When deploying resources in Compute Engine, you choose which region and zone host your resources. This decision determines where your data is stored and processed, and directly affects your application's reliability, latency, cost, and environmental footprint.
This guide outlines criteria and best practices for selecting Google Cloud regions and zones for your workloads.
Terminology
Region: An independent geographic area consisting of zones (typically three or more). Putting resources across different regions provides failure independence and geographic disaster recovery.
Zone: A deployment area within a region. Zones have high-bandwidth, low-latency network connections to other zones in the same region. Google Cloud designs zones to minimize correlated failures caused by physical infrastructure events like power, cooling, or networking outages.
AI zone: A specialized zone geographically separated from standard zones that provides high-density GPU and TPU capacity for large-scale training and batch inference. AI zones share infrastructure with a designated parent zone and require enabling AI zone visibility. For more information, see AI zones.
Round-trip time (RTT): The time it takes to send an IP packet and receive the acknowledgment.
When to choose your regions and zones
Decide how many and which Compute Engine regions and zones to use early in the architecture phase. Review the following considerations:
- Data consistency and replication: Multi-region architectures require mechanisms to synchronize or shard data across geographic distances.
- Cost implications: Resource pricing and cross-region data transfer charges vary by location.
- Migration overhead: Moving live workloads and data between regions is operationally complex and incurs downtime or egress costs.
Factors to consider when selecting regions and zones
Consider the following core factors when planning your deployment:
High availability and failure tolerance
- When you need to survive localized outages: Choose to deploy your instances across multiple zones within a single region so that if one zone goes down, then others remain available.
- When you have business-critical disaster recovery needs: Choose a secondary region to deploy backup services or multi-region replicas.
User and network latency
- When your users are geographically concentrated: Choose a region that is geographically nearby to minimize round-trip time.
- When you serve a global user base from one location: Choose a US region, because they typically offer the best global connectivity.
- When you have an enterprise hybrid environment: Choose your region based on the lowest latency to your on-premises systems connected using Cloud Interconnect or Cloud VPN.
Machine series availability for Artificial Intelligence (AI) and Machine Learning (ML) acceleration
- When you require specific standard machine series: Choose your target zones only after verifying that your required machine series (such as C4, C3, N4, or M3) are supported there.
- When you run AI and ML training workloads: Select regions that support accelerator-optimized machine series (such as A4 with NVIDIA B200 GPUs or A3 with NVIDIA H200/H100 GPUs) or TPUs (such as TPU v6e or TPU v5p). For large-scale training, consider regions with dedicated AI zones or use Dynamic Workload Scheduler (DWS) to secure capacity in advance.
- When you run real-time inference workloads: Prioritize standard zones close to your user base using inference-optimized instances (such as G2 with NVIDIA L4 GPUs) rather than AI zones to prevent inter-zone latency to your backend services.
Pricing and data transfer costs
- When you want to minimize network costs: Choose to keep communication within a single zone (which is free) or within the same region (which incurs minimal costs), as crossing into different regions incurs higher inter-region data transfer charges.
Colocation with other services
- When you rely on other Google Cloud services (such as Spanner, Cloud SQL, BigQuery, or Cloud Storage): Choose to colocate your Compute Engine resources in the same region as those dependent services to avoid cross-service latency and eliminate cross-region egress charges.
Resource quotas and capacity
- When planning a launch: Choose regions where you have verified and requested sufficient regional and zonal resource quota ahead of time.
Carbon-free energy percentage (CFE%)
- When you want to prioritize sustainability: Choose a region with a higher CFE% to reduce your application's gross carbon footprint.
Automated location selection
Compute Engine provides built-in mechanisms to automatically select optimal zones based on real-time capacity and availability:
- Google Cloud console ("Any" zone): When creating an instance in the Google Cloud console, select Any in the zone selector. Compute Engine automatically provisions the instance in a zone with sufficient capacity for the chosen machine type.
- Regional Managed Instance Groups (MIGs): Configure your regional MIG with a
target distribution shape
(
ANY,BALANCED, orANY_SINGLE_ZONE) to let the MIG automatically deploy instances across available zones. - Bulk instance creation API: When creating multiple instances simultaneously, use a regional endpoint with the bulk creation API to automatically distribute instances across zones with available capacity.
Global deployment patterns
Depending on your latency and resiliency requirements, choose one of the following architecture models:
Pattern 1: Single-region deployment
Deploy your application frontend and backend in a single region across multiple zones.
- Best for: Applications with moderate latency tolerance, internal enterprise tools, or workloads with centralized transactional databases.
- Optimization: Combine with Cloud Load Balancing and Cloud CDN to terminate SSL/TCP sessions at edge Points of Presence (PoPs) and cache static assets near users.
Pattern 2: Distributed frontend with centralized backend
Deploy frontend servers in multiple geographic regions close to users while maintaining a centralized backend and database in a single region.
- Best for: Applications that handle mostly read requests, employ aggressive edge or frontend caching, or process writes asynchronously.
- Optimization: Use Cloud CDN to cache static assets and Memorystore to cache frequently accessed data in the same region as the frontend.
Pattern 3: Distributed frontend and backend in multiple regions
Deploy full-stack frontend and backend services across multiple regions.
- Best for: High-throughput, globally distributed applications requiring minimal latency worldwide.
- Data storage: Use globally distributed databases like Spanner multi-region configurations or Cloud Storage dual-region or multi-region locations to provide synchronous local read and write capabilities across continents.
Pattern 4: Multiple parallel regional applications
Run completely independent, self-contained application stacks in different regions, routing users to their local partition with minimal cross-region data sync.
- Best for: Multi-tenant SaaS platforms, multi-region retail stores with localized inventory, or real-time multiplayer gaming worlds.
- Data storage: Use Cloud SQL to provide PostgreSQL, SQL Server, and MySQL databases in each region.
Optimizing and measuring latency
Regardless of your deployment model, you can apply the following techniques to minimize end-user latency:
- Use Premium Tier networking: Premium Tier (the default) routes traffic over Google's private global fiber backbone rather than the public transit internet, ensuring lower latency and reduced packet loss.
- Terminate connections at the edge: Use External Application Load Balancers and Cloud CDN to terminate TLS handshakes and serve cached content from the edge PoP closest to the user.
- Implement in-memory caching: Place Memorystore for Redis or Memcached clusters in the same region as your compute resources to eliminate repeated database round-trips.
- Measure real-world latency:
- Use VPC Flow Logs to inspect TCP round-trip times between clients and instances.
- Enable Custom Request Headers on your load balancer to log client RTT and geographic location data.
- Use Cloud Monitoring metrics for load balancers (
frontend_tcp_rttandbackend_latencies).
Step-by-step selection checklist
Use this step-by-step checklist to finalize your region and zone selection:
- Plan capacity: Use Dynamic Workload Scheduler (DWS) Calendar mode for predictable future reservations or Flex-start for queued batch jobs, and the console Capacity Advisor to discover and plan capacity.
- Determine compliance and data residency requirements: Ensure that the candidate regions comply with relevant data sovereignty and regulatory rules.
- Select primary regions based on user concentration: Place your services in regions that offer the lowest median RTT to your core user base.
- Design for resiliency:
- In every selected region, spread compute instances across at least two to three zones.
- For critical workloads, establish a secondary failover region.
- Evaluate pricing and sustainability: If multiple regions meet your technical criteria, compare resource pricing and Carbon-Free Energy percentages (CFE%) to make the final selection.
Summary of best practices
| Category | Recommended practice |
|---|---|
| Resiliency and high availability | Distribute instances across at least three zones in a region to protect against infrastructure outages. Deploy in a secondary region for cross-region disaster recovery. |
| Automated zone selection | Use the console Any zone option, regional MIG distribution shapes, or the Bulk Instance Creation API to automatically select zones with available capacity. |
| AI and ML workloads | For large-scale training, select regions with specialized AI zones or use Dynamic Workload Scheduler (DWS) to schedule GPU and TPU capacity. For real-time inference, deploy inference-optimized machine series (such as G2) in standard zones close to users. |
| Latency optimization | Use Premium Tier networking, External Application Load Balancing, and Cloud CDN to terminate sessions and cache assets at Google edge PoPs. |
| Cost management | Keep intra-service communication within the same zone or region to avoid cross-region network egress charges. |
| Sustainability | Prioritize regions with higher Carbon-Free Energy percentages (CFE%) to minimize environmental impact. |
What's next
- Review all Available Compute Engine regions and zones.
- Learn how to Design resilient systems on Google Cloud.
- Learn about Optimizing application latency with Cloud Load Balancing.
- Explore Google Cloud region carbon footprints and CFE%.
For architects selecting regions for AI/ML workloads:
- Read about AI zones for GPU and TPU workloads.
- Check GPU model availability by region and zone.
- Check TPU availability by region and zone.
- Learn about Dynamic Workload Scheduler (DWS).