Community

Blog Events Webinars Tutorials Forum
  1. Blog
  2. Events
  3. Webinars
  4. Tutorials
  5. Forum
Create Account
×
Community Blog Optimizing GPU Heterogeneous Parallel Performance (GROMACS) with E-HPC CloudMetrics

Optimizing GPU Heterogeneous Parallel Performance (GROMACS) with E-HPC CloudMetrics

This article uses GROMACS as an example to explain how to use Alibaba Cloud E-HPC's CloudMetrics to optimize GPU heterogeneous parallel performance.

By Mu Di

CloudMetrics is a built-in on-cloud performance monitoring and analysis engine of Alibaba Cloud Elastic High Performance Computing (E-HPC) service. It supports real-time monitoring of cluster resource usage and online analysis of user job running statuses. In a heterogeneous computing scenario where a GPU is used for acceleration, CloudMetrics can monitor the usage of node host resources, as well as the usage of GPU device resources. It shows the changes to performance metrics over time, such as GPU usage, video RAM usage, and PCI-E data transmission bandwidth. This helps users better understand the running status of applications and provides guidance for optimizing applications.

What Is GROMACS?

Molecular Dynamics (MD) simulation is a computer simulation method for studying and analyzing the structure and properties of molecules and molecular systems using classical Newton mechanics. It is the key that can open up the world of microscopic motion. As one of the important areas of research in the HPC field, MD simulation is widely used in the development of new materials, chemical industry simulations, bio-medicine, and other fields.

Relevant research findings have been nominated several times for the ACM Gordon Bell Prize, one of the three major engineering awards in U.S. MD simulation involves massive amounts of intermolecular interaction force computing, and the computing processes are intensive and regular. Therefore, CPU vector computing components and GPUs are suitable for the acceleration of MD simulations. GPUs can be used to accelerate the core computing processes of mainstream MD simulation software, including GROMACS, NAMD, and LAMMPS. The following figure shows the evolution of the biological macromolecule configuration obtained by using the MD simulation method.

1

Full utilization of GPU resources is the key to achieving optimal simulation efficiency in MD software. Using GROMACS as an example, this article uses CloudMetrics to present the runtime performance characteristics of GROMACS from different perspectives and analyze the current performance bottlenecks to optimize the software performance.

GPU Performance Monitoring Function of CloudMetrics

2

Node Dimension: CloudMetrics constructs multiple coordinate planes according to nodes selected by the user. Each coordinate plane shows the GPU resource usage of one node. Performance data of different GPUs on the same node are displayed on the same coordinate plane, with time as its x-axis. This node-based presentation mode helps users compare loads of different GPUs on the same node.

3

Metric Dimension: CloudMetrics constructs multiple coordinate planes based on the GPU sub-devices and performance metrics selected by the user. Each coordinate plane shows the performance of a GPU sub-device or a performance metric on different nodes. This node-based presentation mode helps users compare the GPU loads of different nodes.

4

Using CloudMetrics to Provide Guidance for GROMACS Performance Optimization

In the following example, the GPU version of GROMACS runs on an ecs.gn5 instance (8 cores, 16 vCPUs, 120 GB of memory, and 2 P100s), and the system resource usage information monitored by CloudMetrics is used for optimization.

Node Dimension:

5

Metric Dimension:

6

The two figures indicate that:

  1. The CPU usage is constantly above 85%.
  2. The GPU usage ranges from 45% to 70% and fluctuates significantly.
  3. The video RAM usage of the GPU is about 25%.
  4. The PCI-E transmission bandwidth of the GPU generally does not exceed 2 Gbit/s.

This indicates that the CPU computing resources are almost exhausted and the load is heavy. However, the GPU computing resources, video RAM, and PCI-E bandwidth have not reached their bottlenecks yet and can be utilized further. The GROMACS software itself uses the "CPU + GPU" active/standby collaborative computing mode.

The performance bottleneck of either the CPU or GPU will slow down the overall performance of the software. Therefore, to improve the software execution efficiency of GROMACS, you can move the load of the PME computing process from the CPU to the GPU. This method not only reduces the computing load on the CPU, but also fully utilizes the GPU computing resources, better balancing the computing load between the CPU and GPU. After completing the preceding optimization, you can use CloudMetrics again to monitor the system resource usage of GROMACS when it is running.

Node Dimension:

7

Metric Dimension:

8

The two figures show the performance data after the PME computing process is offloaded from the CPU to the GPU:

  1. The CPU usage drops between 30% and 35%.
  2. The GPU usage is generally above 70%, with small fluctuations.
  3. The video RAM usage of the GPU increases by more than 50%.
  4. The PCI-E transmission bandwidth of the GPU is above 3 Gbit/s.
  5. While the load remains the same, the execution time is reduced by 50% after optimization.

To sum up, for software that adopts the "CPU + GPU" active/standby acceleration mode, you can balance the computing load between the CPU and GPU to greatly improve the execution efficiency of the software.

Conclusion

CloudMetrics can monitor the usage of cluster resources, such as computing, memory, network, disk, and accelerator card resources, and present the results to users in a systematic manner. Users can conveniently compare and analyze the performance data provided by CloudMetrics to quickly locate the performance bottlenecks and optimize their applications accordingly.

ECS GPU Computing Heterogeneous Computing E-HPC High Performance Computing CloudMetrics GROMACS
1 1 1
Share on

Read previous post:

Optimizing WRF Model with Alibaba Cloud E-HPC CloudMetrics

Read next post:

Building a High-Performance Container Solution with Super Computing Cluster and Singularity

Alibaba Cloud ECS

33 posts | 12 followers

You may also like

Comments

Raja_KT March 13, 2019 at 7:49 am

Interesting... routing CPU-GPU thru PCIe :).... lots of catches

Alibaba Cloud ECS

33 posts | 12 followers

Related Products

More Posts by Alibaba Cloud ECS

See All