Server size by concurrent users
Generated from this calculator. These are planning estimates, not benchmarks; load-test before you commit.
| Concurrent users | vCPU | RAM | Bandwidth | Requests/sec |
|---|---|---|---|---|
| 100 | 1 | 2 GB | 11 Mbps | 17 |
| 500 | 1 | 8 GB | 51 Mbps | 84 |
| 1,000 | 2 | 15 GB | 101 Mbps | 167 |
| 5,000 | 7 | 74 GB | 501 Mbps | 834 |
| 10,000 | 13 | 147 GB | 1,001 Mbps | 1,667 |
Related calculators: Bandwidth and CDN · Database Size · Serverless Cost · API Latency
Key Takeaways
- Server capacity planning prevents both over-provisioning (wasted cost) and under-provisioning (poor performance)
- Always include redundancy - N+1 provides 50% extra capacity for failover
- Different workloads have different bottlenecks: web apps need CPU, databases need RAM and fast storage
- Plan for peak traffic, not average - your servers must handle traffic spikes
- Monitor actual usage and adjust - capacity planning is an ongoing process
Understanding Server Capacity Planning
| Application Type | Typical RPS (single server) | Memory per Request | Recommended CPU Cores | Notes |
|---|---|---|---|---|
| Static file server (nginx) | 10,000–50,000 RPS | ~1–5 KB | 4–8 cores | Mostly I/O bound; fast |
| REST API (Node.js/Go) | 1,000–10,000 RPS | ~1–10 MB | 4–8 cores | Async/lightweight |
| REST API (Python/Flask) | 100–1,000 RPS | ~5–50 MB | 4–8 cores | GIL limits concurrency |
| REST API (Java/Spring) | 500–5,000 RPS | ~50–200 MB | 8–16 cores | JVM overhead; scales well |
| Web app (PHP/WordPress) | 100–500 RPS | ~10–50 MB/page | 4–8 cores | Database-bound |
| ML inference (CPU) | 1–10 RPS | ~100 MB–2 GB | 8–32 cores | Very resource intensive |
| ML inference (GPU) | 10–100 RPS | ~1–8 GB GPU VRAM | GPU required | Dramatically faster than CPU |
| Database (PostgreSQL) | 1,000–10,000 QPS | ~1–10 MB/connection | 8–32 cores | Query complexity matters greatly |
| Video streaming | 100–500 concurrent streams | ~1–5 Mbps/stream | 4–8 cores | Network bandwidth often the bottleneck |
| Rule | Description | Formula | Notes |
|---|---|---|---|
| CPU target utilization | Keep CPU at 60–70% peak for headroom | Servers needed = Peak RPS / (Max RPS × 0.65) | Never run >80% sustained |
| Memory headroom | Provision 20–30% extra memory | Memory = (Requests × Memory per req) × 1.25 | Account for OS and caching overhead |
| Concurrency estimation | Thread count = (RPS × Average Response Time) | Little's Law: L = λ × W | L=concurrent users, λ=requests/sec, W=avg response time |
| N+1 redundancy | Always have spare capacity | Minimum servers = calculated + 1 | Allows maintenance without downtime |
| Peak traffic multiplier | Size for 2–5× average traffic | Typical spikes: 2–10× during events | Use autoscaling for variable load |
| Storage headroom | Reserve 20–30% of disk for growth | Disk = Data size × 1.3 | SSDs degrade with high utilization |
| Network bandwidth | Calculate 70% of NIC capacity | Usable bandwidth = NIC rate × 0.7 | Leaves headroom for bursts |
| Factor | Vertical Scaling (Scale Up) | Horizontal Scaling (Scale Out) | Better Choice |
|---|---|---|---|
| What it means | Bigger machine (more CPU/RAM) | More machines | Depends on use case |
| Cost at scale | Diminishing returns; expensive | Linear cost scaling | Horizontal at large scale |
| Implementation | Change instance type | Add instances + load balancer | Vertical simpler initially |
| Downtime required | Often yes (reboot for resize) | No (rolling addition) | Horizontal |
| Stateful applications | Easier (single machine) | Complex (session sharing) | Vertical |
| Maximum limit | Hard ceiling (largest VM) | Virtually unlimited | Horizontal |
| Failure impact | Single point of failure | Distributed (redundant) | Horizontal |
| Database scaling | Vertical often necessary first | Read replicas; then sharding | Hybrid common |
| Best for | Databases, legacy apps, simpler architecture | Web/API servers, microservices, high-traffic | Context-dependent |
Server capacity planning is the process of determining the compute, memory, storage, and network resources required to meet application performance goals. Effective capacity planning balances cost efficiency with the ability to handle peak workloads while maintaining acceptable response times.
Key Factors in Server Sizing
- Concurrent Users: The maximum number of simultaneous active users your application needs to support
- Request Rate: How many requests each user generates per minute or second
- Response Size: The average payload size of server responses affects bandwidth requirements
- Response Time: Your target latency determines how much parallel processing capacity you need
- Workload Type: CPU-bound, memory-bound, or I/O-bound workloads have different resource requirements
Capacity Formulas
Requests/Second = Concurrent Users x Requests/User/Minute / 60
Workload Type Considerations
Web Applications: Generally CPU-bound with moderate memory needs. Focus on fast response times and session management. Caching can significantly reduce server load.
API Servers: High request rates with small payloads. Optimize for throughput and connection handling. Consider rate limiting and request queuing.
Database Servers: Memory-intensive with high I/O requirements. SSD storage is critical. Memory should be sized to hold working dataset for optimal performance.
Compute Intensive: CPU-bound workloads like video encoding, ML inference. Benefit from high core counts and may need GPU acceleration.
Media Streaming: Bandwidth-intensive with moderate CPU needs. Content delivery networks (CDNs) can offload significant traffic.
Redundancy Strategies
- No Redundancy: Single point of failure - acceptable only for development or non-critical systems
- N+1 (Basic): One extra unit for failover - industry standard for most production workloads
- 2N (High Availability): Full duplicate capacity - for business-critical applications
- 3N (Mission Critical): Triple redundancy - for financial systems, healthcare, and life-safety applications
Best Practices
- Start with conservative estimates and scale based on actual metrics
- Use auto-scaling to handle variable traffic patterns cost-effectively
- Implement monitoring and alerting before you need it
- Plan for 3-6 months of growth, not just current needs
- Consider geographic distribution for global applications
- Test your capacity assumptions with load testing tools