Unit 7.3: Performance Engineering
Introduction
FinCorp's portal passed its functional tests and failed its first load test: the month-end batch slowed every customer request. The cause was not one slow component. A database volume was sized for its data, not for its I/O, and every portal replica opened its own database connections. Each was a performance decision taken by default.
This design unit walks the layered architecture of Unit 1.2 from the edge to the disk and names, at each layer, the ceiling that sets performance and the lever that moves it.
1. Where Latency Is Won or Lost
1.1 The Path and Its Ceilings
A request crosses every layer once, and the slowest layer sets the pace:
| Layer | Ceiling to know | Lever |
|---|---|---|
| Edge | ALB processing at Layer 7; NLB at Layer 4 | Keep routing logic at Layer 4 when routing performance is critical (Units 3.3, 3.4) |
| Server network | Up to 6 Gbps internal and up to 6 Gbps external per network card, at MTU 1,500 | NIC Multi-Queue; more servers behind the balancer |
| Kubernetes ingress | A LoadBalancer service is one node's static IP, with 2 Gbit/s of public throughput per node | Spread ingress IPs across several nodes |
| Hybrid link | 1 Gbps per VPN Gateway tunnel (Unit 3.6) | Several tunnels, or a dedicated line |
| Database | One primary serves all reads and writes; connections follow RAM | Pooler and cache (Section 3) |
| Storage | Per-volume IOPS that scale with size, up to a cap | Volume class and size (Section 2) |
The external figure has a qualifier: the maximum external throughput may only be achieved with a corresponding upstream of the provider at the other end. Several of these rows are single-node ceilings that no tuning removes, only scaling out. For close to 5 Gbit/s of Kubernetes ingress, for example, three nodes are needed, each with its own LoadBalancer IP address, assigned to dedicated ingress nodes by node label. The pattern is always the same: give the single node a measured budget and scale out at the layer above before the budget is spent.
1.2 The Fabric Under the Servers
Below the virtual layer, IONOS CLOUD connects the central storage systems and the internal data connections between customer servers with high-speed networks based on InfiniBand or Ethernet technology, and operates redundant networks at each location with connections up to 100 Gbps. You do not choose the fabric. You choose the driver: IONOS CLOUD recommends always using VirtIO to benefit from the full performance of InfiniBand, and IDE is intended for troubleshooting, for example when an operating system has no VirtIO drivers yet. A migrated VM left on IDE (Unit 7.4) gives away storage performance on every request.
1.3 NIC Multi-Queue
By default a network card processes traffic through one queue. NIC Multi-Queue associates multiple queues with a VM based on its number of cores, so that several CPU cores process network packets in parallel. It is off by default and turned on per VM in the DCD or the API, and turning it on restarts the VM. The queue count follows the CPU threads, so after adding cores the VM needs a restart through the DCD or the API; an operating system restart does not update the queues. It needs at least two threads and is not available for Cubes. FinCorp turns it on for the data servers and the network-heavy partner gateway pair, in a maintenance window.
2. Storage Performance Floors and Tiers
2.1 Size Buys Performance
Essential storage is static: 1,100 IOPS regular and 2,500 IOPS burst at 4 KB, independent of size. Balanced and Performance scale with the volume size, which turns capacity into an I/O decision:
| Volume size | Balanced read / write IOPS | Performance read / write IOPS |
|---|---|---|
| 100 GB | 4,000 / 3,000 | 7,500 / 5,000 |
| 300 GB | 12,000 / 9,000 | 22,500 / 15,000 |
| 600 GB and above | 24,000 / 18,000 (cap) | 45,000 / 30,000 (cap) |
The figures follow from 40 read and 30 write IOPS per GB for Balanced and 75 read and 50 write IOPS per GB for Performance, at 4 KB block size. Three rules sit around the table. IONOS CLOUD recommends at least 100 GB for Balanced and Performance volumes to get the full benefit of the SSDs. Volumes above 600 GB reach the maximum expected performance, so growing a volume further adds capacity, not IOPS. And the caps need a minimum server per volume: 2 cores and 2 GB RAM for Balanced, 4 cores and 4 GB RAM for Performance. A small server in front of a large volume never reaches the volume's ceiling.
2.2 Choosing the Tier by Price per IOPS
Balanced costs 0,07 € and Performance 0,15 € per GB per 30 days (Unit 2.4). Up to the Balanced cap, an oversized Balanced volume often buys the same IOPS for less. The self-managed reporting database on one of FinCorp's data servers needs 20,000 read IOPS:
- Balanced: 500 GB delivers 20,000 read IOPS for 35,00 € per 30 days.
- Performance: about 267 GB delivers the same for about 40,05 € per 30 days.
Performance wins when the requirement is above the Balanced caps, or when capacity must stay small. For one volume above 45,000 read IOPS, no single volume suffices: spread the data across several volumes, each with its own ceiling. Remember the asymmetry of Unit 4.2: a volume can grow but never shrink, so over-provisioning storage is cheap to start and permanent.
The managed databases follow the same physics. DBaaS for PostgreSQL recommends a storage size of 100 GB for optimal performance (Unit 5.3), so a small database still gets a 100 GB volume.
3. Throughput Levers in Front of the Database
The data tier has one writer per cluster, and standbys add availability, not read capacity (Units 5.3 and 5.5). Two levers raise throughput without a second writer.
Connection pooling. max_connections in DBaaS for PostgreSQL follows RAM and stops at 1,000 above 8 GB, with 11 reserved for internal use. Each portal replica that opens its own pool multiplies the count: twenty replicas with fifty connections reach the ceiling. The managed pooler in Transaction mode lets many client connections share fewer database connections, which is the right mode for stateless request handling.
Caching. DBaaS for In-Memory DB in front of the database serves repeated reads from memory and takes them off the primary. Its read capacity comes from its primary's cores and RAM, since secondaries are passive. Cache what is read often and changes rarely, and set expiry so that stale data has a bounded lifetime.
Together, the pooler raises how many clients one primary can serve, and the cache lowers how many queries reach it. Measure both in the PostgreSQL dashboard of Unit 7.2: sessions and transaction throughput show whether the pooler or the primary is the limit.
4. Right-Sizing Versus Over-Provisioning
Right-sizing means running close to the measured need and growing when monitoring says so. On this platform growth is cheap and fast: Live Vertical Scaling adds CPU, RAM, NICs and storage to a running Dedicated Core Server or vCPU Server. Shrinking is not: fewer cores or less RAM need a restart, and storage cannot be reduced at all (Unit 4.3). Above 240 GB of RAM, hotplugging is disabled and every RAM increase restarts the VM.
The trade-off follows from that asymmetry:
| Resource | Right-size when | Over-provision when |
|---|---|---|
| CPU and RAM | The workload can take a planned restart to shrink, and alerts fire before saturation | The VM is above 240 GB of RAM, where every increase restarts it |
| Storage | Capacity is the only need | IOPS need the size: buy the size that delivers the IOPS (Section 2) |
| Compute class | Steady, measured load | Latency-sensitive work needs exclusive cores: Dedicated Core instead of shared vCPU (Unit 4.1) |
| Replicas | VM Auto Scaling follows the load (Unit 4.3) | The scale-out reaction time is longer than the spike |
FinCorp applies the table. The portal servers stay right-sized and scale out; the payment switch database keeps Dedicated Cores and a Performance volume sized for I/O; the reporting database gets a 500 GB Balanced volume; and the ingress of fc-k8s-portal stays behind fc-alb-edge, whose target group fc-tg-k8s spreads it across the labelled nodes of both pools.
Decision Summary
| Symptom or requirement | Lever | Hard constraint |
|---|---|---|
| Database I/O is the bottleneck | Grow the Balanced or Performance volume toward its cap; give its server the minimum size | 100 GB recommended; IOPS stop growing above 600 GB |
| Need more than one volume's cap | Several volumes, data spread across them | 24,000 read IOPS Balanced, 45,000 Performance per volume |
| Too many database connections | Transaction pooler | max_connections at most 1,000, 11 reserved |
| Primary saturated by reads | In-Memory DB cache | Standbys do not serve reads |
| Network-bound VM | NIC Multi-Queue, then more servers | Up to 6 Gbps per NIC; restart to enable |
| Kubernetes ingress saturated | Ingress IPs on several labelled nodes | 2 Gbit/s per node |
| Imported VM slow on storage | Switch to VirtIO | IDE is for troubleshooting only |
Summary
Performance on IONOS CLOUD is set by a short list of known ceilings: per-NIC throughput, per-node Kubernetes ingress, per-tunnel VPN throughput, a single database primary, and per-volume IOPS that scale with size up to a cap. The levers are equally concrete: size volumes for I/O and give them a large enough server, pool connections and cache reads in front of the primary, use VirtIO and NIC Multi-Queue, and scale out before a single node's budget is spent. Grow live and shrink in a window.
Key Points:
- Balanced gives 40 read and 30 write IOPS per GB, Performance 75 and 50; both reach their caps at 600 GB and need at least 100 GB to perform.
- Balanced is often cheaper per IOPS up to its cap; Performance wins above it or when capacity must stay small.
- A NIC carries up to 6 Gbps, a Kubernetes ingress node 2 Gbit/s, a VPN tunnel 1 Gbps.
- Pooler and cache raise throughput in front of one writer; standbys do not.
- Use VirtIO, not IDE, and enable NIC Multi-Queue for network-heavy VMs.