Post Syndicated from Neelay Thaker original https://aws.amazon.com/blogs/compute/amazon-ec2-p4d-instances-deep-dive/

This post is contributed by Amr Ragab, Senior Solutions Architect, Amazon EC2

Introduction

AWS is excited to announce that the new Amazon EC2 P4d instances are now generally available. This instance type brings additional benefits with 2.5x higher deep learning performance; adding to the accelerated instances portfolio, new features, and technical breakthroughs that our customers can benefit from with this latest technology. This blog post details some of those key features and how to integrate them into your current workloads and architectures.

Overview

As you can see from the generalized block diagram above, the p4d comes with dual socket Intel Cascade Lake 8275CL processors totaling 96 vCPUs at 3.0 GHz with 1.1 TB of RAM and 8 TB of NVMe local storage. P4d also comes with 8 x 40 GB NVIDIA Tesla A100 GPUs with NVSwitch and 400 Gbps Elastic Fabric Adapter (EFA) enabled networking. This instance configuration represents the latest generation of computing for our customers spanning Machine Learning (ML), High Performance Computing (HPC), and analytics.

One of the improvements of the p4d is in the networking stack.  This new instance type has 400 Gbps with support for EFA and GPUDirect RDMA. Now, on AWS, you can take advantage of point-to-point GPU to GPU communication (across nodes), bypassing the CPU. Look out for additional blogs and webinars detailing use cases of GPUDirect and how this feature helps decrease latency and improve performance for certain workloads.

Let’s look at some new features and performance metrics for the P4d instances.

Features

Local ephemeral NVMe storage
The p4d instance type comes with 8 TB of local NVMe storage. Each device has a maximum read/write throughput of 2.7 GB/s. To create a local namespace and staging area for input into the GPUs, you can create a local RAID 0 of all the drives. This results in aggregate read throughput of about 16 GB/s. The following table summarizes the I/O tests on the NVMe drives in this configuration.

NVSwitch

Introduced with the p4d instance type is NVSwitch. Every GPU in the node is connected to each other in a full mesh topology up to 600 GB/s bidirectional bandwidth. ML frameworks and HPC applications that use NVIDIA communication collectives library (NCCL) can take full advantage of this all-to-all communication layer.

P4d uses a full mesh NVLink topology for optimized all-to-all communication, compared to the previous generation P3/P3dn instances, which have all-to-all communication across various data path domains (NUMA, PCIe switch, NVLink).  This new topology accessed via NCCL will improve performance for multiGPU workloads.
To make optimal use of the NVSwitch ensure that in your instance, all GPUs application boost clocks are set to its maximum values:

sudo nvidia-smi -ac 1215,1410

Multi-Instance GPU (MIG)

It’s now possible, at the user level, to have control of fractionating a GPU into multiple GPU slices, with each GPU slice isolated from each other. This enables multiple users to run different workloads on the same GPU without impacting performance. I walk you through an example implementation of MIG in the following steps:

With every newly launched instance, MIG is disabled. So, you must enable it with the following command:

ubuntu@ip-172-31-34-6:~# sudo nvidia-smi -mig 1 

Enabled MIG Mode for GPU 00000000:10:1C.0
You can get a list of supported MIG profiles.

Next, you can create seven slices, and create compute instances for each slice.

ubuntu@ip-172-31-34-6:~# sudo nvidia-smi mig -cgi 19,19,19,19,19,19,19 
Successfully created GPU instance ID 9 on GPU 0 using profile MIG 1g.5gb (ID 19) 
Successfully created GPU instance ID 7 on GPU 0 using profile MIG 1g.5gb (ID 19) 
Successfully created GPU instance ID 8 on GPU 0 using profile MIG 1g.5gb (ID 19) 
Successfully created GPU instance ID 11 on GPU 0 using profile MIG 1g.5gb (ID 19) 
Successfully created GPU instance ID 12 on GPU 0 using profile MIG 1g.5gb (ID 19) 
Successfully created GPU instance ID 13 on GPU 0 using profile MIG 1g.5gb (ID 19) 
Successfully created GPU instance ID 14 on GPU 0 using profile MIG 1g.5gb (ID 19)

ubuntu@ip-172-31-34-6:~# nvidia-smi mig -cci -gi 7,8,9,11,12,13,14 

Successfully created compute instance ID 0 on GPU 0 GPU instance ID 7 using profile MIG 1g.5gb (ID 0) 
Successfully created compute instance ID 0 on GPU 0 GPU instance ID 8 using profile MIG 1g.5gb (ID 0) 
Successfully created compute instance ID 0 on GPU 0 GPU instance ID 9 using profile MIG 1g.5gb (ID 0) 
Successfully created compute instance ID 0 on GPU 0 GPU instance ID 11 using profile MIG 1g.5gb (ID 0) 
Successfully created compute instance ID 0 on GPU 0 GPU instance ID 12 using profile MIG 1g.5gb (ID 0) 
Successfully created compute instance ID 0 on GPU 0 GPU instance ID 13 using profile MIG 1g.5gb (ID 0) 
Successfully created compute instance ID 0 on GPU 0 GPU instance ID 14 using profile MIG 1g.5gb (ID 0)

You can split a GPU into a maximum of seven slices. To pass the GPU through into a docker container, you can specify the index pair at runtime:

docker run -it --gpus '"device=1:0"' nvcr.io/nvidia/tensorflow:20.09-tf1-py3

With MIG, you can run multiple smaller workloads on the same GPU without compromising performance. We will follow up with additional blogs on this feature as we integrate it with additional AWS services.

NVIDIA GPUDirect RDMA over EFA

For workloads optimized for multiGPU capabilities, we introduced GPUDirect over EFA fabric. This allows direct GPU-GPU communication across multiple p4d nodes for decreased latency and improved performance. Follow this user guide to get started with installing the EFA driver and setting up the environment. The code sample below can be used as a template to use GPUDirect RDMA over EFA.

/opt/amazon/openmpi/bin/mpirun \
     -n ${NUM_PROCS} -N ${NUM_PROCS_NODE} \
     -x RDMAV_FORK_SAFE=1 -x NCCL_DEBUG=info \
     -x FI_EFA_USE_DEVICE_RDMA=1 \
     --hostfile ${HOSTS_FILE} \
     --mca pml ^cm --mca btl tcp,self --mca btl_tcp_if_exclude lo,docker0 --bind-to none \
     $HOME/nccl-tests/build/all_reduce_perf -b 8 -e 4G -f 2 -g 1 -c 1 -n 100

Machine learning Optimizations

You can quickly get started with the all benefits mentioned earlier for the p4d by using our latest Deep Learning AMI (DLAMI). The DLAMI now comes with CUDA11 and the latest NVLink and cuDNN SDKs and drivers to take advantage of the p4d.

TensorFloat32 – TF32

TF32 is a new 19 bit precision datatype from NVIDIA introduced for the first time for the p4d.24xlarge instance. This datatype improves performance with little to no loss of training and validation accuracy for most mainstream models. We have more detailed benchmarks for individual algorithms. But, on the p4d.24xlarge you can achieve approximately a 2.5 fold increase compared to FP32 on the p3dn.24xlarge for mainstream deep learning models.

We have updated our machine learning models here to show examples (see the following chart) of popular algorithms our customers are using today including general DNNs and Bert.

BERT Large – Wikipedia/Books Corpus

You can find other code examples at github.com/NVIDIA/DeepLearningExamples.

If you want to builld your own AMI or extend an AMI maintained by your organization you can use the github repo, which provides Packer scripts to build AMIs for Amazon Linux 2 or Ubuntu 18.04 versions.

https://github.com/aws-samples/aws-efa-nccl-baseami-pipeline

The stack includes the following components:

• NVIDIA Driver 450.80.02
• CUDA 11
• NVIDIA Fabric Manager
• cuDNN 8
• NCCL 2.7.8
• EFA latest driver
• AWS-OFI-NCCL
• FSx kernel and client driver and utilities
• Intel OneDNN
• NVIDIA-runtime Docker

Conclusion

Get started with the new P4d instances with support on Amazon EKS, AWS Batch, and Amazon Sagemaker. We are excited to hear about what you develop and run with the new P4d instances. If you have any questions please reach out to your account team. Now, go power up your ML and HPC workloads with NVIDIA Tesla A100s and the P4d instances.