Benchmark of American Option Pricing Engine

Overview

This is a benchmark of MC (Monte-Carlo) European Engine using the Xilinx Vitis environment to compare with QuantLib. It supports software and hardware emulation as well as running the hardware accelerator on the Alveo U250.

Highlights

The performance of the MCAmericanEngine is shown in the table below, our cold run has 176X and warm run has 529X compared to baseline. Baseline is Quantlib, a Widely Used C++ Open Source Library, running on CPU with 1 thread. The CPU is Intel(R) Xeon(R) CPU E5-2667 v3 @3.200GHz, 8 cores per procssor and 2 threads per core.

Table 20 performance
Platform Execution time
cold run warm run
Baseline 1038.105ms 1038.105ms
Runtime on U250 5.87ms 1.96ms
Accelaration Ratio 176X 529X
Table 21 Application Scenario
Option Type put
strike 40
underlying 36
risk-free rate 6%
volatility 20%
dividend yield 0
maturity 1 year
tolerance 0.02
workload 100 steps, 24576 paths

Profiling

The resource utilization and performance of MCAmericanEngine on U250 FPGA card is listed in the following tables (using viviado 2018.3).

Table 22 Resource utilization report of American Option APIs on U250
Engine APIs LUT FF BRAM URAM DSP
1-kernel ver.
MCAmericanEngine
(UN config: 112)
295301 435249 103 0 923
3-kernel ver.
MCAmericanEnginePreSamples
(UN config: 2)
120756 185169 43 0 416
MCAmericanEngineCalibrate
(UN config: 22)
181793 267405 68 0 462
MCAmericanEnginePricing
(UN config: 4)
251370 368839 71 0 911

Table 27 gives the resource utilization report of four American option APIs. Note that the resource statistics are under specific UN (Unroll Number) configurations. These UN configurations are the templated parameters of the corresponding API.

The complete Vitis demos of American Option Engine are executed with a U250 card on Nimbix. The performance of two demos is listed in Table 23. In this table, kernel execution time and end-to-end execution time (E2E) are calculated. Besides, for end-to-end execution time, since pipeline mode can be used, especially for 3-kernel version American option engine, pipeline mode execution time is also collected.

Table 23 Performance of American Option demos on U250
Engine Configuration(UN) Frequency Execution Time (ms)
kernel E2E E2E pipeline
1-kernel ver. 112 (UN_PATH=1, UN_STEP=1,UN_PRICING=2) 300MHz 7.1 7.25 5.87
3-kernel ver. 2244 (k0:UN=2, k1:UN=2, k2: UN=4, k3: UN=4) 260MHz 5.33 5.45 1.96

We can observe that the 1-kernel version engine may run at the frequency of 300MHz. And the execution time for end-to-end is 7.25ms. However, for pipeline mode, the execution time is 5.87ms. But in fact, only one output data is transferred from device to host. The reason that E2E and E2E pipeline time differentiate so much is that E2E time is the cold run time and the time of E2E pipeline is warm run.

Note

What is cold run and warm run?

  • Cold run means to run the application on board 1 time.
  • Warm run means to run the application multiple times on board. And for pipeline mode, the application is always running multiple times.

For 3-kernel version MCAmerican Engine, in order to maximize the resource utilization on FPGA, the kernel MCAmericanEnginePricing is duplicated. Each Pricing kernel is unrolled with UN 4. Due to place and route on FPGA, the kernel runs at 260MHz finally. However, since the kernel may run in pipeline mode. The final execution time may reach 1.96ms.

Note

Analyzation of the execution time of 3-kernel version American Engine in pipeline mode

The execution time of each kernel is collected and listed in the table below. Because the last kernel takes the longest time while execution, it is duplicated. Which means two MCAmericanEnginePricing kernels, k2 and k3, are included in the final Vitis demo of 3-kernel MCAmerican Engine. The execution time of k2+k3 is 1.56ms. Finally, the end-to-end execution time in pipeline mode is 1.96ms, which is only slightly bigger than the execution time of Pricing kernels.

Table 24 Execution time of each kernel in American Option demo on U250
Kernel Configuration Execution Time - pipeline (ms)
MCAmericanEnginePreSamples UN = 2 1.23
MCAmericanEngineCalibrate Un = 2 1.35
MCAmericanEnginePricing UN = 4, duplicated 1.56

Hint

Why instance two kernels of MCAmericanEnginePricing with UN=4 instead of using one pricing kernel with UN=8?

For one pricing kernel with UN=4, the resource it used is around 74% of one SLR. When UN=8, two SLRs are required to fit 1 kernel. This means that cross SLR for single kernel has to be processed. This may decrease the overall frequency dramatically. In real applications, we are always trying to avoid cross SLR place & routing. Using two exactly the same pricing kernel with UN=4, each kernel can be successfully placed on one SLR. The cross SLR logic can be avoided. Therefore, two pricing kernels with UN=4 are employed in the end.