By Garima Kochhar and Nishanth Dandapanthu

The first half of 2012 has been an exciting time for HPC. Several new technologies have become available with the promise of improved performance and better energy efficiency. These technologies include the Intel microarchitecture codenamed SandyBridge, PCI Gen3 support on the chipset, and Mellanox FDR InfiniBand that can support 56Gbps with its new encoding scheme. Additionally, new servers that support these technologies provide updated BIOS options and more knobs that can be tuned. As engineers, we wanted to quantify these performance claims. How much better performance? How much better power consumption? And how should I tune my system to meet my needs?

In our previous blog [1], we presented initial single server analysis which evaluated a server that includes these new technologies, the Dell 12th Generation PowerEdge R620 server, and quantified the performance advantage provided by this new server over its predecessor.

A cluster-level study is now available for the Dell PowerEdge 12th generation servers. New technologies demand a new server BIOS to support them, and the plethora of available BIOS options can be tuned for performance and energy efficiency. This study quantifies the impact of these BIOS options specific to the HPC domain. Among other things, this paper recommends BIOS settings for certain HPC workloads and also shows the impact of these BIOS options at a cluster level.

Samples of the results are shown below. It is well understood that the Turbo boost feature converts available power into performance. Figure 1 shows the results of the study that quantify exactly how muchperformance improvement is gained by using Turbo Boost, and at the cost of how much extra power for each application. The graph plots the relative energy efficiency of two different BIOS system profiles when Turbo boost is enabled versus disabled. For example, the WRF application has 9% better energy efficiency with Turbo boost disabled when compared to Turbo boost enabled for the DAPC system profile. This improved energy efficiency is made up of two components, a 12% drop in performance with Turbo boost disabled but with a corresponding 24% saving in power! This kind of information can be used to determine how to tune the servers in the data center to meet an organization’s conflicting power consumption and performance goals.

Figure 1. Performance and Energy Efficiency of Turbo Boost

The study analyses several BIOS options in this manner and provides recommendations for different configuration goals. Table 1 below captures the recommended settings.

Table 1. Recommended BIOS settings

* The text in gray indicates options that are pre-selected by the System Profile and cannot be tuned individually. Text in black indicates options that should be explicitly configured.

 

For more details and analysis, the white paper can be accessed here.

For readers interested in tuning their Intel Xeon 5500 and 5600 based systems (architectures code named Nehalem and Westmere), a similar study was done with Dell’s previous 11th generation PowerEdge servers that included recommendations on optimal BIOS settings for several characteristic HPC workloads. Those results can be found at [2].

References:

  1. http://en.community.dell.com/techcenter/high-performance-computing/b/general_hpc/archive/2012/03/06/hpc-performance-on-the-12th-generation-12g-poweredge-servers.aspx
  2. http://content.dell.com/us/en/enterprise/d/business~solutions~whitepapers~en/Documents~11g-optimal-bios-settings-poweredge.pdf.aspx