Infrastructure planning
Bare Metal vs. Hyperscale Public Cloud: When Dedicated Hardware Outperforms Virtualisation
Compare sustained I/O, CPU isolation, tail latency and operating cost for database-heavy workloads using a fair benchmark.
Compare configurations, not labels
Bare metal gives an organisation exclusive access to a physical server. A virtual instance exposes an allocated portion of a host through a hypervisor. That difference matters for hardware control, but it does not predict every performance result.
Modern cloud platforms can deliver efficient virtualisation, and hyperscalers also offer bare-metal instances. Compare CPU generation, storage design and allocation guarantees. An older dedicated server may lose to a newer virtual machine on the same application.
Find the bottleneck in the database’s I/O path
A database committing small transactions stresses durable write latency; a reporting scan stresses throughput. Test the actual read/write mix, block sizes and concurrency. Headline IOPS at a deep queue do not describe the delay of a single synchronous commit.
Dedicated local NVMe can provide direct control over the storage path. Network-attached cloud disks add their own volume and instance limits, but may also provide useful managed durability and resizing. Match the resilience design before calling one option cheaper.
Never compare an unreplicated local disk with a managed replicated service as though they provided the same failure protection. Include recovery capacity, backup traffic and rebuild behaviour in the test.
Account for contention and hardware placement
On shared infrastructure, other workloads can contend for resources that the service does not exclusively allocate. Burstable instances can also have credit limits. Check the instance contract and observe CPU steal time, storage latency and sustained throughput instead of assuming all virtual machines have noisy-neighbour problems.
Bare metal removes other customers from that host, but your own applications can still compete for CPU, memory and I/O. On multi-socket servers, inspect NUMA locality, memory placement and network IRQ affinity. Pinning without measurement can make performance worse.
Run a benchmark that reflects production
Use an isolated test environment and representative data. Do not run destructive disk benchmarks against production volumes.
- Hold database version, durability settings, indexes and application concurrency constant.
- Measure median and p95/p99 transaction latency alongside throughput and errors.
- Run long enough to expose cache effects, credits, thermal behaviour and sustained storage limits.
- Test backups, replication and a recovery event while recording the remaining service capacity.
- Report resource utilisation and the cost of the complete deployment, including spare capacity.
Choose the operating model you can sustain
Steady, demanding databases can make good use of dedicated hardware when the team needs predictable allocation and can operate the stack. Short-lived jobs, highly variable demand or dependencies on managed services can favour public cloud.
Total cost includes compute, storage performance, licences, backups, transfer, support and engineering time. Include idle headroom and failover nodes on both sides. Gloman Dedicated Infrastructure is a candidate for sustained high-I/O workloads; the decision should follow your benchmark and recovery requirements.