Red Hat and Partners Deliver New Performance Records on Prominent  Risk Analytics Benchmark

While performance benchmarks are often application or industry specific they can also provide useful insights that are widely applicable. Risk analytics applications used in financial services industries have performance characteristics similar to many technical computing applications. These applications are large, compute intensive, and take full advantage of parallel processing and compute accelerators.

STAC®, the Securities Technology Analysis Center LLC (www.STACresearch.com), provides technology research and testing tools including benchmarks for measuring system performance on financial workloads. The STAC-A2 benchmark focuses on risk analytics, a large compute intensive workload based on partial differential equations.

While the STAC benchmarks focus on financial applications, these applications are similar to other large, compute intensive applications: they make heavy use of mathematical operations including floating point computation, square roots, exponential and logarithmic calculations. They rely heavily on large matrix operations which stress both memory and compute subsystems in a server. Various tests measure the time to complete a test or the number of test cases that can be completed in a given time. All in all, the STAC benchmarks provide good insight into the technical computing power of a system and a direct comparison between systems at the application level.

A system put together by HPE, NVIDIA and Red Hat set several new records running the STAC-A2 benchmark. Tests were run on an HPE ProLiant XL270d Gen9 server running Red Hat Enterprise Linux 7.4 with 8 x NVIDIA Tesla V100 (Volta) GPUs. The benchmark results are available at http://www.STACresearch.com/news/2017/10/31/NVDA171020 (free registration required to access the full report).

This benchmark highlights several things:

  • The latest generation of NVIDIA GPUs, the Volta V100, running with Red Hat Enterprise Linux on an HPE ProLiant server.
  • Multiple GPUs on a single system – 8 in this benchmark – applied to a single problem.
  • Applications using the NVIDIA CUDA 9 ecosystem with Red Hat Enterprise Linux.
  • Performance leadership achieved through a joint effort by NVIDIA, HPE, and Red Hat.
  • The combination of the enterprise stability and robustness of Red Hat Enterprise Linux with the latest technologies from NVIDIA and HPE.

Summary of Results:

  • Compared to all other publicly reported results to date, this solution set new records in multiple performance benchmarks as well as the energy efficiency benchmark.
  • Compared to all publicly reported results to date on non-NVIDIA based architectures, this solution was:
    • 8.9x the next best throughput (STAC-A2.β2.HPORTFOLIO.SPEED)
    • 6.2x the next best time in warm runs of the baseline Greeks benchmark (STAC-A2.β2.GREEKS.TIME.WARM).
    • 2.7x the next best energy efficiency (STAC-A2.β2.HPORTFOLIO.ENERG_EFF)
    • 1.9x the maximum basket size (STAC-A2.β2.GREEKS.MAX_ASSETS)
    • 1.5x the next best space efficiency (STAC-A2.β2.HPORTFOLIO.SPACE_EFF)
  • Compared to the best performing solution to date using 4 previous-generation NVIDIA Tesla P100 GPUs, this solution with 8 NVIDIA Volta V100 GPUs in an HPE server was:
    • 2.7x the next best throughput (STAC-A2.β2.HPORTFOLIO.SPEED)
    • 2.4x the next best time in warm runs of the baseline Greeks benchmark (STAC-A2.β2.GREEKS.TIME.WARM).
    • 1.5x the maximum basket size (STAC-A2.β2.GREEKS.MAX_ASSETS)
    • 1.2x the next best space efficiency (STAC-A2.β2.HPORTFOLIO.SPACE_EFF)

Full results are available at www.STACresearch.com/news/2017/10/31/NVDA171020

STAC-A2 Benchmark:

The STAC-A2 benchmarks are focused on risk analysis. The algorithms used in market risk analysis are typically a partial differential equation that can not be solved exactly, thus requiring approximation techniques. One of the most popular approaches is to use the Monte Carlo method which runs a large number of randomized simulations. Each simulation calculates a possible outcome based on the randomized input values. The quality of the simulation depends on the accuracy of each calculation and the number of simulations run – more simulations with greater accuracy produces better results, at the cost of requiring more time or more compute resources.

As in many other simulations, greater compute performance allows you to run the same simulation in less time, run more simulations in the same time, or run more accurate simulations.

The STAC benchmarks stress floating point performance, memory performance, and parallel processing capabilities. These characteristics are common to numerical simulations in many industries, not just finance.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s