As a result of verifying the reliability of AWS and GCP, AWS wins overwhelmingly



Amazon Web Services (AWS) and Google Cloud Platform (GCP) are representative examples of cloud computing that uses the computer functions of business operators via networks. Engineer Pierce Freeman verified the reliability of these two.

AWS vs GCP reliability is wildly different | Pierce Freeman

https://freeman.vc/notes/aws-vs-gcp-reliability-is-wildly-different

Freeman points out that in the two services of using the service when needed and paying only for what you use, ``not being able to use when needed'' can lead to resource availability problems. Therefore, we used both services a certain number of times and measured the access time and error rate.

Mr. Freeman accessed T4 GPUs for deep learning at random times of the day, and over the course of two weeks, he accessed 3,000 times per service.

As a result, AWS was able to consistently start the GPU in less than 15 seconds, averaging 11.4 seconds, but GCP took an average of 42.6 seconds. It seems that the number of errors occurred once in two weeks for AWS and 84 times for GCP in two weeks. The graph below shows the startup time on the vertical axis and the time step on the horizontal axis.



Mr. Freeman thought that 'Ready' indicated by AWS was a lie because the difference between AWS and GCP was so extreme, but it seems that there was no problem as a result of verification.

Freeman said, ``Since the services provided by the two cloud vendors are not the same, this may be related to the difference in response time.With GCP, you can attach a GPU as a hardware accelerator to any VM. You can set the number of CPUs separately as needed. AWS only provides predefined VMs with GPUs attached, and here we use g4dn.x series hardware.' also said.

Freeman said, “If you want on-demand computers that you can use when you need them, AWS is the answer. Google may win because of the customizability of wear acceleration.The entire test harness is open source, so we encourage you to do your own analysis, including different GPU configurations and spot instances.'

in Web Service, Posted by log1p_kr