Lizard Results

Applies To: Microsoft HPC Pack 2008, Microsoft HPC Pack 2008 R2, Microsoft HPC Pack 2012, Windows HPC Server 2008, Windows HPC Server 2008 R2

When you run Lizard, it generates two sets of results:

  • The best performance value for your HPC cluster in GFLOPS, and a percentage value for the efficiency that was achieved when the best performance was attained. (This information is listed on the last page of Lizard.)

  • A LINPACK Benchmark output file that can be saved to your computer, with a list of the values that were used for the LINPACK parameters when the best performance of your HPC cluster was attained.

This topic provides some information about how to interpret these two sets of results in the following sections:

  • Performance and efficiency results

  • Output file

Performance and efficiency results

The last page of Lizard lists the following values that are explained in this section:

  • The best performance value for your HPC cluster in GFLOPS

  • An efficiency value, as a percentage, for the efficiency that was achieved when the best performance value was attained on your HPC cluster.

Best performance value

The best performance value is the highest performance value that was returned by the LINPACK Benchmark during the tuning process. Note that this value is not necessarily the performance value from the last time that the LINPACK Benchmark was run during the tuning process, but rather, the highest performance value that was returned after consecutive runs of the LINPACK Benchmark. To report the best performance value, Lizard compares all the performance values that were returned after each time the LINPACK Benchmark ran, and selects the highest value.

Efficiency value

The percentage value for the efficiency of your HPC cluster is calculated by comparing the best performance value attained during the tuning process, with the theoretical peak performance value of your cluster. The theoretical peak performance value of your cluster is calculated by multiplying the theoretical peak performance value of each compute node in your cluster (in GFLOPS), by the number of compute nodes that were used during the tuning process.

Because it is not possible to efficiently determine the theoretical peak performance value of each compute node in your cluster, a peak performance value is calculated using the head node computer as a reference. Although it is not ideal, this generic peak performance value is expected to be a good approximation of the actual peak performance value of the least-efficient compute node in your cluster, which should be the baseline for the calculation of the theoretical peak performance value of the whole cluster.

The peak performance value of the head node computer is calculated by using the number of CPUs in the computer, the clock speed of these CPUs, the number of CPU cores in each CPU, and the number of floating-point operations per clock cycle that each CPU core can perform. For this reason, it is important that you provide the correct processor information if you are prompted by Lizard, as explained in Provide Processor Information.

The following steps describe the process that is used to calculate the peak performance value of the head node computer:

  1. First, the peak performance value of each CPU core is calculated by multiplying the number of floating-point operations per clock cycle that each CPU core is capable of performing by the clock speed of the CPU. For example, if each CPU core is capable of performing 4 floating-point operations per clock cycle, and the clock speed of the CPU is 3 GHz, then the expected peak performance value for each CPU core is 12 GFLOPS.

  2. Next, the peak performance value of each CPU is calculated by multiplying the peak performance value of each CPU core, by the number of cores in each CPU.

  3. Finally, the peak performance value of the computer is calculated by multiplying the peak performance value of each CPU, by the number of CPUs in the computer.

Example calculation of an efficiency value

The following is an example of how the efficiency value is calculated for an HPC cluster in which 100 compute nodes were used during the tuning process:

  • The number of floating-point operations per clock cycle for each CPU core in the head node is identified to be 4, and the clock speed of the CPU is identified to be 3 GHz. Therefore, the peak performance value for each CPU core is estimated to be: 4 floating-point operations per clock cycle x 3 billion cycles per second = 12 billion floating-point operations per second, or 12 GFLOPS.

  • The number of CPU cores in each CPU is identified to be 4, and the number of CPUs in the computer is identified to be 2. Therefore, the peak performance value of the computer is estimated to be: 12 GFLOPS for each CPU core x 4 CPU cores in each CPU x 2 CPUs in the computer = 96 GFLOPS

  • The theoretical performance value that should be expected from this HPC cluster is: 96 GFLOPS for each compute node x 100 compute nodes = 9,600 GFLOPS.

  • After Lizard has completed the tuning process, the best performance value obtained was 7,200 GFLOPS. Therefore, the percentage value for the efficiency of this HPC cluster is: (7,200 GFLOPS / 9,600 GFLOPS) x 100 = 75%.

Output file

After Lizard has completed the tuning process, you can save an output file that contains the performance and efficiency results that were attained by your HPC cluster, and the values that were set for each one of the LINPACK Benchmark parameters when the best performance was attained.

This output file can also be imported in the future as a parameter input file when you run Lizard in advanced tuning mode. You can then use the values for the LINPACK Benchmark parameters as the initial values for the new tuning process.

To save the output file

  1. On the Results page of Lizard, click Save Results and Optimal Tuning Parameters. The Save As dialog box appears.

  2. Browse to the folder where you want to save the output file, type a file name for it, and then click Save.

Additional references