Module 1.4: High-Performance Storage for AI

Цей контент ще не доступний вашою мовою.

Discipline Module | Complexity: [MEDIUM] | Time: 3 hours

Prerequisites

Before starting this module:

Required: Kubernetes storage fundamentals (PersistentVolumes, PersistentVolumeClaims, StorageClasses, CSI drivers)
Required: Basic understanding of ML training data pipelines (datasets, batches, data loaders)
Recommended: Module 1.1: GPU Provisioning — GPU workload basics
Recommended: Experience with object storage (S3, GCS, MinIO)

What You’ll Be Able to Do

After completing this module, you will be able to:

Design high-throughput storage architectures for AI workloads — training data, checkpoints, and model artifacts
Implement storage solutions using CSI drivers, NFS, and object storage optimized for large-scale data access
Configure caching layers that reduce data loading bottlenecks during distributed training
Evaluate storage options — local NVMe, network-attached, cloud object stores — against AI workload I/O patterns

Why This Module Matters

You spent months building a beautiful GPU platform. The GPUs are provisioned, shared efficiently, connected by InfiniBand. Then your ML team starts training and reports this:

“Our 8-GPU job only uses 40% GPU utilization. The GPUs are waiting for data.”

This is the IO bottleneck — the most common and most underestimated performance killer in AI infrastructure. Your $300,000 DGX node is sitting idle 60% of the time because the storage system cannot feed data to the GPUs fast enough.

The numbers are stark:

Component	Throughput	Latency
GPU compute (A100 BF16)	312 TFLOPS	nanoseconds
GPU memory (HBM3)	2 TB/s	nanoseconds
NVMe SSD (local)	7 GB/s	10-100 μs
Network storage (CephFS)	1-5 GB/s	0.5-5 ms
Object storage (S3)	100-500 MB/s	10-100 ms

There is a 1,000x gap between GPU memory speed and network storage speed. Bridging this gap is what this module is about.

The IO Bottleneck in ML Workloads

Where IO Happens

Every training step involves IO at multiple stages:

┌─────────────────────────────────────────────────────────────┐
│                      Training Loop                           │
│                                                               │
│  1. Load batch          2. Transfer to GPU    3. Compute     │
│  ┌──────────────┐      ┌──────────────┐     ┌────────────┐ │
│  │ Read from    │      │ CPU RAM →    │     │ Forward +  │ │
│  │ storage      │ ──→  │ GPU VRAM     │ ──→ │ Backward   │ │
│  │ (IO bound)   │      │ (PCIe bound) │     │ (compute)  │ │
│  └──────────────┘      └──────────────┘     └────────────┘ │
│  100ms - 5s              1-10ms               10-100ms       │
│                                                               │
│  ← This dominates when storage is slow                       │
└─────────────────────────────────────────────────────────────┘

Workload IO Profiles

Different ML workloads have radically different IO characteristics:

Workload	Data Size	Access Pattern	Read Size	Throughput Need
Image classification (ImageNet)	150 GB	Random, small files	100-500 KB	2-5 GB/s
Object detection (COCO)	20 GB	Random, medium files	200 KB - 5 MB	1-3 GB/s
NLP pre-training (C4)	800 GB	Sequential, large files	1-100 MB	5-20 GB/s
Video training	5-50 TB	Sequential, very large	50-500 MB	10-50 GB/s
LLM fine-tuning (tokenized)	10-100 GB	Sequential	1-10 MB	1-5 GB/s
Checkpoint save	1-50 GB per save	Sequential write	Full model	5-20 GB/s burst

The key insight: image training does millions of small random reads (hard for network storage), while LLM training does large sequential reads (easier to cache).

Profiling IO Bottlenecks

Before optimizing, measure. Run your training job with GPU utilization monitoring:

# Monitor GPU utilization during training
# If GPU util is < 80% and you're not memory-bound, you're IO-bound

# Quick check: watch nvidia-smi during training
kubectl exec -it training-pod -- watch -n 1 'nvidia-smi --query-gpu=utilization.gpu,utilization.memory --format=csv,noheader'

# Better: check PyTorch DataLoader timing
# Add this to your training script:
# import time
# for batch in dataloader:
#     load_end = time.time()
#     print(f"Data load: {load_end - load_start:.3f}s")
#     # ... training step ...
#     load_start = time.time()

Storage Tiers for AI

The Storage Pyramid

AI workloads need a multi-tier storage architecture:

                    ┌─────────────┐
                    │   GPU VRAM   │  2 TB/s, μs latency
                    │  (training)  │  Managed by framework
                    ├─────────────┤
                 ┌──┤  Local NVMe  │  3-14 GB/s, 10-100 μs
                 │  │  (hot cache) │  TopoLVM, OpenEBS LVM
                 │  ├─────────────┤
              ┌──┤  │ Distributed  │  1-10 GB/s, 0.5-5 ms
              │  │  │  FS (warm)   │  CephFS, GlusterFS, JuiceFS
              │  │  ├─────────────┤
           ┌──┤  │  │   Object     │  100 MB-5 GB/s, 10-100 ms
           │  │  │  │  Storage     │  S3, GCS, MinIO
           │  │  │  │  (cold)      │
           │  │  │  ├─────────────┤
           │  │  │  │   Tape/      │  Archival
           │  │  │  │  Archive     │  Glacier, Coldline
           │  │  │  └─────────────┘
           │  │  │
      Cost │  │  │ Speed
       ▼   │  │  │   ▲
       $   $$  $$$  $$$$

The platform team’s job is to build infrastructure that automatically moves data between tiers based on access patterns.

Local NVMe Caching

Why Local Storage Matters

A modern NVMe SSD delivers 3-7 GB/s sequential read and 500K-1M IOPS random read. This is 10-50x faster than network storage for the random small-file reads that image training demands.

The strategy: keep the active dataset (or a cache of it) on local NVMe while the canonical copy lives in object storage.

TopoLVM: Topology-Aware Local Volumes

TopoLVM is a CSI driver that provisions PersistentVolumes from local LVM volume groups, with topology awareness — it ensures Pods are scheduled on nodes that have available local storage.

# Install TopoLVM
helm repo add topolvm https://topolvm.github.io/topolvm
helm repo update

helm install topolvm topolvm/topolvm \
  --namespace topolvm-system \
  --create-namespace \
  --set controller.replicaCount=2 \
  --set node.volumeGroup.name=nvme-vg    # LVM VG name on each node

Create a StorageClass:

apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: nvme-local
provisioner: topolvm.io
parameters:
  topolvm.io/device-class: nvme       # Maps to a device class in TopoLVM config
volumeBindingMode: WaitForFirstConsumer # Delay binding until Pod is scheduled
allowVolumeExpansion: true
reclaimPolicy: Delete

Use in a training Pod:

apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: training-cache
  namespace: ml-training
spec:
  storageClassName: nvme-local
  accessModes:
    - ReadWriteOnce
  resources:
    requests:
      storage: 500Gi
---
apiVersion: v1
kind: Pod
metadata:
  name: image-trainer
  namespace: ml-training
spec:
  containers:
    - name: trainer
      image: nvcr.io/nvidia/pytorch:24.09-py3
      volumeMounts:
        - name: cache
          mountPath: /data/cache
        - name: dataset
          mountPath: /data/s3     # S3 FUSE mount or pre-downloaded
      resources:
        limits:
          nvidia.com/gpu: 4
  volumes:
    - name: cache
      persistentVolumeClaim:
        claimName: training-cache
    - name: dataset
      persistentVolumeClaim:
        claimName: imagenet-s3

OpenEBS LVM Local PV

OpenEBS provides a simpler alternative for local NVMe provisioning:

# Install OpenEBS LVM LocalPV
helm repo add openebs https://openebs.github.io/openebs
helm repo update

helm install openebs openebs/openebs \
  --namespace openebs \
  --create-namespace \
  --set lvm-localpv.enabled=true \
  --set engines.replicated.mayastor.enabled=false

apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: openebs-nvme
provisioner: local.csi.openebs.io
parameters:
  storage: "lvm"
  vgPattern: "nvme-vg"          # LVM volume group pattern
  fsType: "xfs"                  # XFS recommended for large files
volumeBindingMode: WaitForFirstConsumer

Init Container Pattern for Data Staging

A common pattern: use an init container to stage data from object storage to local NVMe before training begins:

apiVersion: batch/v1
kind: Job
metadata:
  name: training-with-staging
  namespace: ml-training
spec:
  template:
    spec:
      initContainers:
        - name: stage-data
          image: amazon/aws-cli:2.17
          command: ["sh", "-c"]
          args:
            - |
              echo "Staging dataset from S3..."
              start=$(date +%s)
              aws s3 sync s3://my-datasets/imagenet/ /data/cache/imagenet/ \
                --no-sign-request --quiet
              end=$(date +%s)
              size=$(du -sh /data/cache/imagenet/ | cut -f1)
              echo "Staged $size in $((end-start)) seconds"
          volumeMounts:
            - name: cache
              mountPath: /data/cache
          resources:
            requests:
              cpu: "4"
              memory: 8Gi
      containers:
        - name: trainer
          image: nvcr.io/nvidia/pytorch:24.09-py3
          command: ["torchrun", "--nproc_per_node=4", "train.py", "--data_dir=/data/cache/imagenet"]
          volumeMounts:
            - name: cache
              mountPath: /data/cache
          resources:
            limits:
              nvidia.com/gpu: 4
      volumes:
        - name: cache
          persistentVolumeClaim:
            claimName: training-cache
      restartPolicy: OnFailure

Distributed Filesystems

CephFS

Ceph is the most widely deployed distributed storage system in Kubernetes. CephFS provides a POSIX-compatible filesystem backed by the Ceph cluster.

Strengths for AI:

POSIX semantics (training frameworks expect filesystem APIs)
Scalable metadata server (can handle millions of small files)
Multi-reader access (ReadWriteMany) for data-parallel training
Integrated with Rook for Kubernetes-native deployment

Weaknesses for AI:

Latency: 0.5-5ms per operation (100-1000x slower than local NVMe)
Throughput ceiling: limited by network and OSD count
Small file performance: poor for millions of tiny files (image datasets)

# Rook-Ceph CephFS StorageClass
apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: cephfs-ai
provisioner: rook-ceph.cephfs.csi.ceph.com
parameters:
  clusterID: rook-ceph
  fsName: ai-filesystem
  pool: ai-data-pool
  csi.storage.k8s.io/provisioner-secret-name: rook-csi-cephfs-provisioner
  csi.storage.k8s.io/provisioner-secret-namespace: rook-ceph
  csi.storage.k8s.io/node-stage-secret-name: rook-csi-cephfs-node
  csi.storage.k8s.io/node-stage-secret-namespace: rook-ceph
mountOptions:
  - noatime                    # Disable access time updates (major perf win)
  - nodiratime

JuiceFS

JuiceFS is a cloud-native distributed filesystem purpose-built for the gap between object storage and high-performance compute. It separates metadata (stored in Redis, PostgreSQL, or TiKV) from data (stored in any object storage).

┌─────────────────────────────────────────────────────┐
│                    JuiceFS Architecture               │
│                                                       │
│  ┌────────────┐    ┌──────────────┐    ┌──────────┐ │
│  │  POSIX     │    │   Metadata   │    │  Object  │ │
│  │  Client    │──→ │   Engine     │    │  Storage │ │
│  │ (FUSE/CSI) │    │ (Redis/PG)   │    │  (S3)    │ │
│  │            │    └──────────────┘    │          │ │
│  │            │──────────────────────→ │          │ │
│  └────────────┘     Data path          └──────────┘ │
│       │                                              │
│       ▼                                              │
│  ┌────────────┐                                     │
│  │ Local Cache │  ← NVMe or RAM                    │
│  │ (read/write)│                                    │
│  └────────────┘                                     │
└─────────────────────────────────────────────────────┘

Why JuiceFS excels for AI:

Transparent caching: Reads are cached on local NVMe. Second read of same file is at NVMe speed.
POSIX compatible: Drop-in replacement for local filesystem in training scripts.
Any object store backend: S3, GCS, Azure Blob, MinIO — your data stays where it is.
Metadata engine flexibility: Redis for speed, PostgreSQL for durability, TiKV for scale.
Kubernetes-native: CSI driver with dynamic provisioning.

Installing JuiceFS CSI Driver

# Install JuiceFS CSI Driver
helm repo add juicefs https://juicedata.github.io/charts/
helm repo update

helm install juicefs-csi juicefs/juicefs-csi-driver \
  --namespace kube-system \
  --set storageClasses[0].name=juicefs-sc \
  --set storageClasses[0].enabled=true \
  --set storageClasses[0].backend.name=ai-data \
  --set storageClasses[0].backend.metaurl=redis://:password@redis-master:6379/1 \
  --set storageClasses[0].backend.storage=s3 \
  --set storageClasses[0].backend.bucket=s3://my-ai-datasets \
  --set storageClasses[0].backend.accessKey=AKIAIOSFODNN7EXAMPLE \
  --set storageClasses[0].backend.secretKey=wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY \
  --set storageClasses[0].cachePVC=juicefs-cache

JuiceFS StorageClass with Caching

apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: juicefs-ai
provisioner: csi.juicefs.com
parameters:
  csi.storage.k8s.io/provisioner-secret-name: juicefs-secret
  csi.storage.k8s.io/provisioner-secret-namespace: kube-system
  juicefs/mount-options: |
    cache-dir=/var/jfsCache
    cache-size=512000          # 500GB local cache
    buffer-size=1024           # 1GB read-ahead buffer
    prefetch=3                 # Prefetch 3 blocks ahead
    max-uploads=40             # Parallel upload threads
    metacache-expire=300       # Metadata cache TTL (seconds)
    open-cache=300             # Open file handle cache
reclaimPolicy: Retain
volumeBindingMode: Immediate

Dataset Caching: Fluid and Alluxio

The Caching Problem

Consider this scenario: 50 ML engineers share a 500GB ImageNet dataset stored in S3. Without caching:

Engineer 1 training job: Downloads 500GB from S3 → 30 min
Engineer 2 training job: Downloads 500GB from S3 → 30 min
...
Engineer 50 training job: Downloads 500GB from S3 → 30 min

Total: 25 TB downloaded, 25 hours of wait time, $50+ in S3 egress

With a caching layer:

Engineer 1 training job: Downloads 500GB from S3 → 30 min (cold)
Engineer 2 training job: Reads from cache → 2 min (warm)
...
Engineer 50 training job: Reads from cache → 2 min (warm)

Total: 500GB downloaded, 2 hours total wait, $1 in S3 egress

Fluid: Kubernetes-Native Dataset Orchestration

Fluid is a CNCF sandbox project that brings dataset-aware scheduling to Kubernetes. It manages datasets as first-class resources and uses cache engines (Alluxio, JuiceFS, JindoFS) under the hood.

# Install Fluid
helm repo add fluid https://fluid-cloudnative.github.io/charts
helm repo update

helm install fluid fluid/fluid \
  --namespace fluid-system \
  --create-namespace

Define a dataset and its caching runtime:

apiVersion: data.fluid.io/v1alpha1
kind: Dataset
metadata:
  name: imagenet
  namespace: ml-training
spec:
  mounts:
    - mountPoint: s3://my-datasets/imagenet/
      name: imagenet
      options:
        aws.accessKeyId: AKIAIOSFODNN7EXAMPLE
        aws.region: us-east-1
      encryptOptions:
        - name: aws.secretAccessKey
          valueFrom:
            secretKeyRef:
              name: s3-credentials
              key: secretAccessKey
---
apiVersion: data.fluid.io/v1alpha1
kind: AlluxioRuntime
metadata:
  name: imagenet
  namespace: ml-training
spec:
  replicas: 3                    # 3 cache workers
  tieredstore:
    levels:
      - mediumtype: SSD
        path: /dev/shm,/var/cache/alluxio
        quota: 100Gi,400Gi       # 100GB RAM + 400GB SSD cache per worker
        high: "0.95"
        low: "0.7"
  fuse:
    args:
      - fuse
      - --attr-timeout=7200s
      - --entry-timeout=7200s
    cleanPolicy: OnDemand
  properties:
    alluxio.user.metadata.cache.enabled: "true"
    alluxio.user.metadata.cache.expireTime: "2day"
    alluxio.user.streaming.data.timeout: "300sec"

Use the dataset in a training Pod:

apiVersion: batch/v1
kind: Job
metadata:
  name: imagenet-training
  namespace: ml-training
spec:
  template:
    spec:
      containers:
        - name: trainer
          image: nvcr.io/nvidia/pytorch:24.09-py3
          command: ["python", "train.py", "--data_dir=/data/imagenet"]
          volumeMounts:
            - name: imagenet
              mountPath: /data/imagenet
              readOnly: true
          resources:
            limits:
              nvidia.com/gpu: 4
      volumes:
        - name: imagenet
          persistentVolumeClaim:
            claimName: imagenet    # Automatically created by Fluid
      restartPolicy: OnFailure

Fluid’s Data-Aware Scheduling

Fluid tracks where cached data resides and preferentially schedules Pods on nodes that already have the data cached:

# Fluid automatically injects scheduling hints
# Pods using the 'imagenet' dataset prefer nodes where Alluxio workers
# have already cached ImageNet data

# You can also trigger pre-warming:
apiVersion: data.fluid.io/v1alpha1
kind: DataLoad
metadata:
  name: imagenet-warmup
  namespace: ml-training
spec:
  dataset:
    name: imagenet
    namespace: ml-training
  loadMetadata: true
  target:
    - path: /
      replicas: 2    # Cache 2 copies for fault tolerance

Alluxio Standalone (without Fluid)

For teams that want more control, Alluxio can be deployed independently:

helm repo add alluxio https://alluxio-charts.storage.googleapis.com/openSource
helm repo update

helm install alluxio alluxio/alluxio \
  --namespace alluxio \
  --create-namespace \
  --set master.count=1 \
  --set worker.count=3 \
  --set worker.resources.limits.memory=32Gi \
  --set tieredStore.levels[0].level=0 \
  --set tieredStore.levels[0].mediumtype=MEM \
  --set tieredStore.levels[0].path=/dev/shm \
  --set tieredStore.levels[0].quota=16Gi \
  --set tieredStore.levels[1].level=1 \
  --set tieredStore.levels[1].mediumtype=SSD \
  --set tieredStore.levels[1].path=/mnt/nvme/alluxio \
  --set tieredStore.levels[1].quota=500Gi \
  --set properties."alluxio.underfs.s3.region"=us-east-1

Checkpoint Storage

Why Checkpoint IO Matters

During training, checkpoints must be saved periodically. A checkpoint for a 70B parameter model is:

Model parameters:   70B × 2 bytes (BF16)  = 140 GB
Optimizer state:    70B × 8 bytes (Adam)   = 560 GB
Total:              ~700 GB per checkpoint

If your storage can write at 2 GB/s, saving one checkpoint takes 350 seconds — almost 6 minutes. During this time, GPUs are either idle (synchronous checkpoint) or must continue while carefully not overwriting the in-flight checkpoint (asynchronous).

Strategies for Fast Checkpointing

Synchronous (simple, slow):

# Training pauses during save
torch.save(model.state_dict(), "/checkpoints/latest.pt")
# 6 minutes of idle GPUs

Asynchronous with background thread:

import threading

def save_async(state_dict, path):
    torch.save(state_dict, path)

# Clone state dict to CPU, then save in background
state_dict_cpu = {k: v.cpu().clone() for k, v in model.state_dict().items()}
thread = threading.Thread(target=save_async, args=(state_dict_cpu, path))
thread.start()
# Training continues immediately

Sharded checkpoints (PyTorch FSDP / DeepSpeed):

# Each GPU saves its own shard in parallel
# 8 GPUs writing 87.5 GB each at 2 GB/s = 44 seconds (vs 350 seconds serial)
from torch.distributed.checkpoint import save
save(model.state_dict(), checkpoint_id=f"/checkpoints/step_{step}")

Recommended Checkpoint Storage

Storage Type	Write Speed	Best For
Local NVMe (TopoLVM)	5-7 GB/s	Fastest saves; risk of data loss on node failure
CephFS / GlusterFS (RWX)	1-5 GB/s	Shared access, multi-node distributed saves
JuiceFS (NVMe cache + S3)	3-7 GB/s local, async to S3	Best of both: fast writes, durable storage
NFS	0.5-2 GB/s	Simple, widely available; potential bottleneck

Try This: Measure Your Storage Performance

Run this inside a Pod on your cluster to understand your storage baseline:

# Create a test pod with your storage class
cat <<'EOF' | kubectl apply -f -
apiVersion: v1
kind: Pod
metadata:
  name: storage-bench
  namespace: default
spec:
  containers:
    - name: bench
      image: ubuntu:22.04
      command: ["sleep", "infinity"]
      volumeMounts:
        - name: test-vol
          mountPath: /data
      resources:
        requests:
          cpu: "4"
          memory: 8Gi
  volumes:
    - name: test-vol
      persistentVolumeClaim:
        claimName: bench-pvc
EOF

# Inside the pod:
kubectl exec -it storage-bench -- bash

apt-get update && apt-get install -y fio

# Sequential read (simulates loading a large dataset)
fio --name=seq-read --rw=read --bs=1M --size=10G \
  --numjobs=4 --direct=1 --directory=/data \
  --runtime=60 --time_based --group_reporting

# Random read (simulates image dataset loading)
fio --name=rand-read --rw=randread --bs=256K --size=10G \
  --numjobs=8 --direct=1 --directory=/data \
  --runtime=60 --time_based --group_reporting \
  --iodepth=32

# Sequential write (simulates checkpoint saves)
fio --name=seq-write --rw=write --bs=1M --size=10G \
  --numjobs=4 --direct=1 --directory=/data \
  --runtime=60 --time_based --group_reporting

Did You Know?

ImageNet, the dataset that launched the deep learning revolution, contains 14 million images totaling about 150GB. But the images are tiny JPEG files (average ~10KB each). This means loading ImageNet requires 14 million random reads — a worst case for any storage system. This is why ImageNet training was one of the first workloads to expose storage bottlenecks in GPU clusters.
The concept of “data gravity” is literal in AI infrastructure. Moving a 10TB dataset across the internet takes hours to days, but computing on it takes seconds to minutes. This is why cloud providers offer “data import” services where they physically ship hard drives. Google’s Transfer Appliance can hold 1 PB and ships via FedEx — sometimes the highest-bandwidth network is a truck full of disks.
Meta reported that during Llama 3 training, their storage system served 240 PB of data over the 54-day run — roughly 4.4 PB per day, or 51 GB per second sustained. This required a custom distributed filesystem (Tectonic) because no off-the-shelf system could handle this throughput at this scale.

War Story: The Cache That Saved $200K

A medical imaging startup trained models on a 2TB dataset of CT scans stored in Google Cloud Storage (GCS). They had 20 GPU nodes, each running training jobs that loaded the full dataset.

Before caching: Each job downloaded 2TB from GCS at ~500 MB/s = 67 minutes startup time. With 20 nodes running 3 jobs/day each, they downloaded 120 TB/day from GCS.

GCS egress cost: $0.12/GB × 120,000 GB/day = $14,400/day = $432,000/month
GPU idle time during downloads: 20 nodes × 3 jobs × 67 min = 67 GPU-hours/day wasted

After JuiceFS with NVMe cache: First job on each node downloads from GCS (cold cache). Subsequent jobs read from local NVMe cache at 5 GB/s = 7 minutes.

GCS egress: 20 nodes × 2TB × 1 download/week = 40TB/week = $19,200/month
GPU idle time: negligible (7 min cached vs 67 min uncached)

Monthly savings: $432,000 - $19,200 = $412,800/month. The caching infrastructure (JuiceFS + NVMe on each node) cost $3,000/month.

Lesson: In AI infrastructure, the most impactful optimization is often the simplest: cache the dataset close to the GPUs.

Common Mistakes

Mistake	Problem	Solution
Using S3 FUSE mounts for training	FUSE adds 2-10x latency overhead per IO operation	Use JuiceFS or Alluxio with local NVMe cache; or download to local disk first
Network storage for ImageNet-style training	Millions of small random reads kill network storage	Cache dataset on local NVMe; or use WebDataset/TFRecord for sequential access
Synchronous checkpoints on slow storage	GPUs idle for minutes during each checkpoint save	Use async checkpointing or sharded distributed checkpoints
No `noatime` mount option	Every file read triggers a metadata write (access time update)	Always mount with `noatime,nodiratime` for training volumes
RWO volumes for multi-node training	ReadWriteOnce cannot be mounted on multiple nodes	Use RWX storage (CephFS, NFS, JuiceFS) or local cache per node
Ignoring storage class `volumeBindingMode`	PVC binds to wrong node before Pod is scheduled	Always use `WaitForFirstConsumer` for local storage
Not pre-warming cache before training	First epoch runs at cold-cache speed, skewing benchmark results	Use Fluid’s `DataLoad` CRD or an init container to warm cache
Using ext4 for large files	ext4 fragments large sequential writes	Use XFS for datasets and checkpoint volumes; it handles large files better

Quiz: Check Your Understanding

Question 1

Why is storing a dataset of 14 million small images on S3 problematic for GPU training?

Show Answer

Three compounding issues:

Latency: Each S3 GET request has 10-100ms latency. With 14M random reads, the cumulative latency is enormous even with parallelism.
Request overhead: S3 is optimized for throughput on large objects, not IOPS on small objects. Each 10KB image requires a full HTTP GET with TLS handshake, authentication, etc. The protocol overhead exceeds the data size.
No prefetching: S3 has no concept of “read the next file” — each read is independent. Local filesystems and caching layers can prefetch adjacent files, but S3 cannot.

Solutions: (a) Convert to sequential formats like WebDataset or TFRecord, (b) cache on local NVMe with JuiceFS/Alluxio, or (c) download the entire dataset to local storage before training.

Question 2

Explain the difference between JuiceFS and Alluxio as caching solutions for AI workloads.

Show Answer

JuiceFS is a full POSIX filesystem with caching:

Separates metadata (Redis/PostgreSQL) from data (any object store)
Provides a complete filesystem (create, write, read, delete, rename)
Client-side caching on local NVMe
Can be used as primary storage, not just a cache layer
Simpler architecture (no separate master/worker topology)

Alluxio is a caching middleware layer:

Sits between compute and existing storage systems
Master/worker architecture with distributed cache
Does not store data itself — always backed by an “under filesystem” (S3, HDFS)
Richer data management: pinning, TTL, replication policies
More complex to operate but more features for large-scale deployments

When to choose which:

JuiceFS: when you need a filesystem that also caches, or when simplicity matters
Alluxio (via Fluid): when you need dataset-aware scheduling, multi-tier caching, or already have complex data infrastructure

Question 3

A training job saves a 700GB checkpoint every 1000 steps. Steps take 2 seconds each. Checkpoint save takes 350 seconds on the current storage. What percentage of GPU time is wasted on checkpointing, and how would you reduce it?

Show Answer

Waste calculation:

Steps between checkpoints: 1000 × 2s = 2000s of training
Checkpoint time: 350s
Waste: 350 / (2000 + 350) = 14.9% of total time

Reduction strategies:

Sharded checkpoints: 8 GPUs each save 87.5GB in parallel → 44s instead of 350s → waste drops to 2.1%
Async checkpointing: Clone state dict to CPU RAM (takes ~10s), save in background thread while training continues → waste drops to ~0.5%
Faster storage: Local NVMe at 7 GB/s → 100s → waste drops to 4.8%. Combined with sharding: 12.5s → 0.6%
Less frequent checkpoints: Every 2000 steps instead of 1000 → halves the waste, but doubles max lost work on failure

The best approach combines sharded + async: each GPU clones its shard to CPU, then background threads write to fast storage while training continues. This achieves <1% waste.

Question 4

What does Fluid’s data-aware scheduling do that a regular PVC does not?

Show Answer

A regular PVC binds to a volume and any node that can access that volume can run the Pod. It has no awareness of data locality — a Pod might run on a node that has no cached data, causing a cold-cache start.

Fluid’s data-aware scheduling:

Tracks cache location: Knows which nodes’ Alluxio workers have cached which datasets
Prefers warm nodes: Injects scheduling hints (nodeAffinity) so Pods prefer nodes where their dataset is already cached
Enables pre-warming: DataLoad CRD can prefill cache before Pods start
Manages cache lifecycle: Evicts stale data, rebalances across workers, manages multi-tier (RAM + SSD) caching
Abstracts the cache engine: User sees a PVC; Fluid manages Alluxio/JuiceFS/JindoFS underneath

Hands-On Exercise: JuiceFS Cache Over S3 with Latency Measurement

Objective

Deploy JuiceFS with a local NVMe cache backed by S3-compatible object storage, load a dataset, and measure the difference between cold-cache and warm-cache read performance.

Environment

Kubernetes cluster with at least one node
MinIO or any S3-compatible storage (we will deploy MinIO for this exercise)
A node with local storage available (emptyDir is acceptable for the exercise)

Step 1: Deploy MinIO (S3-compatible Object Store)

# Install MinIO for local S3-compatible storage
kubectl create namespace storage

cat <<'EOF' | kubectl apply -f -
apiVersion: apps/v1
kind: Deployment
metadata:
  name: minio
  namespace: storage
spec:
  replicas: 1
  selector:
    matchLabels:
      app: minio
  template:
    metadata:
      labels:
        app: minio
    spec:
      containers:
        - name: minio
          image: minio/minio:RELEASE.2024-10-13T13-34-11Z
          args: ["server", "/data", "--console-address", ":9001"]
          env:
            - name: MINIO_ROOT_USER
              value: minioadmin
            - name: MINIO_ROOT_PASSWORD
              value: minioadmin123
          ports:
            - containerPort: 9000
            - containerPort: 9001
          volumeMounts:
            - name: data
              mountPath: /data
      volumes:
        - name: data
          emptyDir:
            sizeLimit: 20Gi
---
apiVersion: v1
kind: Service
metadata:
  name: minio
  namespace: storage
spec:
  ports:
    - port: 9000
      targetPort: 9000
      name: api
    - port: 9001
      targetPort: 9001
      name: console
  selector:
    app: minio
EOF

kubectl -n storage wait --for=condition=Ready pod -l app=minio --timeout=120s

Step 2: Create a Test Dataset in MinIO

# Create a bucket and upload test data
kubectl -n storage run mc --rm -it --restart=Never \
  --image=minio/mc:RELEASE.2024-10-08T09-37-26Z -- bash -c '
  mc alias set local http://minio:9000 minioadmin minioadmin123
  mc mb local/ai-datasets

  # Create a 1GB test dataset (256 files of 4MB each)
  for i in $(seq 1 256); do
    dd if=/dev/urandom of=/tmp/data_${i}.bin bs=4M count=1 2>/dev/null
    mc cp /tmp/data_${i}.bin local/ai-datasets/training/
    rm /tmp/data_${i}.bin
  done

  echo "Dataset created:"
  mc ls local/ai-datasets/training/ | wc -l
  mc du local/ai-datasets/
'

Step 3: Deploy Redis (JuiceFS Metadata Engine)

cat <<'EOF' | kubectl apply -f -
apiVersion: apps/v1
kind: Deployment
metadata:
  name: redis
  namespace: storage
spec:
  replicas: 1
  selector:
    matchLabels:
      app: redis
  template:
    metadata:
      labels:
        app: redis
    spec:
      containers:
        - name: redis
          image: redis:7-alpine
          ports:
            - containerPort: 6379
---
apiVersion: v1
kind: Service
metadata:
  name: redis
  namespace: storage
spec:
  ports:
    - port: 6379
  selector:
    app: redis
EOF

Step 4: Install JuiceFS CSI Driver

helm repo add juicefs https://juicedata.github.io/charts/
helm repo update

# Create the JuiceFS secret
cat <<'EOF' | kubectl apply -f -
apiVersion: v1
kind: Secret
metadata:
  name: juicefs-secret
  namespace: storage
type: Opaque
stringData:
  name: ai-data
  metaurl: redis://redis.storage.svc:6379/1
  storage: s3
  bucket: http://minio.storage.svc:9000/ai-datasets
  access-key: minioadmin
  secret-key: minioadmin123
EOF

# Install the CSI driver
helm install juicefs-csi juicefs/juicefs-csi-driver \
  --namespace kube-system \
  --version v0.24.8

Step 5: Create JuiceFS StorageClass and PVC

cat <<'EOF' | kubectl apply -f -
apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
  name: juicefs-cache
provisioner: csi.juicefs.com
parameters:
  csi.storage.k8s.io/provisioner-secret-name: juicefs-secret
  csi.storage.k8s.io/provisioner-secret-namespace: storage
  csi.storage.k8s.io/node-publish-secret-name: juicefs-secret
  csi.storage.k8s.io/node-publish-secret-namespace: storage
  juicefs/mount-options: "cache-size=10240,buffer-size=512"
reclaimPolicy: Delete
volumeBindingMode: Immediate
---
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: juicefs-data
  namespace: storage
spec:
  storageClassName: juicefs-cache
  accessModes:
    - ReadWriteMany
  resources:
    requests:
      storage: 20Gi
EOF

Step 6: Measure Cold vs Warm Cache Performance

cat <<'BENCHEOF' | kubectl apply -f -
apiVersion: v1
kind: Pod
metadata:
  name: cache-benchmark
  namespace: storage
spec:
  containers:
    - name: bench
      image: ubuntu:22.04
      command: ["bash", "-c"]
      args:
        - |
          apt-get update -qq && apt-get install -y -qq time bc 2>/dev/null

          echo "=== COLD CACHE READ (first access, data fetched from S3) ==="
          sync; echo 3 > /proc/sys/vm/drop_caches 2>/dev/null || true

          cold_start=$(date +%s%N)
          total_bytes=0
          for f in /data/training/data_*.bin; do
            cat "$f" > /dev/null 2>&1
            total_bytes=$((total_bytes + $(stat -c%s "$f" 2>/dev/null || echo 0)))
          done
          cold_end=$(date +%s%N)

          cold_ms=$(( (cold_end - cold_start) / 1000000 ))
          cold_mbps=$(echo "scale=2; $total_bytes / 1048576 / ($cold_ms / 1000)" | bc 2>/dev/null || echo "N/A")
          echo "Cold read: ${cold_ms}ms for $(echo "$total_bytes / 1048576" | bc)MB = ${cold_mbps} MB/s"
          echo ""

          echo "=== WARM CACHE READ (second access, data from local cache) ==="
          warm_start=$(date +%s%N)
          for f in /data/training/data_*.bin; do
            cat "$f" > /dev/null 2>&1
          done
          warm_end=$(date +%s%N)

          warm_ms=$(( (warm_end - warm_start) / 1000000 ))
          warm_mbps=$(echo "scale=2; $total_bytes / 1048576 / ($warm_ms / 1000)" | bc 2>/dev/null || echo "N/A")
          echo "Warm read: ${warm_ms}ms for $(echo "$total_bytes / 1048576" | bc)MB = ${warm_mbps} MB/s"
          echo ""

          speedup=$(echo "scale=1; $cold_ms / $warm_ms" | bc 2>/dev/null || echo "N/A")
          echo "=== RESULT: Warm cache is ${speedup}x faster than cold ==="

          sleep 3600
      volumeMounts:
        - name: data
          mountPath: /data
      resources:
        requests:
          cpu: "2"
          memory: 4Gi
  volumes:
    - name: data
      persistentVolumeClaim:
        claimName: juicefs-data
  restartPolicy: Never
BENCHEOF

# Wait and check results
kubectl -n storage wait --for=condition=Ready pod/cache-benchmark --timeout=300s
sleep 60
kubectl -n storage logs cache-benchmark

Step 7: Cleanup

kubectl delete namespace storage

Success Criteria

You have completed this exercise when:

MinIO is running and contains a 1GB test dataset (256 files)
Redis metadata engine is running
JuiceFS CSI driver is installed and StorageClass is created
PVC is bound and mountable
Cold cache read time is measured (expect: 30-120 seconds for 1GB)
Warm cache read time is measured (expect: 2-10 seconds for 1GB)
Warm cache read is at least 3x faster than cold cache read
You can explain why the speedup occurs (local NVMe/memory vs S3 network round-trip)

Key Takeaways

IO is the most common bottleneck in GPU training — if GPU utilization is below 80%, investigate storage first
Local NVMe is 10-50x faster than network storage for the random small-file reads that image training demands
JuiceFS bridges the gap between object storage (cheap, durable, slow) and local NVMe (fast, ephemeral)
Fluid/Alluxio add data-aware scheduling — Pods prefer nodes that already have their dataset cached
Checkpointing must be fast — synchronous saves to slow storage can waste 15%+ of GPU time; use sharded async checkpoints
Mount with noatime — a one-line fix that eliminates unnecessary metadata writes on every file read
Measure before optimizing — use fio to benchmark your storage and nvidia-smi to correlate with GPU utilization
The init container pattern for data staging is simple, reliable, and often sufficient for datasets under 1TB

Summary

Storage is the hidden bottleneck that prevents expensive GPUs from reaching their potential. A multi-tier approach — local NVMe for hot data, distributed filesystem for warm data, object storage for cold data — combined with intelligent caching (JuiceFS, Fluid/Alluxio) bridges the 1,000x performance gap between GPU memory and network storage. Fast checkpoint storage with async and sharded writes minimizes GPU idle time during saves. Measure, cache, and measure again.

Next Module

Continue to Module 1.5: Serving LLMs at Scale to learn how to deploy large language models for inference with vLLM, continuous batching, and KEDA autoscaling.

“Data is the new oil, but storage is the pipeline. A clogged pipeline makes the oil worthless.” — Anonymous infrastructure engineer

Module 1.4: High-Performance Storage for AI

Prerequisites

What You’ll Be Able to Do

Why This Module Matters

The IO Bottleneck in ML Workloads

Where IO Happens

Workload IO Profiles

Profiling IO Bottlenecks

Storage Tiers for AI

The Storage Pyramid

Local NVMe Caching

Why Local Storage Matters

TopoLVM: Topology-Aware Local Volumes

OpenEBS LVM Local PV

Init Container Pattern for Data Staging

Distributed Filesystems

CephFS

JuiceFS

Installing JuiceFS CSI Driver

JuiceFS StorageClass with Caching

Dataset Caching: Fluid and Alluxio

The Caching Problem

Fluid: Kubernetes-Native Dataset Orchestration

Fluid’s Data-Aware Scheduling

Alluxio Standalone (without Fluid)

Checkpoint Storage

Why Checkpoint IO Matters

Strategies for Fast Checkpointing

Recommended Checkpoint Storage

Try This: Measure Your Storage Performance

Did You Know?

War Story: The Cache That Saved $200K

Common Mistakes

Quiz: Check Your Understanding

Question 1

Question 2

Question 3

Question 4

Hands-On Exercise: JuiceFS Cache Over S3 with Latency Measurement

Objective

Environment

Step 1: Deploy MinIO (S3-compatible Object Store)

Step 2: Create a Test Dataset in MinIO

Step 3: Deploy Redis (JuiceFS Metadata Engine)

Step 4: Install JuiceFS CSI Driver

Step 5: Create JuiceFS StorageClass and PVC

Step 6: Measure Cold vs Warm Cache Performance

Step 7: Cleanup

Success Criteria

Key Takeaways

Further Reading

Summary

Next Module