4 ms·

Compute Engine Preemptible VMs Are Now Generally Available

39 points by boulos 11y ago

WestCoastJustin 11y ago
TL/DR - Preemptible VMs are the same as regular instances except for one key difference - they may be shut down at any time [1, 2, 3]. Similar to early AWS spot instances from what I gather (minus the bidding exchange). [1] http://googlecloudplatform.blogspot.ca/2015/05/Introducing-Preemptible-VMs-a-new-class-of-compute-available-at-70-off-standard-pricing.html http://googlecloudplatform.blogspot.ca/2015/05/Introducing-P... [2] https://cloud.google.com/compute/docs/instances/preemptible https://cloud.google.com/compute/docs/instances/preemptible [3] https://news.ycombinator.com/item?id=9564122 https://news.ycombinator.com/item?id=9564122
- boulos 11y ago
  And of course we give you a 70% discount ;). Disclaimer: I worked on Preemptible VMs.
  - boundlessdreamz 11y ago
    I love Google Cloud except for Google Cloud Storage request pricing. Cloud Storage considers retrieving an object through its HTTP url is considered a class B XML request type which is priced at $0.01/10,000 ops. This is 1.3x-2.5x more expensive than cloudfront and s3 respectively Can you get somone who works on Cloud Storage to take a look at the pricing please? :) It seems like an anomaly since Google Cloud is usually at par or cheaper than AWS equivalents. Also, the pricing on this page https://cloud.google.com/storage/pricing#operations-pricing https://cloud.google.com/storage/pricing#operations-pricing specifies only XML API pricing. I didn't expect HTTP GET requests to be counted as XML API requests and that caused me some trouble. We moved from cloudfront to Cloud Storage and the bill increased
    - boulos 11y ago
      Sorry for the confusion; I'll file a docs bug to make it clear that direct web serving of GCS buckets triggers class B get requests. The pricing difference has been a long battle...
mjn 11y ago
Pretty reasonable pricing in return for the preemptible nature. Dividing the numbers on the pricing page [1]: If you use them as a replacement for occasional (full-price) instances, it's a 70% discount across the board. If you use them as a replacement for mostly-on instances that would qualify for the sustained-usage discount, it's a 57% discount on top of that. So for example an always-on n1-standard-1 goes from about $25/mo to $11/mo. That ends up actually being cheaper than an equivalent DigitalOcean server, if your workload is ok with being preempted. [1] https://cloud.google.com/compute/#pricing https://cloud.google.com/compute/#pricing
- MichaelGG 11y ago
  I think they don't allow you to run for more than 24-hours straight though. Probably to make sure people aren't unhappily surprised when they do need to preempt.
  - boulos 11y ago
    That's exactly right: I don't want you to be surprised when a few weeks after we've landed a lot of capacity that you can't run things for a week straight. If you're interested we explained some of this in an interview on InfoQ: http://www.infoq.com/news/2015/05/google-preemptible-vm http://www.infoq.com/news/2015/05/google-preemptible-vm
- boulos 11y ago
  I personally prefer to think of the 1 to 1.75 cents per core-hour as being competitive with "roll your own": if you're thinking about building out more servers or using preemptible, I think preemptible comes out ahead a lot. The distinction though is that these are from our spare capacity, so it's not a replacement for a 24x7 anything (whether that's our regular VMs or a DO droplet). But if you're prepared for disruptions, come and play!
  - mjn 11y ago
    That's a good point, and sort of what I was using the monthly rate to eyeball. I think in (amortized) monthly costs for other ways of provisioning bulk computation, e.g. owning+powering your own hardware, renting a dedicated server, or VPSs. Not really equivalent, since those are better options if 24/7 uptime is needed, while a physically owned server or rented dedicated server are poorly priced when only occasional usage is needed. But I have a feel for what bulk compute currently costs me per-month with full utilization, as a starting point for comparing other options.
teraflop 11y ago
The documentation says that you get 30 seconds warning before your VM is forcibly terminated. Does anyone know if that's enough time to spin up a new instance to take over from the one being preempted? I haven't messed with GCE much, so I don't know what the typical startup latency is.
- boulos 11y ago
  That's just about our median time to "create to ssh". I wouldn't encourage betting on being able to spin up a replacement when you're getting shot: that's precisely when we might be running out of capacity! If you're looking to checkpoint and save state though, you can mount your persistent disks separately and store whatever you need. You can also start your shot VM back up later when capacity is available again.
  - toomuchtodo 11y ago
    Your team might want to consider extending the warning. AWS gives two minutes notice: https://aws.amazon.com/blogs/aws/new-ec2-spot-instance-termination-notices/ https://aws.amazon.com/blogs/aws/new-ec2-spot-instance-termi...
    - boulos 11y ago
      Thanks for the feedback. We explained the reasoning behind 30 seconds vs "longer" in the InfoQ interview I mentioned in a different thread. But basically if you're getting preempted it's often because some regular VM is trying to land on the box and needs to be kicked off. Our goal here is to make sure we don't impact even our 95th-percentile VM creation time for regular VMs. We create and boot VMs so much faster (I think even p90 is 36s right now) that "2 minutes" is more than it sounds like. Finally, one of the things we tracked carefully during Beta is how quickly people actually shut down. Median on Linux was something like 5 seconds. Even the 90th percentile was something like 10 or 15 seconds. There might be people avoiding the product because they "need" more time, but honestly most of those use cases aren't really prepared for preemption and should probably stick with a different solution.
MichaelGG 11y ago
I love Google Compute Engine. On projects where we've been offered significant Azure credits, I strongly recommend to still use GCE. It's fast, it's cheap, and the management UI is simple and makes sense. Oh and their networking is great (fast and reliable).
- boulos 11y ago
  Glad to hear it! For those that are curious we have a $300 free trial for 60 days, as well as a $100,000 startups program for those that qualify (https://cloud.google.com/developers/startups/ https://cloud.google.com/developers/startups/)

39 points by boulos 11y ago

TL/DR - Preemptible VMs are the same as regular instances except for one key difference - they may be shut down at any time [1, 2, 3]. Similar to early AWS spot instances from what I gather (minus the bidding exchange). [1] http://googlecloudplatform.blogspot.ca/2015/05/Introducing-Preemptible-VMs-a-new-class-of-compute-available-at-70-off-standard-pricing.html http://googlecloudplatform.blogspot.ca/2015/05/Introducing-P... [2] https://cloud.google.com/compute/docs/instances/preemptible https://cloud.google.com/compute/docs/instances/preemptible [3] https://news.ycombinator.com/item?id=9564122 https://news.ycombinator.com/item?id=9564122

boulos 11y ago
And of course we give you a 70% discount ;). Disclaimer: I worked on Preemptible VMs.
- boundlessdreamz 11y ago
  I love Google Cloud except for Google Cloud Storage request pricing. Cloud Storage considers retrieving an object through its HTTP url is considered a class B XML request type which is priced at $0.01/10,000 ops. This is 1.3x-2.5x more expensive than cloudfront and s3 respectively Can you get somone who works on Cloud Storage to take a look at the pricing please? :) It seems like an anomaly since Google Cloud is usually at par or cheaper than AWS equivalents. Also, the pricing on this page https://cloud.google.com/storage/pricing#operations-pricing https://cloud.google.com/storage/pricing#operations-pricing specifies only XML API pricing. I didn't expect HTTP GET requests to be counted as XML API requests and that caused me some trouble. We moved from cloudfront to Cloud Storage and the bill increased
  - boulos 11y ago
    Sorry for the confusion; I'll file a docs bug to make it clear that direct web serving of GCS buckets triggers class B get requests. The pricing difference has been a long battle...

mjn 11y ago

Pretty reasonable pricing in return for the preemptible nature. Dividing the numbers on the pricing page [1]: If you use them as a replacement for occasional (full-price) instances, it's a 70% discount across the board. If you use them as a replacement for mostly-on instances that would qualify for the sustained-usage discount, it's a 57% discount on top of that. So for example an always-on n1-standard-1 goes from about $25/mo to $11/mo. That ends up actually being cheaper than an equivalent DigitalOcean server, if your workload is ok with being preempted. [1] https://cloud.google.com/compute/#pricing https://cloud.google.com/compute/#pricing

MichaelGG 11y ago
I think they don't allow you to run for more than 24-hours straight though. Probably to make sure people aren't unhappily surprised when they do need to preempt.
- boulos 11y ago
  That's exactly right: I don't want you to be surprised when a few weeks after we've landed a lot of capacity that you can't run things for a week straight. If you're interested we explained some of this in an interview on InfoQ: http://www.infoq.com/news/2015/05/google-preemptible-vm http://www.infoq.com/news/2015/05/google-preemptible-vm
boulos 11y ago
I personally prefer to think of the 1 to 1.75 cents per core-hour as being competitive with "roll your own": if you're thinking about building out more servers or using preemptible, I think preemptible comes out ahead a lot. The distinction though is that these are from our spare capacity, so it's not a replacement for a 24x7 anything (whether that's our regular VMs or a DO droplet). But if you're prepared for disruptions, come and play!
- mjn 11y ago
  That's a good point, and sort of what I was using the monthly rate to eyeball. I think in (amortized) monthly costs for other ways of provisioning bulk computation, e.g. owning+powering your own hardware, renting a dedicated server, or VPSs. Not really equivalent, since those are better options if 24/7 uptime is needed, while a physically owned server or rented dedicated server are poorly priced when only occasional usage is needed. But I have a feel for what bulk compute currently costs me per-month with full utilization, as a starting point for comparing other options.

teraflop 11y ago

The documentation says that you get 30 seconds warning before your VM is forcibly terminated. Does anyone know if that's enough time to spin up a new instance to take over from the one being preempted? I haven't messed with GCE much, so I don't know what the typical startup latency is.

boulos 11y ago
That's just about our median time to "create to ssh". I wouldn't encourage betting on being able to spin up a replacement when you're getting shot: that's precisely when we might be running out of capacity! If you're looking to checkpoint and save state though, you can mount your persistent disks separately and store whatever you need. You can also start your shot VM back up later when capacity is available again.
- toomuchtodo 11y ago
  Your team might want to consider extending the warning. AWS gives two minutes notice: https://aws.amazon.com/blogs/aws/new-ec2-spot-instance-termination-notices/ https://aws.amazon.com/blogs/aws/new-ec2-spot-instance-termi...
  - boulos 11y ago
    Thanks for the feedback. We explained the reasoning behind 30 seconds vs "longer" in the InfoQ interview I mentioned in a different thread. But basically if you're getting preempted it's often because some regular VM is trying to land on the box and needs to be kicked off. Our goal here is to make sure we don't impact even our 95th-percentile VM creation time for regular VMs. We create and boot VMs so much faster (I think even p90 is 36s right now) that "2 minutes" is more than it sounds like. Finally, one of the things we tracked carefully during Beta is how quickly people actually shut down. Median on Linux was something like 5 seconds. Even the 90th percentile was something like 10 or 15 seconds. There might be people avoiding the product because they "need" more time, but honestly most of those use cases aren't really prepared for preemption and should probably stick with a different solution.

MichaelGG 11y ago

I love Google Compute Engine. On projects where we've been offered significant Azure credits, I strongly recommend to still use GCE. It's fast, it's cheap, and the management UI is simple and makes sense. Oh and their networking is great (fast and reliable).

boulos 11y ago
Glad to hear it! For those that are curious we have a $300 free trial for 60 days, as well as a $100,000 startups program for those that qualify (https://cloud.google.com/developers/startups/ https://cloud.google.com/developers/startups/)