in the article. How much more (or less) are you paying for the apparent increase in performance?
- Evbn 14y ago
  It's like all those Mac reviews. Boot up time and the appearance of the default desktop is all the reviewer cares about. If you want real useful info, go to a industry site not a pop tech chat board like HN.
ScotterC 14y ago
I'd love to see Azure's numbers added. Does anyone use Azure yet?
- ceph_ 14y ago
  Didn't they just have a 16 hour long outage this week?
Rezo 14y ago
On the surface of it, a GCE n1-standard-1 and a EC2 m1.medium may look comparable. Each cost $0.12/hour, both have the same 3.75 GB of RAM etc. Sure, there's some differences; GCE may very well be faster, and the EC2 instance has local storage available, but basically they're comparable. However, when you factor in EC2 Reserved Instances pricing, the EC2 price drops to $0.028/hour. For a small startup with a couple of instances to start with, you can run your entire EC2 setup for the price of one entry-level GCE instance.
- oijaf888 14y ago
  So your total spend on 1 reserved EC2 m1.medium instance is $579.92 a year, not including any bandwidth or EBS storage. On the other hand you could go to OVH and get a dedicated server with 8gb of ram and 2TB of disk and significantly faster processor for $468/year and that includes 5TB of bandwidth a month. (source: http://www.ovh.com/us/dedicated-servers/kimsufi.xml http://www.ovh.com/us/dedicated-servers/kimsufi.xml) Seems like unless you are going to need the ability to quickly scale the number of instances up and down it makes more sense to put your core services on dedicated servers rather than in the cloud.
  - coldtea 14y ago
    >Seems like unless you are going to need the ability to quickly scale the number of instances up and down it makes more sense to put your core services on dedicated servers rather than in the cloud. Which is almost always the case for startups (needing to scale at will).
    - oijaf888 14y ago
      I don't know too many startups that need to scale in minutes as opposed to hours. Even sites that have taken off have usually had growth over a period of months or years. I do think it makes sense if you are doing something massively cpu intensive that has hugely varied demand (video encoding, etc).
tszming 14y ago
If I am leaving EC2 to another cloud provider, one of the reason I can think of is Linode start offering real cloud instance..
davenaff 14y ago
Interesting analysis for sure, but I think the only solid takeaway here is that (unsurprisingly) Google's private backbone between regions offers a huge performance advantage over the public Internet between EC3 regions. As the analysis stands, the individual IO stats aren't meaningful. EC2 instances offer a range of I/O throughput in different instances: a t1.micro has 'low' I/O, an m1.small has 'moderate' and an hi1.4xlarge offers 1.1 GB/s write and 2.0 GB/s write. Cost has to figure into a benchmarked analysis here. And CPU. One of Google's big claims at launch was that CPU time was notably less expensive than on EC2. I'd love to see some data here and clicked the link expecting to find that analysis.
kicktheshoe 14y ago
This benchmark provides some interesting data, but would have been far more useful if they had provided more details. There are many unanswered questions here: How even is a comparison when GCE hardly has any customers sharing the same physical host and network at this stage? It's rather interesting that EC2 read I/O performance was higher despite this consideration. What will happen when you start sharing a physical host on GCE with many other customers? Will you all still get 157 MB per second disk write throughput simultaneously? What type of instances were used (and were they EC2's first generation or second generation instances?). Given the fact that they are looking at ephemeral disk performance it's probably a 1st generation instance. I see no mention of EC2 EBS provisioned IOPs. If performance was a goal then surely having guaranteed provisioned I/O should have been considered? Does GCE even have that feature? Surely IOPS are more important than sequential I/O for most tasks? When benchmarking EBS I/O were EBS-optimized instances selected? Why no mention of EC2 cross region AMI copy? Also GCE is currently in the US only. EC2 is on 5 different continents currently. What are the costs of the instances (using reserved instance pricing for EC2)?
- signifiers 14y ago
  > GCE is currently in the US only. Not correct. They've had EU zones since November: https://developers.google.com/compute/docs/faq#datacenters https://developers.google.com/compute/docs/faq#datacenters https://developers.google.com/compute/docs/pricing#table https://developers.google.com/compute/docs/pricing#table
  - bloke8888 14y ago
    Seems the FAQ has recently been updated, a week ago there was no mention of Europe. Odd.
JDDunn9 14y ago
Of course, if you do switch and become invested in Google's technology, they'll probably just shut it down.

134 points by namityadav 14y ago

"Multi-region images We’re not quite sure why AWS doesn’t support this, but images on GCE are multi-region (“multi-zone” in their terms)" Whilst images are not multi-region, you can now copy them[1] across regions with AWS. [1]http://aws.amazon.com/about-aws/whats-new/2013/03/12/announcing-ami-copy-for-amazon-ec2/ http://aws.amazon.com/about-aws/whats-new/2013/03/12/announc...

count 14y ago
Keeping the image in one region allows the customer to guarantee data provenance - EU folks don't want their data accidentally getting stuck on US servers (and the other way around).

codesuela 14y ago

Looks impressive but would I be wrong to assume that disk write performance will go down as GCE adoption spreads? As far as I know providing virtualized storage is not something that you can consistently limit like RAM for example. Still competition is always a good thing and especially if the giants are fighting it out (Amazon, Google, Microsoft)

RyanZAG 14y ago

The biggest problem GCE has is not performance or anything technical - it is the terrible, terrible support that Google offers (or fails to offer). I'm pretty sure that most people who have used any of Google's pay for services knows just how bad that support can be. What happens if something goes wrong with your billing on EC2? No problem, you call them up and sort it out. What happens if something goes wrong with your billing on any Google product? You'll be down for days while attempting to get Google to even acknowledge a problem. Until Google fixes their support issues, anybody using a Google service for anything business critical needs to be very careful in weighing up pros and cons.

cnlwsu 14y ago
Funny considering the main issue we had with EC2 was amazons horrible support. Even spending a lot (lot) per month our only option was to post on the forums and wait. We ended up at Rackspace after a few years of frustrating issues.
- MBCook 14y ago
  You didn't sign up for the paid support?
  - elq 14y ago
    From my limited experience - I'm at best on the periphery of dealing with AWS support; I can characterize aws support as such - Free tier - occasionally get answers. answers are generally vague. questions to forums. Paid tier - always get answers. answers are generally vague. you get a support portal. Top tier paid - always get answers. answers are generally vague. amazon employees will fly out to your site to give vague answers.
    - makeshifthoop 14y ago
      So true. I ask a question on the paid support plan and I got an answer copied from a blog I already googled. After several back and forth and escalations, I keep getting the same unhelpful answers and gave up.
    - bborud 14y ago
      Welcome to 2013. To be successful you need to address an immensely huge audience. It is just not physically possible to provide everyone with the kind of customer support you got back when you bought meat from the person who had killed the pig.
  - cnlwsu 14y ago
    We did. And we were paying enough (6 digit monthly bill, not huge but significant I would think) that we did get some nice treatment, although we were definitely not a big enough customer apparently. We adopted a lot of things right away when it came out (ie windows instances, elb). I think that was a mistake and lead to a lot of issues that resulted in responses acting like its not a problem until 6 months later when a fix came out. EC2 was a lot better then my experiences with Terremark though, but we were definitely pennies to them.

rdw 14y ago

It seems like the inter-region bandwidth test would be especially likely to be negatively affected by increased adoption of GCE. They lay a fixed amount of capacity, and if there's relatively few other customers, you get a proportionately larger slice of it. Hard to prove either way, but it seems that hitching your wagon to GCE based off of these performance numbers might not pan out for the long term.

crucifiction 14y ago
one part not accounted for is that the wide-open internet, while slower and with more latency, is going to be pretty much 100% reliable compared to their dark fiber link between regions.

count 14y ago

So, uh, could they show what they used to get those numbers? A EC2 micro vs. the GCE big version (I'm not familiar enough with GCE). I'll note a HUGE factor is missing: the word 'cost' isn't even in the article. How much more (or less) are you paying for the apparent increase in performance?

Evbn 14y ago
It's like all those Mac reviews. Boot up time and the appearance of the default desktop is all the reviewer cares about. If you want real useful info, go to a industry site not a pop tech chat board like HN.

ScotterC 14y ago

I'd love to see Azure's numbers added. Does anyone use Azure yet?

ceph_ 14y ago
Didn't they just have a 16 hour long outage this week?

Rezo 14y ago

On the surface of it, a GCE n1-standard-1 and a EC2 m1.medium may look comparable. Each cost $0.12/hour, both have the same 3.75 GB of RAM etc. Sure, there's some differences; GCE may very well be faster, and the EC2 instance has local storage available, but basically they're comparable. However, when you factor in EC2 Reserved Instances pricing, the EC2 price drops to $0.028/hour. For a small startup with a couple of instances to start with, you can run your entire EC2 setup for the price of one entry-level GCE instance.

oijaf888 14y ago
So your total spend on 1 reserved EC2 m1.medium instance is $579.92 a year, not including any bandwidth or EBS storage. On the other hand you could go to OVH and get a dedicated server with 8gb of ram and 2TB of disk and significantly faster processor for $468/year and that includes 5TB of bandwidth a month. (source: http://www.ovh.com/us/dedicated-servers/kimsufi.xml http://www.ovh.com/us/dedicated-servers/kimsufi.xml) Seems like unless you are going to need the ability to quickly scale the number of instances up and down it makes more sense to put your core services on dedicated servers rather than in the cloud.
- coldtea 14y ago
  >Seems like unless you are going to need the ability to quickly scale the number of instances up and down it makes more sense to put your core services on dedicated servers rather than in the cloud. Which is almost always the case for startups (needing to scale at will).
  - oijaf888 14y ago
    I don't know too many startups that need to scale in minutes as opposed to hours. Even sites that have taken off have usually had growth over a period of months or years. I do think it makes sense if you are doing something massively cpu intensive that has hugely varied demand (video encoding, etc).

tszming 14y ago

If I am leaving EC2 to another cloud provider, one of the reason I can think of is Linode start offering real cloud instance..

davenaff 14y ago

Interesting analysis for sure, but I think the only solid takeaway here is that (unsurprisingly) Google's private backbone between regions offers a huge performance advantage over the public Internet between EC3 regions. As the analysis stands, the individual IO stats aren't meaningful. EC2 instances offer a range of I/O throughput in different instances: a t1.micro has 'low' I/O, an m1.small has 'moderate' and an hi1.4xlarge offers 1.1 GB/s write and 2.0 GB/s write. Cost has to figure into a benchmarked analysis here. And CPU. One of Google's big claims at launch was that CPU time was notably less expensive than on EC2. I'd love to see some data here and clicked the link expecting to find that analysis.

kicktheshoe 14y ago

This benchmark provides some interesting data, but would have been far more useful if they had provided more details. There are many unanswered questions here: How even is a comparison when GCE hardly has any customers sharing the same physical host and network at this stage? It's rather interesting that EC2 read I/O performance was higher despite this consideration. What will happen when you start sharing a physical host on GCE with many other customers? Will you all still get 157 MB per second disk write throughput simultaneously? What type of instances were used (and were they EC2's first generation or second generation instances?). Given the fact that they are looking at ephemeral disk performance it's probably a 1st generation instance. I see no mention of EC2 EBS provisioned IOPs. If performance was a goal then surely having guaranteed provisioned I/O should have been considered? Does GCE even have that feature? Surely IOPS are more important than sequential I/O for most tasks? When benchmarking EBS I/O were EBS-optimized instances selected? Why no mention of EC2 cross region AMI copy? Also GCE is currently in the US only. EC2 is on 5 different continents currently. What are the costs of the instances (using reserved instance pricing for EC2)?

signifiers 14y ago
> GCE is currently in the US only. Not correct. They've had EU zones since November: https://developers.google.com/compute/docs/faq#datacenters https://developers.google.com/compute/docs/faq#datacenters https://developers.google.com/compute/docs/pricing#table https://developers.google.com/compute/docs/pricing#table
- bloke8888 14y ago
  Seems the FAQ has recently been updated, a week ago there was no mention of Europe. Odd.

JDDunn9 14y ago

Of course, if you do switch and become invested in Google's technology, they'll probably just shut it down.