Initial commit: nccl-perftest Helm chart
- Original K8s manifest (hostdev-node1-pod1.yaml) - Helm chart with templated ConfigMap, Service, StatefulSet - README with install guide and values reference
This commit is contained in:
Cloud User
commit
18d9630d08
|
|
@ -0,0 +1,2 @@
|
|||
.env
|
||||
*.tgz
|
||||
|
|
@ -0,0 +1,119 @@
|
|||
# nccl-perftest Helm Chart
|
||||
|
||||
Multi-node NCCL 분산 학습 성능 테스트를 위한 Helm chart.
|
||||
VGG11 + CIFAR-10 기반 torchrun 분산 학습을 StatefulSet으로 배포한다.
|
||||
|
||||
## Prerequisites
|
||||
|
||||
- Kubernetes 1.24+
|
||||
- Helm 3.x
|
||||
- NVIDIA GPU Operator 설치
|
||||
- Multus CNI + hostdevice NAD 구성
|
||||
- InfiniBand 장치 (`/dev/infiniband`)
|
||||
|
||||
## Helm Repo (Harbor OCI)
|
||||
|
||||
```bash
|
||||
# Harbor 로그인
|
||||
helm registry login harbor.inje-private.com -u admin
|
||||
|
||||
# Chart pull
|
||||
helm pull oci://harbor.inje-private.com/nccl-perftest/nccl-perftest --version 0.1.0
|
||||
```
|
||||
|
||||
## Install
|
||||
|
||||
```bash
|
||||
# 기본 설치 (values.yaml 기본값 사용)
|
||||
helm install nccl-perftest ./nccl-perftest
|
||||
|
||||
# 릴리스 이름 지정
|
||||
helm install my-release ./nccl-perftest
|
||||
|
||||
# OCI 레지스트리에서 직접 설치
|
||||
helm install nccl-perftest oci://harbor.inje-private.com/nccl-perftest/nccl-perftest --version 0.1.0
|
||||
```
|
||||
|
||||
## Uninstall
|
||||
|
||||
```bash
|
||||
helm uninstall nccl-perftest
|
||||
```
|
||||
|
||||
## Values 설정
|
||||
|
||||
### 커스텀 values 파일 사용
|
||||
|
||||
```bash
|
||||
helm install nccl-perftest ./nccl-perftest -f my-values.yaml
|
||||
```
|
||||
|
||||
### CLI로 개별 값 오버라이드
|
||||
|
||||
```bash
|
||||
# 노드 수 변경 (3노드)
|
||||
helm install nccl-perftest ./nccl-perftest --set replicaCount=3
|
||||
|
||||
# GPU 수 + 노드 수 동시 변경
|
||||
helm install nccl-perftest ./nccl-perftest \
|
||||
--set replicaCount=4 \
|
||||
--set resources.gpu=4 \
|
||||
--set training.nprocPerNode=4
|
||||
|
||||
# 이미지 태그 변경
|
||||
helm install nccl-perftest ./nccl-perftest --set image.tag="24.12-py3"
|
||||
|
||||
# NCCL 디버그 옵션
|
||||
helm install nccl-perftest ./nccl-perftest \
|
||||
--set nccl.ibHCA="mlx5_0\,mlx5_1\,mlx5_2\,mlx5_3" \
|
||||
--set nccl.debugFile="/tmp/nccl-%h-%p.log"
|
||||
```
|
||||
|
||||
## Values Reference
|
||||
|
||||
| Key | Default | Description |
|
||||
|-----|---------|-------------|
|
||||
| `nameOverride` | `nccl-test` | 리소스 이름 |
|
||||
| `replicaCount` | `2` | 노드(Pod) 수. NNODES에 자동 반영 |
|
||||
| `image.repository` | `nvcr.io/nvidia/pytorch` | 컨테이너 이미지 |
|
||||
| `image.tag` | `24.10-py3` | 이미지 태그 |
|
||||
| `image.pullPolicy` | `IfNotPresent` | 이미지 풀 정책 |
|
||||
| `training.nprocPerNode` | `8` | 노드당 GPU 프로세스 수 |
|
||||
| `training.masterPort` | `29500` | 마스터 rendezvous 포트 |
|
||||
| `training.batchSize` | `1200` | 글로벌 배치 사이즈 |
|
||||
| `training.epochs` | `90` | 학습 에포크 수 |
|
||||
| `training.lr` | `0.01` | 학습률 |
|
||||
| `resources.gpu` | `8` | nvidia.com/gpu 요청 수 |
|
||||
| `resources.hostdev` | `2` | nvidia.com/hostdev 요청 수 |
|
||||
| `shmSize` | `128Gi` | /dev/shm 크기 |
|
||||
| `network.nadName` | `hostdevice-net` | Multus NAD 이름 |
|
||||
| `nccl.debug` | `INFO` | NCCL_DEBUG 레벨 |
|
||||
| `nccl.debugSubsys` | `INIT,NET,IB` | NCCL_DEBUG_SUBSYS |
|
||||
| `nccl.ibDisable` | `0` | NCCL_IB_DISABLE (0=IB 사용) |
|
||||
| `nccl.socketIfname` | `net` | NCCL_SOCKET_IFNAME |
|
||||
| `nccl.ibHCA` | `""` | NCCL_IB_HCA (비어있으면 미설정) |
|
||||
| `nccl.debugFile` | `""` | NCCL_DEBUG_FILE (비어있으면 미설정) |
|
||||
| `volumes.infiniband.hostPath` | `/dev/infiniband` | IB 장치 경로 |
|
||||
| `volumes.dataset.hostPath` | `/home/ubuntu/cifar-10-batches-py` | CIFAR-10 데이터셋 경로 |
|
||||
|
||||
## 예시: 3노드 4GPU 설정
|
||||
|
||||
```yaml
|
||||
# my-values.yaml
|
||||
replicaCount: 3
|
||||
|
||||
training:
|
||||
nprocPerNode: 4
|
||||
batchSize: 600
|
||||
|
||||
resources:
|
||||
gpu: 4
|
||||
hostdev: 1
|
||||
|
||||
nccl:
|
||||
ibHCA: "mlx5_0,mlx5_1"
|
||||
```
|
||||
|
||||
```bash
|
||||
helm install nccl-3node ./nccl-perftest -f my-values.yaml
|
||||
```
|
||||
|
|
@ -0,0 +1,376 @@
|
|||
apiVersion: v1
|
||||
kind: ConfigMap
|
||||
metadata:
|
||||
name: nccl-test-scripts
|
||||
data:
|
||||
run.sh: |
|
||||
#!/usr/bin/env bash
|
||||
set -ex
|
||||
|
||||
# ----- StatefulSet Pod 이름에서 node_rank 추출 -----
|
||||
# Pod 이름: nccl-test-0, nccl-test-1, nccl-test-2 ...
|
||||
NODE_RANK=${HOSTNAME##*-}
|
||||
echo "[INFO] HOSTNAME=$HOSTNAME, NODE_RANK=$NODE_RANK"
|
||||
|
||||
# ----- NCCL 환경변수 -----
|
||||
export NCCL_DEBUG=INFO
|
||||
export NCCL_DEBUG_SUBSYS=INIT,NET,IB
|
||||
export NCCL_IB_DISABLE=0
|
||||
export NCCL_SOCKET_IFNAME="net"
|
||||
# export NCCL_IB_HCA="mlx5_0,mlx5_1,mlx5_2,mlx5_3"
|
||||
# export NCCL_DEBUG_FILE="/tmp/nccl-%h-%p.log" # 주석처리: tee로 파일+stdout 동시 출력
|
||||
|
||||
# ----- 학습 실행 -----
|
||||
torchrun \
|
||||
--nnodes=${NNODES} \
|
||||
--nproc_per_node=${NPROC_PER_NODE} \
|
||||
--rdzv_backend=c10d \
|
||||
--node_rank=${NODE_RANK} \
|
||||
--rdzv_endpoint=${MASTER_ADDR}:${MASTER_PORT} \
|
||||
/workspace/train_nccl.py 2>&1 | tee /tmp/nccl-aggregate.log
|
||||
|
||||
train_nccl.py: |
|
||||
import os
|
||||
import time
|
||||
import torch
|
||||
import argparse
|
||||
import torch.distributed as dist
|
||||
import torch.multiprocessing as mp
|
||||
from torch.optim.lr_scheduler import StepLR
|
||||
|
||||
# for dataset
|
||||
from torchvision.datasets.cifar import CIFAR10
|
||||
import torchvision.transforms as tfs
|
||||
from torch.utils.data import DataLoader
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
# for model
|
||||
from torchvision.models import vgg11
|
||||
from torch.nn.parallel import DistributedDataParallel as DDP
|
||||
|
||||
import numpy as np
|
||||
import random
|
||||
import datetime
|
||||
|
||||
|
||||
def set_random_seeds(random_seed=0):
|
||||
torch.manual_seed(random_seed)
|
||||
torch.backends.cudnn.deterministic = True
|
||||
torch.backends.cudnn.benchmark = False
|
||||
np.random.seed(random_seed)
|
||||
random.seed(random_seed)
|
||||
|
||||
|
||||
def get_args_parser():
|
||||
parser = argparse.ArgumentParser(add_help=False)
|
||||
parser.add_argument('--lr', type=float, default=0.01)
|
||||
parser.add_argument('--epoch', type=int, default=90)
|
||||
parser.add_argument('--batch_size', type=int, default=1200)
|
||||
parser.add_argument('--global_rank', type=int, default=0)
|
||||
parser.add_argument('--vis_step', type=int, default=10)
|
||||
parser.add_argument('--num_workers', type=int, default=24)
|
||||
parser.add_argument("--local_rank", type=int,
|
||||
help="Local rank. Necessary for using the torch.distributed.launch utility.")
|
||||
parser.add_argument('--world_size', type=int, default=0)
|
||||
parser.add_argument('--port', type=int, default=2022)
|
||||
parser.add_argument('--root', type=str, default='data')
|
||||
parser.add_argument('--start_epoch', type=int, default=0)
|
||||
parser.add_argument('--save_path', type=str, default='./save')
|
||||
parser.add_argument('--save_file_name', type=str, default='vgg_cifar')
|
||||
return parser
|
||||
|
||||
|
||||
def main(opts):
|
||||
# 1. set random seeds
|
||||
set_random_seeds(random_seed=0)
|
||||
|
||||
# 2. initialization
|
||||
init_for_distributed(opts)
|
||||
|
||||
# 3. visdom
|
||||
vis = None
|
||||
|
||||
# 4. data set
|
||||
transform_train = tfs.Compose([
|
||||
tfs.Resize(256),
|
||||
tfs.RandomCrop(224),
|
||||
tfs.RandomHorizontalFlip(),
|
||||
tfs.ToTensor(),
|
||||
tfs.Normalize(mean=(0.4914, 0.4822, 0.4465),
|
||||
std=(0.2023, 0.1994, 0.2010)),
|
||||
])
|
||||
|
||||
transform_test = tfs.Compose([
|
||||
tfs.Resize(256),
|
||||
tfs.CenterCrop(224),
|
||||
tfs.ToTensor(),
|
||||
tfs.Normalize(mean=(0.4914, 0.4822, 0.4465),
|
||||
std=(0.2023, 0.1994, 0.2010)),
|
||||
])
|
||||
|
||||
train_set = CIFAR10(root=opts.root,
|
||||
train=True,
|
||||
transform=transform_train,
|
||||
download=True)
|
||||
|
||||
test_set = CIFAR10(root=opts.root,
|
||||
train=False,
|
||||
transform=transform_test,
|
||||
download=True)
|
||||
|
||||
train_sampler = DistributedSampler(dataset=train_set, shuffle=True)
|
||||
test_sampler = DistributedSampler(dataset=test_set, shuffle=False)
|
||||
|
||||
train_loader = DataLoader(dataset=train_set,
|
||||
batch_size=int(opts.batch_size / opts.world_size),
|
||||
shuffle=False,
|
||||
num_workers=int(opts.num_workers / opts.world_size),
|
||||
sampler=train_sampler,
|
||||
pin_memory=True)
|
||||
|
||||
test_loader = DataLoader(dataset=test_set,
|
||||
batch_size=int(opts.batch_size / opts.world_size),
|
||||
shuffle=False,
|
||||
num_workers=int(opts.num_workers / opts.world_size),
|
||||
sampler=test_sampler,
|
||||
pin_memory=True)
|
||||
|
||||
# 5. model
|
||||
model = vgg11(pretrained=False)
|
||||
model = model.cuda(opts.local_rank)
|
||||
model = DDP(module=model,
|
||||
device_ids=[opts.local_rank])
|
||||
|
||||
# 6. criterion
|
||||
criterion = torch.nn.CrossEntropyLoss().to(opts.local_rank)
|
||||
|
||||
# 7. optimizer
|
||||
optimizer = torch.optim.SGD(params=model.parameters(),
|
||||
lr=0.01,
|
||||
weight_decay=0.0005,
|
||||
momentum=0.9)
|
||||
|
||||
# 8. scheduler
|
||||
scheduler = StepLR(optimizer=optimizer,
|
||||
step_size=30,
|
||||
gamma=0.1)
|
||||
|
||||
if opts.start_epoch != 0:
|
||||
|
||||
checkpoint = torch.load(os.path.join(opts.save_path, opts.save_file_name) + '.{}.pth.tar'
|
||||
.format(opts.start_epoch - 1),
|
||||
map_location=torch.device('cuda:{}'.format(opts.local_rank)))
|
||||
model.load_state_dict(checkpoint['model_state_dict']) # load model state dict
|
||||
optimizer.load_state_dict(checkpoint['optimizer_state_dict']) # load optim state dict
|
||||
scheduler.load_state_dict(checkpoint['scheduler_state_dict']) # load sched state dict
|
||||
if opts.global_rank == 0:
|
||||
print('\nLoaded checkpoint from epoch %d.\n' % (int(opts.start_epoch) - 1))
|
||||
|
||||
for epoch in range(opts.start_epoch, opts.epoch):
|
||||
|
||||
# 9. train
|
||||
tic = time.time()
|
||||
model.train()
|
||||
train_sampler.set_epoch(epoch)
|
||||
|
||||
for i, (images, labels) in enumerate(train_loader):
|
||||
images = images.to(opts.local_rank)
|
||||
labels = labels.to(opts.local_rank)
|
||||
outputs = model(images)
|
||||
|
||||
# ----------- update -----------
|
||||
optimizer.zero_grad()
|
||||
loss = criterion(outputs, labels)
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
|
||||
# get lr
|
||||
for param_group in optimizer.param_groups:
|
||||
lr = param_group['lr']
|
||||
|
||||
# time
|
||||
toc = time.time()
|
||||
|
||||
# visualization
|
||||
if (i % opts.vis_step == 0 or i == len(train_loader) - 1):
|
||||
print('GPU[{0}] Epoch [{1}/{2}], Iter [{3}/{4}], Loss: {5:.4f}, LR: {6:.5f}, Time: {7:.2f}'.format(opts.global_rank,
|
||||
epoch,
|
||||
opts.epoch,
|
||||
i,
|
||||
len(train_loader),
|
||||
loss.item(),
|
||||
lr,
|
||||
toc - tic))
|
||||
|
||||
# save pth file
|
||||
if opts.local_rank == 0:
|
||||
if not os.path.exists(opts.save_path):
|
||||
os.mkdir(opts.save_path)
|
||||
|
||||
checkpoint = {'epoch': epoch,
|
||||
'model_state_dict': model.state_dict(),
|
||||
'optimizer_state_dict': optimizer.state_dict(),
|
||||
'scheduler_state_dict': scheduler.state_dict()}
|
||||
|
||||
torch.save(checkpoint, os.path.join(opts.save_path, opts.save_file_name + '.{}.pth.tar'.format(epoch)))
|
||||
print("save pth.tar {} epoch!".format(epoch))
|
||||
|
||||
# 10. test
|
||||
model.eval()
|
||||
|
||||
val_avg_loss = 0
|
||||
correct_top1 = 0
|
||||
correct_top5 = 0
|
||||
total = 0
|
||||
|
||||
with torch.no_grad():
|
||||
for i, (images, labels) in enumerate(test_loader):
|
||||
images = images.to(opts.local_rank) # [100, 3, 224, 224]
|
||||
labels = labels.to(opts.local_rank) # [100]
|
||||
outputs = model(images)
|
||||
loss = criterion(outputs, labels)
|
||||
val_avg_loss += loss.item()
|
||||
# ------------------------------------------------------------------------------
|
||||
# rank 1
|
||||
_, pred = torch.max(outputs, 1)
|
||||
total += labels.size(0)
|
||||
correct_top1 += (pred == labels).sum().item()
|
||||
|
||||
# ------------------------------------------------------------------------------
|
||||
# rank 5
|
||||
_, rank5 = outputs.topk(5, 1, True, True)
|
||||
rank5 = rank5.t()
|
||||
correct5 = rank5.eq(labels.view(1, -1).expand_as(rank5))
|
||||
|
||||
# ------------------------------------------------------------------------------
|
||||
for k in range(5): # 0, 1, 2, 3, 4, 5
|
||||
correct_k = correct5[:k+1].reshape(-1).float().sum(0, keepdim=True)
|
||||
correct_top5 += correct_k.item()
|
||||
|
||||
accuracy_top1 = correct_top1 / total
|
||||
accuracy_top5 = correct_top5 / total
|
||||
|
||||
val_avg_loss = val_avg_loss / len(test_loader) # make mean loss
|
||||
|
||||
print("top-1 percentage : {0:0.3f}%".format(correct_top1 / total * 100))
|
||||
print("top-5 percentage : {0:0.3f}%".format(correct_top5 / total * 100))
|
||||
scheduler.step()
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
def init_for_distributed(opts):
|
||||
|
||||
# 1. setting for distributed training
|
||||
opts.global_rank = int(os.environ['RANK'])
|
||||
opts.local_rank = int(os.environ['LOCAL_RANK'])
|
||||
opts.world_size = int(os.environ['WORLD_SIZE'])
|
||||
torch.cuda.set_device(opts.local_rank)
|
||||
if opts.global_rank is not None and opts.local_rank is not None:
|
||||
print("Use GPU: [{}/{}] for training".format(opts.global_rank, opts.local_rank))
|
||||
|
||||
# 2. init_process_group
|
||||
dist.init_process_group(
|
||||
backend="nccl",
|
||||
rank=opts.global_rank,
|
||||
world_size=opts.world_size,
|
||||
device_id=torch.device(f"cuda:{opts.local_rank}"),
|
||||
timeout=datetime.timedelta(seconds=60)
|
||||
)
|
||||
|
||||
return
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
|
||||
parser = argparse.ArgumentParser('vgg11 cifar training', parents=[get_args_parser()])
|
||||
opts = parser.parse_args()
|
||||
main(opts)
|
||||
---
|
||||
apiVersion: v1
|
||||
kind: Service
|
||||
metadata:
|
||||
name: nccl-test
|
||||
labels: { app: nccl-test }
|
||||
spec:
|
||||
clusterIP: None # Headless Service
|
||||
selector: { app: nccl-test }
|
||||
ports:
|
||||
- name: rdzv
|
||||
port: 29500
|
||||
---
|
||||
apiVersion: apps/v1
|
||||
kind: StatefulSet
|
||||
metadata:
|
||||
name: nccl-test
|
||||
spec:
|
||||
serviceName: nccl-test
|
||||
replicas: 2 # 3노드 (ib-1, ib-2, ib-3)
|
||||
selector:
|
||||
matchLabels: { app: nccl-test }
|
||||
template:
|
||||
metadata:
|
||||
labels: { app: nccl-test }
|
||||
annotations:
|
||||
k8s.v1.cni.cncf.io/networks: hostdevice-net
|
||||
spec:
|
||||
containers:
|
||||
- name: worker
|
||||
image: nvcr.io/nvidia/pytorch:24.10-py3
|
||||
imagePullPolicy: IfNotPresent
|
||||
securityContext:
|
||||
privileged: true
|
||||
capabilities:
|
||||
add:
|
||||
- IPC_LOCK
|
||||
command:
|
||||
- bash
|
||||
- -lc
|
||||
- |
|
||||
cp /config_scripts/* /workspace/ && \
|
||||
chmod +x /workspace/run.sh && \
|
||||
cd /workspace && ./run.sh; \
|
||||
sleep infinity
|
||||
env:
|
||||
# ----- 분산 학습 설정 -----
|
||||
- name: NNODES
|
||||
value: "2"
|
||||
- name: NPROC_PER_NODE
|
||||
value: "8"
|
||||
- name: MASTER_ADDR
|
||||
value: "nccl-test-0.nccl-test"
|
||||
- name: MASTER_PORT
|
||||
value: "29500"
|
||||
resources:
|
||||
limits:
|
||||
nvidia.com/gpu: "8"
|
||||
nvidia.com/hostdev: 2
|
||||
requests:
|
||||
nvidia.com/gpu: "8"
|
||||
nvidia.com/hostdev: 2
|
||||
volumeMounts:
|
||||
- name: scripts
|
||||
mountPath: /config_scripts
|
||||
- name: shared-memory
|
||||
mountPath: /dev/shm
|
||||
- name: infiniband
|
||||
mountPath: /dev/infiniband
|
||||
- name: cifar-data
|
||||
mountPath: /workspace/data/cifar-10-batches-py
|
||||
volumes:
|
||||
- name: scripts
|
||||
configMap:
|
||||
name: nccl-test-scripts
|
||||
defaultMode: 0755
|
||||
- name: shared-memory
|
||||
emptyDir:
|
||||
medium: Memory
|
||||
sizeLimit: 128Gi # GPU 8장 x 32GB = 256GB, 권장: 50% = 128Gi
|
||||
- name: infiniband
|
||||
hostPath:
|
||||
path: /dev/infiniband
|
||||
type: Directory
|
||||
- name: cifar-data
|
||||
hostPath:
|
||||
path: /home/ubuntu/cifar-10-batches-py
|
||||
type: Directory
|
||||
|
|
@ -0,0 +1,6 @@
|
|||
apiVersion: v2
|
||||
name: nccl-perftest
|
||||
description: Multi-node NCCL distributed training performance test
|
||||
type: application
|
||||
version: 0.1.0
|
||||
appVersion: "1.0.0"
|
||||
|
|
@ -0,0 +1,33 @@
|
|||
{{/*
|
||||
Chart name, overridden by nameOverride.
|
||||
*/}}
|
||||
{{- define "nccl-perftest.name" -}}
|
||||
{{- default .Chart.Name .Values.nameOverride | trunc 63 | trimSuffix "-" }}
|
||||
{{- end }}
|
||||
|
||||
{{/*
|
||||
Fully qualified app name (release-aware).
|
||||
*/}}
|
||||
{{- define "nccl-perftest.fullname" -}}
|
||||
{{- if .Values.nameOverride }}
|
||||
{{- .Values.nameOverride | trunc 63 | trimSuffix "-" }}
|
||||
{{- else }}
|
||||
{{- printf "%s-%s" .Release.Name .Chart.Name | trunc 63 | trimSuffix "-" }}
|
||||
{{- end }}
|
||||
{{- end }}
|
||||
|
||||
{{/*
|
||||
Common labels.
|
||||
*/}}
|
||||
{{- define "nccl-perftest.labels" -}}
|
||||
app: {{ include "nccl-perftest.name" . }}
|
||||
helm.sh/chart: {{ printf "%s-%s" .Chart.Name .Chart.Version | replace "+" "_" | trunc 63 | trimSuffix "-" }}
|
||||
app.kubernetes.io/managed-by: {{ .Release.Service }}
|
||||
{{- end }}
|
||||
|
||||
{{/*
|
||||
Selector labels.
|
||||
*/}}
|
||||
{{- define "nccl-perftest.selectorLabels" -}}
|
||||
app: {{ include "nccl-perftest.name" . }}
|
||||
{{- end }}
|
||||
|
|
@ -0,0 +1,294 @@
|
|||
apiVersion: v1
|
||||
kind: ConfigMap
|
||||
metadata:
|
||||
name: {{ include "nccl-perftest.name" . }}-scripts
|
||||
labels:
|
||||
{{- include "nccl-perftest.labels" . | nindent 4 }}
|
||||
data:
|
||||
run.sh: |
|
||||
#!/usr/bin/env bash
|
||||
set -ex
|
||||
|
||||
# ----- StatefulSet Pod 이름에서 node_rank 추출 -----
|
||||
# Pod 이름: nccl-test-0, nccl-test-1, nccl-test-2 ...
|
||||
NODE_RANK=${HOSTNAME##*-}
|
||||
echo "[INFO] HOSTNAME=$HOSTNAME, NODE_RANK=$NODE_RANK"
|
||||
|
||||
# ----- NCCL 환경변수 -----
|
||||
export NCCL_DEBUG={{ .Values.nccl.debug }}
|
||||
export NCCL_DEBUG_SUBSYS={{ .Values.nccl.debugSubsys }}
|
||||
export NCCL_IB_DISABLE={{ .Values.nccl.ibDisable }}
|
||||
export NCCL_SOCKET_IFNAME="{{ .Values.nccl.socketIfname }}"
|
||||
{{- if .Values.nccl.ibHCA }}
|
||||
export NCCL_IB_HCA="{{ .Values.nccl.ibHCA }}"
|
||||
{{- end }}
|
||||
{{- if .Values.nccl.debugFile }}
|
||||
export NCCL_DEBUG_FILE="{{ .Values.nccl.debugFile }}"
|
||||
{{- end }}
|
||||
|
||||
# ----- 학습 실행 -----
|
||||
torchrun \
|
||||
--nnodes=${NNODES} \
|
||||
--nproc_per_node=${NPROC_PER_NODE} \
|
||||
--rdzv_backend=c10d \
|
||||
--node_rank=${NODE_RANK} \
|
||||
--rdzv_endpoint=${MASTER_ADDR}:${MASTER_PORT} \
|
||||
/workspace/train_nccl.py 2>&1 | tee /tmp/nccl-aggregate.log
|
||||
|
||||
train_nccl.py: |
|
||||
import os
|
||||
import time
|
||||
import torch
|
||||
import argparse
|
||||
import torch.distributed as dist
|
||||
import torch.multiprocessing as mp
|
||||
from torch.optim.lr_scheduler import StepLR
|
||||
|
||||
# for dataset
|
||||
from torchvision.datasets.cifar import CIFAR10
|
||||
import torchvision.transforms as tfs
|
||||
from torch.utils.data import DataLoader
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
# for model
|
||||
from torchvision.models import vgg11
|
||||
from torch.nn.parallel import DistributedDataParallel as DDP
|
||||
|
||||
import numpy as np
|
||||
import random
|
||||
import datetime
|
||||
|
||||
|
||||
def set_random_seeds(random_seed=0):
|
||||
torch.manual_seed(random_seed)
|
||||
torch.backends.cudnn.deterministic = True
|
||||
torch.backends.cudnn.benchmark = False
|
||||
np.random.seed(random_seed)
|
||||
random.seed(random_seed)
|
||||
|
||||
|
||||
def get_args_parser():
|
||||
parser = argparse.ArgumentParser(add_help=False)
|
||||
parser.add_argument('--lr', type=float, default={{ .Values.training.lr }})
|
||||
parser.add_argument('--epoch', type=int, default={{ .Values.training.epochs }})
|
||||
parser.add_argument('--batch_size', type=int, default={{ .Values.training.batchSize }})
|
||||
parser.add_argument('--global_rank', type=int, default=0)
|
||||
parser.add_argument('--vis_step', type=int, default=10)
|
||||
parser.add_argument('--num_workers', type=int, default=24)
|
||||
parser.add_argument("--local_rank", type=int,
|
||||
help="Local rank. Necessary for using the torch.distributed.launch utility.")
|
||||
parser.add_argument('--world_size', type=int, default=0)
|
||||
parser.add_argument('--port', type=int, default=2022)
|
||||
parser.add_argument('--root', type=str, default='data')
|
||||
parser.add_argument('--start_epoch', type=int, default=0)
|
||||
parser.add_argument('--save_path', type=str, default='./save')
|
||||
parser.add_argument('--save_file_name', type=str, default='vgg_cifar')
|
||||
return parser
|
||||
|
||||
|
||||
def main(opts):
|
||||
# 1. set random seeds
|
||||
set_random_seeds(random_seed=0)
|
||||
|
||||
# 2. initialization
|
||||
init_for_distributed(opts)
|
||||
|
||||
# 3. visdom
|
||||
vis = None
|
||||
|
||||
# 4. data set
|
||||
transform_train = tfs.Compose([
|
||||
tfs.Resize(256),
|
||||
tfs.RandomCrop(224),
|
||||
tfs.RandomHorizontalFlip(),
|
||||
tfs.ToTensor(),
|
||||
tfs.Normalize(mean=(0.4914, 0.4822, 0.4465),
|
||||
std=(0.2023, 0.1994, 0.2010)),
|
||||
])
|
||||
|
||||
transform_test = tfs.Compose([
|
||||
tfs.Resize(256),
|
||||
tfs.CenterCrop(224),
|
||||
tfs.ToTensor(),
|
||||
tfs.Normalize(mean=(0.4914, 0.4822, 0.4465),
|
||||
std=(0.2023, 0.1994, 0.2010)),
|
||||
])
|
||||
|
||||
train_set = CIFAR10(root=opts.root,
|
||||
train=True,
|
||||
transform=transform_train,
|
||||
download=True)
|
||||
|
||||
test_set = CIFAR10(root=opts.root,
|
||||
train=False,
|
||||
transform=transform_test,
|
||||
download=True)
|
||||
|
||||
train_sampler = DistributedSampler(dataset=train_set, shuffle=True)
|
||||
test_sampler = DistributedSampler(dataset=test_set, shuffle=False)
|
||||
|
||||
train_loader = DataLoader(dataset=train_set,
|
||||
batch_size=int(opts.batch_size / opts.world_size),
|
||||
shuffle=False,
|
||||
num_workers=int(opts.num_workers / opts.world_size),
|
||||
sampler=train_sampler,
|
||||
pin_memory=True)
|
||||
|
||||
test_loader = DataLoader(dataset=test_set,
|
||||
batch_size=int(opts.batch_size / opts.world_size),
|
||||
shuffle=False,
|
||||
num_workers=int(opts.num_workers / opts.world_size),
|
||||
sampler=test_sampler,
|
||||
pin_memory=True)
|
||||
|
||||
# 5. model
|
||||
model = vgg11(pretrained=False)
|
||||
model = model.cuda(opts.local_rank)
|
||||
model = DDP(module=model,
|
||||
device_ids=[opts.local_rank])
|
||||
|
||||
# 6. criterion
|
||||
criterion = torch.nn.CrossEntropyLoss().to(opts.local_rank)
|
||||
|
||||
# 7. optimizer
|
||||
optimizer = torch.optim.SGD(params=model.parameters(),
|
||||
lr=0.01,
|
||||
weight_decay=0.0005,
|
||||
momentum=0.9)
|
||||
|
||||
# 8. scheduler
|
||||
scheduler = StepLR(optimizer=optimizer,
|
||||
step_size=30,
|
||||
gamma=0.1)
|
||||
|
||||
if opts.start_epoch != 0:
|
||||
|
||||
checkpoint = torch.load(os.path.join(opts.save_path, opts.save_file_name) + '.{}.pth.tar'
|
||||
.format(opts.start_epoch - 1),
|
||||
map_location=torch.device('cuda:{}'.format(opts.local_rank)))
|
||||
model.load_state_dict(checkpoint['model_state_dict']) # load model state dict
|
||||
optimizer.load_state_dict(checkpoint['optimizer_state_dict']) # load optim state dict
|
||||
scheduler.load_state_dict(checkpoint['scheduler_state_dict']) # load sched state dict
|
||||
if opts.global_rank == 0:
|
||||
print('\nLoaded checkpoint from epoch %d.\n' % (int(opts.start_epoch) - 1))
|
||||
|
||||
for epoch in range(opts.start_epoch, opts.epoch):
|
||||
|
||||
# 9. train
|
||||
tic = time.time()
|
||||
model.train()
|
||||
train_sampler.set_epoch(epoch)
|
||||
|
||||
for i, (images, labels) in enumerate(train_loader):
|
||||
images = images.to(opts.local_rank)
|
||||
labels = labels.to(opts.local_rank)
|
||||
outputs = model(images)
|
||||
|
||||
# ----------- update -----------
|
||||
optimizer.zero_grad()
|
||||
loss = criterion(outputs, labels)
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
|
||||
# get lr
|
||||
for param_group in optimizer.param_groups:
|
||||
lr = param_group['lr']
|
||||
|
||||
# time
|
||||
toc = time.time()
|
||||
|
||||
# visualization
|
||||
if (i % opts.vis_step == 0 or i == len(train_loader) - 1):
|
||||
print('GPU[{0}] Epoch [{1}/{2}], Iter [{3}/{4}], Loss: {5:.4f}, LR: {6:.5f}, Time: {7:.2f}'.format(opts.global_rank,
|
||||
epoch,
|
||||
opts.epoch,
|
||||
i,
|
||||
len(train_loader),
|
||||
loss.item(),
|
||||
lr,
|
||||
toc - tic))
|
||||
|
||||
# save pth file
|
||||
if opts.local_rank == 0:
|
||||
if not os.path.exists(opts.save_path):
|
||||
os.mkdir(opts.save_path)
|
||||
|
||||
checkpoint = {'epoch': epoch,
|
||||
'model_state_dict': model.state_dict(),
|
||||
'optimizer_state_dict': optimizer.state_dict(),
|
||||
'scheduler_state_dict': scheduler.state_dict()}
|
||||
|
||||
torch.save(checkpoint, os.path.join(opts.save_path, opts.save_file_name + '.{}.pth.tar'.format(epoch)))
|
||||
print("save pth.tar {} epoch!".format(epoch))
|
||||
|
||||
# 10. test
|
||||
model.eval()
|
||||
|
||||
val_avg_loss = 0
|
||||
correct_top1 = 0
|
||||
correct_top5 = 0
|
||||
total = 0
|
||||
|
||||
with torch.no_grad():
|
||||
for i, (images, labels) in enumerate(test_loader):
|
||||
images = images.to(opts.local_rank) # [100, 3, 224, 224]
|
||||
labels = labels.to(opts.local_rank) # [100]
|
||||
outputs = model(images)
|
||||
loss = criterion(outputs, labels)
|
||||
val_avg_loss += loss.item()
|
||||
# ------------------------------------------------------------------------------
|
||||
# rank 1
|
||||
_, pred = torch.max(outputs, 1)
|
||||
total += labels.size(0)
|
||||
correct_top1 += (pred == labels).sum().item()
|
||||
|
||||
# ------------------------------------------------------------------------------
|
||||
# rank 5
|
||||
_, rank5 = outputs.topk(5, 1, True, True)
|
||||
rank5 = rank5.t()
|
||||
correct5 = rank5.eq(labels.view(1, -1).expand_as(rank5))
|
||||
|
||||
# ------------------------------------------------------------------------------
|
||||
for k in range(5): # 0, 1, 2, 3, 4, 5
|
||||
correct_k = correct5[:k+1].reshape(-1).float().sum(0, keepdim=True)
|
||||
correct_top5 += correct_k.item()
|
||||
|
||||
accuracy_top1 = correct_top1 / total
|
||||
accuracy_top5 = correct_top5 / total
|
||||
|
||||
val_avg_loss = val_avg_loss / len(test_loader) # make mean loss
|
||||
|
||||
print("top-1 percentage : {0:0.3f}%".format(correct_top1 / total * 100))
|
||||
print("top-5 percentage : {0:0.3f}%".format(correct_top5 / total * 100))
|
||||
scheduler.step()
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
def init_for_distributed(opts):
|
||||
|
||||
# 1. setting for distributed training
|
||||
opts.global_rank = int(os.environ['RANK'])
|
||||
opts.local_rank = int(os.environ['LOCAL_RANK'])
|
||||
opts.world_size = int(os.environ['WORLD_SIZE'])
|
||||
torch.cuda.set_device(opts.local_rank)
|
||||
if opts.global_rank is not None and opts.local_rank is not None:
|
||||
print("Use GPU: [{}/{}] for training".format(opts.global_rank, opts.local_rank))
|
||||
|
||||
# 2. init_process_group
|
||||
dist.init_process_group(
|
||||
backend="nccl",
|
||||
rank=opts.global_rank,
|
||||
world_size=opts.world_size,
|
||||
device_id=torch.device(f"cuda:{opts.local_rank}"),
|
||||
timeout=datetime.timedelta(seconds=60)
|
||||
)
|
||||
|
||||
return
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
|
||||
parser = argparse.ArgumentParser('vgg11 cifar training', parents=[get_args_parser()])
|
||||
opts = parser.parse_args()
|
||||
main(opts)
|
||||
|
|
@ -0,0 +1,13 @@
|
|||
apiVersion: v1
|
||||
kind: Service
|
||||
metadata:
|
||||
name: {{ include "nccl-perftest.name" . }}
|
||||
labels:
|
||||
{{- include "nccl-perftest.labels" . | nindent 4 }}
|
||||
spec:
|
||||
clusterIP: None
|
||||
selector:
|
||||
{{- include "nccl-perftest.selectorLabels" . | nindent 4 }}
|
||||
ports:
|
||||
- name: rdzv
|
||||
port: {{ .Values.training.masterPort }}
|
||||
|
|
@ -0,0 +1,78 @@
|
|||
apiVersion: apps/v1
|
||||
kind: StatefulSet
|
||||
metadata:
|
||||
name: {{ include "nccl-perftest.name" . }}
|
||||
labels:
|
||||
{{- include "nccl-perftest.labels" . | nindent 4 }}
|
||||
spec:
|
||||
serviceName: {{ include "nccl-perftest.name" . }}
|
||||
replicas: {{ .Values.replicaCount }}
|
||||
selector:
|
||||
matchLabels:
|
||||
{{- include "nccl-perftest.selectorLabels" . | nindent 6 }}
|
||||
template:
|
||||
metadata:
|
||||
labels:
|
||||
{{- include "nccl-perftest.selectorLabels" . | nindent 8 }}
|
||||
annotations:
|
||||
k8s.v1.cni.cncf.io/networks: {{ .Values.network.nadName }}
|
||||
spec:
|
||||
containers:
|
||||
- name: worker
|
||||
image: {{ .Values.image.repository }}:{{ .Values.image.tag }}
|
||||
imagePullPolicy: {{ .Values.image.pullPolicy }}
|
||||
securityContext:
|
||||
privileged: true
|
||||
capabilities:
|
||||
add:
|
||||
- IPC_LOCK
|
||||
command:
|
||||
- bash
|
||||
- -lc
|
||||
- |
|
||||
cp /config_scripts/* /workspace/ && \
|
||||
chmod +x /workspace/run.sh && \
|
||||
cd /workspace && ./run.sh; \
|
||||
sleep infinity
|
||||
env:
|
||||
- name: NNODES
|
||||
value: {{ .Values.replicaCount | quote }}
|
||||
- name: NPROC_PER_NODE
|
||||
value: {{ .Values.training.nprocPerNode | quote }}
|
||||
- name: MASTER_ADDR
|
||||
value: "{{ include "nccl-perftest.name" . }}-0.{{ include "nccl-perftest.name" . }}"
|
||||
- name: MASTER_PORT
|
||||
value: {{ .Values.training.masterPort | quote }}
|
||||
resources:
|
||||
limits:
|
||||
nvidia.com/gpu: {{ .Values.resources.gpu | quote }}
|
||||
nvidia.com/hostdev: {{ .Values.resources.hostdev }}
|
||||
requests:
|
||||
nvidia.com/gpu: {{ .Values.resources.gpu | quote }}
|
||||
nvidia.com/hostdev: {{ .Values.resources.hostdev }}
|
||||
volumeMounts:
|
||||
- name: scripts
|
||||
mountPath: /config_scripts
|
||||
- name: shared-memory
|
||||
mountPath: /dev/shm
|
||||
- name: infiniband
|
||||
mountPath: /dev/infiniband
|
||||
- name: cifar-data
|
||||
mountPath: /workspace/data/cifar-10-batches-py
|
||||
volumes:
|
||||
- name: scripts
|
||||
configMap:
|
||||
name: {{ include "nccl-perftest.name" . }}-scripts
|
||||
defaultMode: 0755
|
||||
- name: shared-memory
|
||||
emptyDir:
|
||||
medium: Memory
|
||||
sizeLimit: {{ .Values.shmSize }}
|
||||
- name: infiniband
|
||||
hostPath:
|
||||
path: {{ .Values.volumes.infiniband.hostPath }}
|
||||
type: Directory
|
||||
- name: cifar-data
|
||||
hostPath:
|
||||
path: {{ .Values.volumes.dataset.hostPath }}
|
||||
type: Directory
|
||||
|
|
@ -0,0 +1,37 @@
|
|||
nameOverride: "nccl-test"
|
||||
replicaCount: 2
|
||||
|
||||
image:
|
||||
repository: nvcr.io/nvidia/pytorch
|
||||
tag: "24.10-py3"
|
||||
pullPolicy: IfNotPresent
|
||||
|
||||
training:
|
||||
nprocPerNode: 8
|
||||
masterPort: 29500
|
||||
batchSize: 1200
|
||||
epochs: 90
|
||||
lr: 0.01
|
||||
|
||||
resources:
|
||||
gpu: 8
|
||||
hostdev: 2
|
||||
|
||||
shmSize: "128Gi"
|
||||
|
||||
network:
|
||||
nadName: "hostdevice-net"
|
||||
|
||||
nccl:
|
||||
debug: "INFO"
|
||||
debugSubsys: "INIT,NET,IB"
|
||||
ibDisable: 0
|
||||
socketIfname: "net"
|
||||
ibHCA: "" # 비어있으면 설정 안 함
|
||||
debugFile: "" # 비어있으면 설정 안 함
|
||||
|
||||
volumes:
|
||||
infiniband:
|
||||
hostPath: "/dev/infiniband"
|
||||
dataset:
|
||||
hostPath: "/home/ubuntu/cifar-10-batches-py"
|
||||
Loading…
Reference in New Issue