Cloud Computing: High Performance Computing (HPC)

Proprietary Trading Simulation

Overview:

Solid Logic created the IT infrastructure to test an equity trading system across approximately 7.4 Billion data points. Based on internal estimates, the tests would have taken  1-2 months using existing infrastructure. Using our expertise, we were able to complete the test in 45 hours, at a cost savings of $231k (> 99%) vs purchasing in-house hardware.

Business Challenge:

  • Characterize the performance and sensitivity of an equity trading system across input parameters and market conditions
  • Optimize parameters based on profit and risk measures
  • Estimated runtime is unacceptable on local workstation (>1 month)
  • Primary bottlenecks are in dense linear algebra operations
  • Spectral decomposition (ARPACK)
  • Pairwise comparison of higher-order distribution moments (M-M arithmetic)

Problem Scope:

  • Assets – 62
  • Tests/asset – 96
  • Total tests – 5,952

 Test Information:

  • Mean components/asset – 395
  • Points/component – 3,135
  • Points/test – 1,238,325
  •  Total elements   7,370,510,400

Potential Solutions:

  • Run on existing hardware – wait for results
  • Physical or virtualized servers with supporting job schedulers – requires hardware, software, and specialized labor
  • Setup cloud infrastructure to process work – requires software and specialized labor

Chosen Technology Solution:

  • Built an optimized simulation environment as virtual image  (AWS EC2 AMI)
  • Provisioned and configured centralized storage (AWS S3)
  • Experiment configuration
  • Simulation input
  • Simulation output
  • Post-processed results
  • Fully automated deployment of simulation to instances through master source control system (git)

Compute Instance (x16):

  • 88 Elastic Compute Units (ECU)
  • 2x Xeon E5-2670s-16 cores
  • 60.5GB RAM
  • 10GbE, dual NIC
  • 3+TB instance scratch

Total Compute Resources:

  • 1408 ECUs (1 ECU~=5GFlops)
  • 512 concurrent threads (HT)
  • 968GB RAM

Run Time:

  • Cloud: 45 hours
  • Single-seat: 1-2 months
  • Order of magnitude improvement in time!

Cost Savings:

Cost Model Details

  • Cost estimates using assumptions and calculations in the AWS Cost Comparison Worksheet

    • Costs represent one year annualized costs. Assumes a useful life of three years for purchased equipment
    • 1 = Cost savings using On-Site as baseline
    • 2 = On-Site and Co-Location assume 100% usage
    • 3 = Based on actual 686 machine hours used