Google ACE Practice Question: Designing a GKE cluster for a workload that…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  N2 machine family with custom vCPU and memory configuration

  Why it's wrong here

  The N2 machine family does support custom vCPU and memory ratios, but its maximum memory per VM is 896 GB, and standard N2 configurations top out at 8 GB per vCPU unless you request a custom ratio that raises the memory-to-vCPU ceiling. Even with a custom config that approaches the cap, N2 instances are priced on a per-vCPU and per-GB basis that does not include the memory-optimized price discounts or the specialized high-bandwidth memory architecture of the M series. For a workload that needs more than 768 GB of RAM, N2 is not the intended choice: you would pay significantly more for the same memory footprint and still hit a hard limit below the largest M2 offerings. Thus, while N2 is flexible, it is neither cost-effective nor purpose-built for in-memory analytics at this scale.

• ✓

  Memory-optimized (M1 or M2) machine family

  Why this is correct

  The M1 and M2 machine families are the correct choice because they are explicitly engineered for high-memory workloads like SAP HANA, large in-memory databases, and real-time analytics. M1 offers m1-megamem (up to 1.75 TB) and m1-ultramem (up to 3.75 TB), while M2 adds m2-hypermem with up to 12 TB of RAM, all using Intel Xeon Scalable processors with twice the memory bandwidth of general-purpose families. These machines are priced per GB of memory at a lower rate than custom N2 or C2 configurations, making them the most cost-effective option when you need hundreds of GB to terabytes of RAM. Additionally, M-series VMs support live migration and sole-tenant nodes, giving you the same operational flexibility as other families while delivering the required 768 GB+ capacity.

• ✗

  Compute-optimized (C2) machine family

  Why it's wrong here

  The C2 (compute-optimized) family is built around high all-core turbo frequencies and a 3.8 GHz sustained clock, making it ideal for CPU-bound HPC, gaming, or media transcoding—not for massive memory footprints. C2 instances offer only up to 3.8 GB per vCPU in many configurations, and even the largest C2 standard machine tops out at 240 GB of RAM with 60 vCPUs. This memory ceiling is far below the 768 GB+ requirement, so a C2 would force you to scale out across multiple VMs, adding network overhead and licensing complexity. The C2 design philosophy prioritizes raw compute density over memory capacity, and its memory bandwidth is not tuned for the random-access patterns of large in-memory datasets.

• ✗

  Accelerator-optimized (A2) machine family

  Why it's wrong here

  The A2 accelerator-optimized family is centered around NVIDIA A100 GPUs, which are meant for GPU-accelerated compute such as deep learning training, scientific simulation, and CUDA-based workloads. While A2 instances do offer up to 96 vCPUs and 680 GB of system memory in the largest a2-ultragpu configuration, that memory is a secondary resource to the attached GPUs, and the per-GPU pricing makes it prohibitively expensive for pure in-memory analytics. If your workload does not use CUDA or GPU processing, you would be paying a large premium for unused accelerators and still not exceeding 768 GB RAM without scaling out. The A2 family also lacks the memory-optimized NUMA layout and high memory-to-vCPU ratios found in the M series, so it is a poor fit for memory cap-exceeding workloads that do not require GPU compute.

Go deeper

Related to this question