STAC Report: Intel 1U server under STAC-A3 (strategy backtesting)

Multiple records in performance and storage efficiency

3 June 2020

STAC recently performed STAC-A3 benchmarks on a stack consisting of the STAC-A3 Pack for Intel® Parallel Studio XE (Rev A) over Red Hat Enterprise Linux 7.8 with 2 x Intel® Xeon® Platinum 9242 Processors in an Intel® Server System S9200WK SDP with 2 x 1.6TB Intel® DC P3700 SSDs for application data. Taking half of a 2U chassis, this solution is considered to occupy 1 rack unit. The STAC Report is available here.

STAC-A3 simulates workloads common in the refinement and backtesting of trading strategies. As a rate-limiting step in a firm's response to changing market conditions, the performance of backtesting infrastructure has a top-line impact. Like other STAC Benchmarks, STAC-A3 is agnostic to architecture.

Of the many results in the report, Intel chose to highlight the following:

• Compared to a solution involving 16 GPUs (SUT ID NVDA190425), this solution showed:
• 11.4x speedup for 1,000 simulations and 50 instruments*
• 7.6x speedup for 10,000 simulations and 50 instruments*
• Processing of 100,000 simulations on 50 instruments in less time than NVDA190425 processed 10,000 simulations on 48 instruments.
• 3.1 x the storage efficiency
• Results for over 10x the number of instruments (1,024 vs 96)

• Compared to the previous best solution not involving GPUs but involving 6 servers and an all-flash storage array (SUT ID LEVX180608), this single-server solution showed:
• 4,500x the space efficiency (STAC-A3.β1.SWEEP.SPACE_EFF)
• 1,563x speedup in max simulations for a basket of 50 instruments in 60 minutes (STAC-A3.β1.SWEEP.MAX60)
• 268x to 752x speedup in every other named throughput benchmark
• 211x speedup for 500 instruments**
• 171x speedup for 1000 instruments***
• More than 5x the storage efficiency

• Measure of linearity in throughput (instrument simulations per second):
• At 10,000 simulations per instrument, throughput varied by only 11% while increasing the number of instruments by 20x, from 50 to 1,024.
• At 100,000 simulations per instrument, throughput varied by only 3% while increasing the number of symbols by 10x, from 50 to 500.

• This SUT set a new record in storage efficiency (STAC-A3.β1.SWEEP.STORAGE.EFF)
• * Comparing the 50-instrument results from this SUT with the 48-instrument results of NVDA190425.
• ** Comparing the 500-instrument results from this SUT with to the 480-instrument results from LEVX180608.
• *** Comparing the 1,024-instrument results from this SUT with to the 960-instrument results from LEVX180608.

For details, please see the report at the link above. Firms with subscriptions to the Analytics STAC Track also have access to the configuration details, implementation code, and test-harness software. To learn about subscription options, please visit us in the STAC booth or contact us.