part one, I introduced the two most popular HPC networking technologies — Ethernet (GigE and 10GigE) and Infiniband. We also compared latency, bandwidth, and the N/2 performance of these technologies. While these numbers give a general feel for performance, there is no easy way to determine the actual performance of your application.

" />
x
Loading
 Loading
Hello, Guest | Login | Register
Today's HPC Clusters Resource Center

The Network IS the Cluster: Infiniband and Ethernet Network Fabric Solutions for HPC (Part Two)

In part one, I introduced the two most popular HPC networking technologies — Ethernet (GigE and 10GigE) and Infiniband. We also compared latency, bandwidth, and the N/2 performance of these technologies. While these numbers give a general feel for performance, there is no easy way to determine the actual performance of your application.

In part one, I introduced the two most popular HPC networking technologies — Ethernet (GigE and 10GigE) and Infiniband. We also compared latency, bandwidth, and the N/2 performance of these technologies. While these numbers give a general feel for performance, there is no easy way to determine the actual performance of your application. Benchmarking your application suite is really the only way to know for sure if a networking technology works best for your needs. Other issues, such as drivers, manageability, infrastructure, etc. may also be important in your choice.

In order to look at actual performance, I chose two popular commercial applications with published benchmark sets and results. As with most things HPC, the results are never quite what you would expect.

Application Performance

Now that we know the performance of the various interconnects (using micro-benchmarks) and the approximate per port or per node costs, let’s take a look at real application performance and see how well the interconnects perform.

I’m going to look at two Computer Aided Engineering (CAE) applications: Fluent and LS-Dyna. Fluent is a Computational Fluid Dynamics (CFD) application used in many industry sectors. LS-Dyna is what some people call a crash or impact code because it can model objects impacting each other using Finite Elements. It also has some other capabilities, but the impact analysis seems to have gained most notoriety. Who doesn’t like to see things smash together?

In the case of Fluent, the parent company ANSYS, provides benchmark examples and allows…

Please log in to view this content.

Not Yet a Member?

Register with LinuxMagazine.com and get free access to the entire archive, including:

  • Hands-on Content
  • White Papers
  • Community Features
  • And more.
Already a Member?
Log in!
Username

Password

Remember me

Forgotten your password?
Forgotten your username?
Read More
  1. Cluster 3.0: Dynamic Provisioning with MOAB and XCAT
  2. InfiniBand Interconnects for Computing Clusters
  3. Optimizing Performance for HPC: Part 2 - Interconnect with InfiniBand
  4. Optimizing the Nehalem for HPC
  5. Sledgehammer HPC