The Google Compute Engine (GCE) infrastructure as a service (IaaS) is now in general availability. The launch also brings price cuts, a new storage model, expanded operating system support and live migration capabilities for transparent maintenance.
Like Microsoft’s Azure, Google had previously been offering some discount against Amazon’s pricing by having per minute billing. Google have now made a more decisive move on pricing, by reducing on demand instance charges by 10% for the ‘most popular standard Compute Engine instances’. A Google n1-standard-4 instance with 4 vCPU and 15GB RAM is priced at $0.415/hr, which compares to a (reduced just last month) m3.xlarge from Amazon at $0.450/hr. Other comparisons are less easy as although both providers have similar offerings in terms of CPU and RAM most Amazon types have bundled storage and Google have now unbundled storage. Netflix cloud architect Adrian Cockcroft has charted the various options.
Across the entire range of GCE machines a new persistent disk model is offered for storage with a flat pricing of $0.04 GB/month. There are no additional charges for IO operations, but there’s a performance cap that scales with each TB of persistent disk. This allows 300 IOPS for small reads, 1500 IOPS for small writes, 120 MB/s streaming read throughput and 90 MB/s stream write throughput. GCE allows bursting above these limits on boot disk volumes to allow for faster boot up and occasional installation of software. There are also published limits for IOPS and throughput per virtual machine, with some scaling based on cores. Overall this makes Google’s storage more comparable to Amazon’s Elastic Block Store (EBS) with provisioned IOPS (PIOPS). Making that comparison, 1TB with 300 IOPS is $40/month on Google and $151.80 at Amazon. Mileage will vary for real world use cases as Amazon allows flexible scaling of size and performance whilst Google ties size to performance, which might entail overprovisioning in one dimension to get the right capacity for a given application. Google have deprecated their scratch disks, which were similar to Amazon’s instance store.
GCE has moved away from its limited choice of Linux kernels in order to support a broader range of operating systems. At the time of writing only Debian 7 and Centos 6 are available in the web console, but Google have announced a partnership with Red Hat to deliver their Enterprise Linux (RHEL), which is available in limited preview. Support for SUSE and FreeBSD has also been announced, and the service should also run any other Linux distribution, including the Docker focussed CoreOS. Any imported virtual machine images now need their own kernel and init ramdisk in order to run on GCE’s KVM hypervisor.
A feature that was announced before general availability is live migration and automatic restart. This allows virtual machines to be moved to a different availability zone whilst scheduled maintenance is taking place. The fact that GCE has scheduled maintenance was seen as a weakness versus other public IaaS that operate on a continuous basis. Live migration gets around that, and will in many cases be preferable to receiving notice that a VM is scheduled for redundancy and must be manually moved, which is what happens on other services. Google’s ability to do this is helped by their network model, which is able to span availability zones and regions, so a VM can be moved without tearing it away from its IP address and the network it connected to. Their use of KVM is also helpful, as it is much less fussy about machine specifications for migration (VMware requires identical hardware down to the stepping level on the CPU). This means that Google will be able to continuously upgrade the underlying hardware for GCE without causing outages to its users.
Larger instance types, with 16 vCPUs are now available as part of a limited preview. These machines come in 3 different types determined by RAM: standard, 30GB; high memory, 104GB and high CPU, 14.4GB.
Google have previously targeted sales at very large cloud users needing huge quantities of CPU and/or storage, which is borne out in case studies like Mendelics, who are using GCE for DNA sequencing. General availability, price cuts and better operating system support should help to broaden its appeal, though there’s no sign of Windows support yet.