Overall Satisfaction with Google Compute Engine
Google Compute Engine (GCE) is used for most of the AI workload spanning both on-prem private cloud and public cloud. It is used for both onetime training phase for our Deep Learning workload as well as ongoing Deep Learning inference for customer facing applications.
- East interface to scale up and down the compute capacity
- Easy, straight forward billing and chargeback capabilities
- Reliability / uptime is great and had no issues so far on uptime
- Works well in multi cloud environments
- Although not always used, there is room for adding more detailed and granular management console when things go wrong (and sometime they do)
- Documentation can sometime be hard to find especially for using GCE for time critical, large scale deployments
- There are also some compatibility issues when running custom libraries over GCE. Support for third party drivers and libraries can be improved.
- Positive impact on the OpEx with reduction in CapEx resulting from reducing the time to move a workload from on-prem to the cloud
- Incased RoI by reducing need for on-prem compute
- Improved agility by providing the option to take on new AI workloads for test and dev without the need for upfront investment in
- Amazon Elastic Compute Cloud (EC2), Microsoft Azure and Oracle Cloud Infrastructure (OCI)
After all the discounts, GCE is a bit cheaper with much less incidental expenses to deploy and maintain compared to Amazon, Microsoft Azur and Oracle (OCI). It is also easy to manage as the interface is simpler compared to AWS or Azur.
Support so far has been excellent with direct help from Google engineering and support engineering teams.
Do you think Google Compute Engine delivers good value for the price?
Are you happy with Google Compute Engine's feature set?
Did Google Compute Engine live up to sales and marketing promises?
Did implementation of Google Compute Engine go as expected?
Would you buy Google Compute Engine again?
GCE is very well suited to small to mid scale deployments. It is also run very well with AI workloads especially when using Google TenserFlow et al. It is less appropriate for extremest scale deployments that spam multiple data center (probably because of lack of document on best practices)