GPU Cluster Design and Tuning for ML (NVIDIA A100/H100)

GPU idling due to network bottlenecks or misconfigured drivers is a common scenario—an 8-node A100 cluster delivering only 30% utilization. The root cause is rarely the cards themselves but the environment: slow interconnect, improper NCCL parameters, or unoptimized storage. We have tuned 20+ GPU clusters over 5+ years and know how to squeeze every ampere-hour. Our approach is simple: eliminate bottlenecks across the entire chain—from drivers to job scheduler.

Why hardware is only half the story

Even top-tier GPUs won't boost performance if the rest of the system is unbalanced. The NVIDIA A100 (80GB SXM) with NVLink (600 GB/s intra-node) and H100 (80GB SXM5) with HBM3 (3.35 TB/s) are powerful, but they require matching infrastructure. Without InfiniBand and a parallel filesystem (GPFS, Lustre), you get 30–50% GPU utilization instead of 85%+. We design clusters around your workloads: for LLMs with tensor parallelism, NVLink speed is critical; for data parallelism, InfiniBand bandwidth between nodes matters most.

How to verify cluster performance after tuning

After every setup we run AllReduce tests using nccl-tests. For 8× A100, expected bandwidth is >280 GB/s at 1GB message size. If lower, we hunt for the bottleneck: NUMA affinity, driver version, switch configuration. We also launch a benchmark training run for your model (e.g., GPT-2 or BERT) and compare throughput against expectations. According to NVIDIA, proper NUMA tuning can yield up to 20% improvement.

Our GPU cluster tuning process

1. Audit current infrastructure and requirements—dataset size, model types, training frequency.
2. Design—select GPUs, number of nodes, interconnect type, filesystem.
3. Install drivers and CUDA—production versions, enable persistence mode, optimize power limits.
4. Tune NCCL—fine-tune parameters, test AllReduce bandwidth (target >280 GB/s on 8× A100).
5. Integrate with scheduler—Slurm for batch training or Kubernetes + GPU Operator for containerization.
6. Monitoring and optimization—DCGM, Prometheus, dashboards with key metrics.
7. Documentation and team training—how to launch jobs, diagnose issues.

Example: driver and CUDA installation

# Ubuntu 22.04
apt install linux-headers-$(uname -r) nvidia-driver-535
wget https://developer.download.nvidia.com/compute/cuda/12.3.0/local_installers/cuda_12.3.0_545.23.06_linux.run
sh cuda_12.3.0_545.23.06_linux.run --silent --toolkit
# cuDNN
tar -xvf cudnn-linux-x86_64-8.9.7.29_cuda12-archive.tar.xz
cp cuda/include/cudnn*.h /usr/local/cuda/include
cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
ldconfig
nvidia-smi; nvcc --version

NCCL tuning and interconnect testing

apt install libnccl2 libnccl-dev
git clone https://github.com/NVIDIA/nccl-tests
cd nccl-tests && make
./build/all_reduce_perf -b 1G -e 4G -f 2 -g 8
# Expected: 1GB ~280 GB/s, 4GB ~300 GB/s (algbw)

Interconnect choice: InfiniBand vs Ethernet

For multi-node training with tensor parallelism, InfiniBand is mandatory. Ethernet is acceptable only for small clusters (2–4 nodes) or inference.

Configuration comparison: single-node vs multi-node

NCCL tuning details

NCCL uses Tree, Ring, and NVLS algorithms. For H100, we recommend enabling NVLS (NVLink Shared) to speed up all-reduce. The parameter NCCL_ALGO=NVLS can yield 10–15% improvement. Also important is NCCL_IB_HCA to specify InfiniBand interfaces. More details can be found in the official NCCL repository.

Orchestration: Slurm or Kubernetes?

Slurm is the HPC standard, best for long batch jobs with fixed GPU count. Kubernetes + GPU Operator suits containerized, dynamic resource allocation. We help you choose and configure gang scheduling so all GPU pods launch simultaneously.

GPU Operator installation (Helm)

helm repo add nvidia https://helm.ngc.nvidia.com/nvidia
helm install gpu-operator nvidia/gpu-operator --namespace gpu-operator --create-namespace --set driver.enabled=true --set toolkit.enabled=true

Example Slurm job

#!/bin/bash
#SBATCH --nodes=4
#SBATCH --ntasks-per-node=8
#SBATCH --gres=gpu:8
#SBATCH --partition=a100
#SBATCH --time=48:00:00
srun python train.py --nproc_per_node=8 --nnodes=4

Monitoring: DCGM Exporter and metrics

helm install dcgm-exporter nvidia/dcgm-exporter

Key metrics: GPU utilisation (>85%), memory copy utilisation, NVLink bandwidth, power usage.

Common tuning mistakes

• Skipping NUMA affinity configuration—costs 10–20% performance.
• Using a single filesystem partition for both datasets and checkpoints—creates an IO bottleneck.
• Not running AllReduce tests between nodes—often only discovered in production.
• Wrong scheduler parameters (timeout, backfill)—GPUs sit idle.

Results and guarantees

After tuning, your cluster will deliver:

• GPU utilization ≥85% under standard loads.
• Scaling efficiency of 85–90% for multi-node training.
• Documented deployment and monitoring procedures.

We guarantee stable operation and provide support under a service agreement. We will estimate your project within 1–2 days. Contact us for a consultation and get a preliminary assessment. Order tuning and forget about GPU downtime.

MLOps: Infrastructure for Training, Deploying, and Monitoring ML Models

The model is trained, metrics — F1 0.94 on validation. Three months later in production, quality drops by 12%. No one knows when — there is no monitoring. It's impossible to retrain quickly — the training script is in a Jupyter notebook of a data scientist who has already left. Data for retraining is collected manually from three disparate systems. About half of the projects come to us with this pain. We build a turnkey MLOps platform: from experiment tracking to automatic deployment and data drift monitoring. We will assess your infrastructure in 1–2 weeks, and in 4–6 weeks you will get a basic MLOps core running in production. Our team has 10+ years of experience in ML infrastructure, over 50 implementations.

How does MLOps infrastructure benefit your ML projects?

Experiment Tracking and Reproducibility

Without tracking, an ML project turns into chaos: it's unclear which checkpoint is better, which hyperparameters were used, which dataset. Reproducing a result a month later is a quest.

Why is experiment tracking the foundation of reproducibility?

MLflow is an open source standard for tracking. It logs parameters, metrics, artifacts (models, graphs), and code. MLflow Model Registry is a centralized model storage with versioning and lifecycle stages (Staging → Production → Archived). Deployment via MLflow Serving or integration with external systems.

Typical initialization in code:

import mlflow

mlflow.set_experiment("fraud-detection-v2")
with mlflow.start_run():
    mlflow.log_params({"learning_rate": 3e-4, "batch_size": 64, "epochs": 10})
    mlflow.log_metric("val_f1", val_f1, step=epoch)
    mlflow.pytorch.log_model(model, "model")

This is the minimum. In production, we add logging of system metrics (GPU utilization, memory), dataset (hash, version), code (git commit hash). Weights & Biases — richer UI, collaboration features, sweep for hyperparameter optimization. MLflow — for on-premise deployment without external dependencies.

DVC (Data Version Control) — versioning of data and models on top of git. Data is stored in S3/GCS/Azure Blob, only metadata (hashes) in git. dvc repro reproduces the entire pipeline from raw data to metrics.

To ensure reproducibility of training, fix random seeds (torch.manual_seed, numpy.random.seed, random.seed) and record them in experiment metadata. Without this, debugging irregular results is painful. Log the dataset version (DVC hash) and git commit — then any experiment can be reproduced down to the byte.

Pipeline Orchestration: Kubeflow, Airflow, Prefect

A pipeline orchestrator becomes necessary when: A 100-line training script in cron is fine for simple tasks. But as soon as you have a multi-step pipeline (data loading → preprocessing → feature engineering → training → validation → deployment if quality above threshold), you need an orchestrator with retry logic, visualization, and alerts.

Kubeflow — Kubernetes-native orchestrator for ML (see Kubeflow). Each step is a Docker container. Supports parallel steps, conditional branches, artifacts between steps. Integrates with Katib (AutoML), KServe (serving), Feast (feature store).

Apache Airflow — more general DAG orchestrator. Wide ecosystem of operators (S3, Spark, DBT, Kubernetes). Easier to deploy if Airflow already exists in the company.

Prefect / Metaflow — less boilerplate. Prefect 2.x with @flow and @task decorators — quick start for small teams.

Typical training pipeline architecture on Kubeflow:

1. Data ingestion component — fetches data from S3/DB, validates schema via Great Expectations
2. Preprocessing component — transformations, normalization, train/val/test split
3. Training component — training on GPU, logging to MLflow
4. Evaluation component — metric calculation, comparison with baseline in Model Registry
5. Conditional deployment — deploy only if new model is better than current by >2% F1

Each component is a separate Docker image. Pipeline is versioned in git. Scheduled run (retraining once a week on new data) or manual.

Model Registry and Lifecycle Management

Model Registry is not just a checkpoint store. It is a centralized system that knows:

• Which model is currently in production (and with what metrics)
• History of all versions with training parameters
• Metadata: dataset, git commit, validation results
• Lifecycle stage: None → Staging → Production → Archived

MLflow Model Registry — standard. For enterprise — Vertex AI Model Registry (GCP), SageMaker Model Registry (AWS), Azure ML Model Registry.

Model promotion through stages: automatically move model to Staging after successful eval, then manual or automatic (during A/B test) promotion to Production. Rollback — switch to previous Production version in seconds.

Serving: From FastAPI to Triton Inference Server

Simple case. FastAPI + PyTorch/ONNX on one server — 80% of production ML deployments are exactly that. Sufficient for most tasks with load up to 100 req/s.

from fastapi import FastAPI
import onnxruntime as ort

app = FastAPI()
session = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

@app.post("/predict")
async def predict(request: PredictRequest):
    inputs = preprocess(request.text)
    outputs = session.run(None, {"input_ids": inputs})
    return {"label": postprocess(outputs)}

Triton Inference Server — production standard for high loads (500+ req/s). Dynamic batching, concurrent model execution, model ensemble. Supports TensorRT, ONNX, PyTorch TorchScript, TensorFlow SavedModel.

KServe — Kubernetes-native ML serving with autoscaling, canary deployments, A/B testing out of the box. Scale-to-zero for inactive models — savings on infrastructure up to 40% annually for a project with 10 models.

Monitoring: Data Drift, Model Drift, Infrastructure Metrics

Monitoring — what is usually done last and regretted first. Three levels.

Infrastructure monitoring. Latency (P50/P95/P99), throughput (req/s), error rate (4xx, 5xx), GPU/CPU utilization. Prometheus + Grafana — standard. Alert when P99 latency > threshold or error rate > 1%.

Data drift monitoring. Distribution of input data changes over time. Detect via PSI (Population Stability Index) for numerical features: PSI > 0.2 — strong drift. Chi-squared test for categorical, Kolmogorov-Smirnov test for continuous. Evidently AI — open source library with ready-made drift tests.

Model drift monitoring. If ground truth is delayed (e.g., we know conversion after a week) — monitor real metrics. If not — surrogate metrics: distribution of prediction scores, proportion of confident predictions.

Alerting. Three levels: INFO (minor drift, log it), WARNING (significant, notify team), CRITICAL (quality dropped below threshold — automatic switch to fallback model).

Why is data drift monitoring important?

Without it, you learn about model degradation only from user complaints or ringing SLA. A drift alert allows you to retrain the model in advance, before errors start causing losses. In one of our projects, PSI monitoring detected drift 2 days after a data source change — this saved the campaign.

Feature Store

Feature Store solves the training-serving skew problem. If preprocessing during training and inference is implemented in two different places — divergence is inevitable.

A Feature Store is needed when:

• Several models use the same features
• Features are computed from streaming data (real-time)
• Large team with different people on feature engineering and model training

Feast — open source Feature Store. Offline store (S3 + Parquet) for training, online store (Redis, DynamoDB) for low-latency inference. Feature definitions as code, materialization job syncs offline → online.

Tecton (commercial), Vertex AI Feature Store (GCP), SageMaker Feature Store (AWS) — managed options with less ops overhead.

CI/CD for ML

ML CI/CD is regular CI/CD plus specific ML steps.

ML-specific checks in CI:

• Reproducibility check: run training with a fixed seed, result must match
• Data validation: Great Expectations or Pandera on schema/distribution checks
• Model performance check: automatic eval on holdout, block merge if degradation > threshold
• Latency regression test: inference must meet SLA

GitOps for deployment. Merge to main → CI triggers training → eval → if passes → automatic deployment to Staging → smoke tests → manual promotion to Production or automatic upon successful canary.

Tools: GitHub Actions / GitLab CI for CI, ArgoCD for GitOps deployment on Kubernetes.

What's Included in MLOps Platform Development

We provide a full cycle of work, documentation, and team training.

Total time from audit to full MLOps platform: 3–5 months. Also possible phased launch: basic level (tracking + serving) in 4–6 weeks.

Cost is calculated individually based on data volume, number of models, and infrastructure requirements. Order an MLOps infrastructure audit — get a roadmap in 1–2 weeks. Contact us for a project assessment — we will send a preliminary estimate within 2 business days.

Note: warranty on architectural solutions — 12 months. We provide integration certificates with major cloud providers (AWS, GCP, Azure). During our work, we have not lost a single client after the first implementation — the experience of 50+ successful MLOps projects speaks for itself. Get a consultation on building an MLOps platform today.