Persistent Disk Performance Depends on vCPUs and Capacity

What Persistent Disk Performance Limits Depend On

In Google Cloud Compute Engine, the actual I/O performance (IOPS and throughput) achievable by a Persistent Disk (PD) is governed by the minimum of two independent limits: the provisioned disk capacity (which scales IOPS and throughput linearly per GiB) and the VM machine type limits (determined by the number of vCPUs assigned to the VM instance).

How It Meets the Requirements

• Target IOPS (25,000) & Throughput (800 MiBps): For modern compute instances (such as T2A, N2, or Z3), an instance with only 4 vCPUs is strictly limited to maximum write/read IOPS around 15,000–20,000 and maximum throughput around 240–400 MiBps, regardless of how large the attached disk is.
• vCPU scaling: Upgrading the VM instance to at least 8 vCPUs raises the VM-level ceiling to 25,000 IOPS and 800 MiBps write/read throughput.
• Capacity scaling: Ensuring the pd-ssd disk has sufficient GiB provisioned guarantees that the disk itself can generate the required IOPS and throughput baseline to saturate the upgraded VM interface limits.

Key Advantages of This Approach

• Balanced Resource Allocation: Eliminates the bottleneck at both the compute hypervisor layer (vCPU limit) and the block storage allocation layer (per-GiB disk limit).
• Predictable Performance: Guarantees sustained performance for latency-sensitive transactional workloads without relying on burst credits.
• Non-Disruptive Capacity Expansion: Persistent Disks can be resized online without stopping the virtual machine.

Why This Is Optimal

Upgrading both the disk capacity and the VM core count directly addresses the dual-bottleneck architecture of Compute Engine block storage, allowing the database to sustain 25,000 IOPS and 800 MiBps.