enhancing pricing performance and quants productivity in a...
TRANSCRIPT
Enhancing pricing performance and Quants productivity in a Cloud based development environment
Nicolas BLANC Software Engineer
Misys at a glance
Who we are
What we offer
A global leader in financial
services software with
4,500+ employees
50+ countries. The broadest and deepest portfolio of financial services software on the
market. Misys solutions cover retail and corporate banking, lending, treasury,
capital markets, investment management and enterprise-wide risk
management Who we serve
2K clients 12 of the top 20
asset managers $ 46 of the world’s
top 50 banks
2 © Misys 2017 9 May, 2017
“We are transforming the global financial services
industry by making financial institutions more
resilient, more efficient and more competitive.”
Goals and challenges
3
• Run current and future pricing models on GPUs to dramatically speed up processing time for complex trading analytics and risk management calculations,
• Make it available in the cloud for a large ecosystem.
Challenges
Our goals
GPU code is complex to write Requires specific skillset Maintenance and extensibility is often difficult Not everybody can access professional grade GPUs
© Misys 2017 9 May, 2017
Write financial GPU code ?
4
Possible solutions
Software engineers Quantitative analysts
Experienced in GPU progamming Less experienced in finance
Experienced in finance Less experienced GPU programming
Each team gains experience in both fields ?
Create an abstraction layer
Migrate legacy code ?
© Misys 2017 9 May, 2017
Fusion parallel processor
Parallel Processing framework concept
5
Software engineers Quantitative analysts
Build a scripting engine Write pricing models Write payoffs
Translate a groovy script to OpenCL / CUDA / java bytecode
Easy to write and maintain Runs fast
© Misys 2017 9 May, 2017
Parallelization
6
Easy to write and maintain Runs fast
• Price a whole portfolio of deals • Price a deal using multiple market data (Montecarlo) • Price a deal at several dates (PFE)
© Misys 2017 9 May, 2017
Where does the workload come from ?
We separate the logic of the script from its parallelization axis
Script example
7
// Retrieve rate
def rate = data0D(“INTEREST_RATE”, [currency1], fixingDate)
// Compute daycount fraction
def daycount = dayCountFraction(effectiveDate, maturityDate, currency1)
// Compute discount factor
def factor = discountFactor(calculationDate(), maturityDate, basis)
// Compute the price
def price = notional * daycount * (rate - fixedRate) * factor
// return price
return price
Script Deal
{
"currency1": "USD",
"currency2": “EUR",
"fixingDate": "2015-03-17",
"effectiveDate": "2015-03-01",
"maturityDate": "2015-03-01",
"basis": "ACTUAL_360",
"notional": 1000000
}
© Misys 2017 9 May, 2017
Deal axis
Script example
8
// Retrieve rate
def rate = data0D(“INTEREST_RATE”, [currency1], fixingDate)
// Compute daycount fraction
def daycount = dayCountFraction(effectiveDate, maturityDate, currency1)
// Compute discount factor
def factor = discountFactor(calculationDate(), maturityDate, basis)
// Compute the price
def price = notional * daycount * (rate - fixedRate) * factor
// return price
return price
Script Deal
{
"currency1": "USD",
"currency2": “EUR",
"fixingDate": "2015-03-17",
"effectiveDate": "2015-03-01",
"maturityDate": "2015-03-01",
"basis": "ACTUAL_360",
"notional": 1000000
}
Market data axis
© Misys 2017 9 May, 2017
Deal axis
Script example
9
// Retrieve rate
def rate = data0D(“INTEREST_RATE”, [currency1], fixingDate)
// Compute daycount fraction
def daycount = dayCountFraction(effectiveDate, maturityDate, currency1)
// Compute discount factor
def factor = discountFactor(calculationDate(), maturityDate, basis)
// Compute the price
def price = notional * daycount * (rate - fixedRate) * factor
// return price
return price
Script Deal
{
"currency1": "USD",
"currency2": “EUR",
"fixingDate": "2015-03-17",
"effectiveDate": "2015-03-01",
"maturityDate": "2015-03-01",
"basis": "ACTUAL_360",
"notional": 1000000
}
Market data axis
© Misys 2017 9 May, 2017
Deal axis Date axis
Script example
10
// Retrieve rate
def rate = data0D(“INTEREST_RATE”, [currency1], fixingDate)
// Compute daycount fraction
def daycount = dayCountFraction(effectiveDate, maturityDate, currency1)
// Compute discount factor
def factor = discountFactor(calculationDate(), maturityDate, basis)
// Compute the price
def price = notional * daycount * (rate - fixedRate) * factor
// return price
return price
Script Deal
{
"currency1": "USD",
"currency2": “EUR",
"fixingDate": "2015-03-17",
"effectiveDate": "2015-03-01",
"maturityDate": "2015-03-01",
"basis": "ACTUAL_360",
"notional": 1000000
}
Market data axis User defined functions
© Misys 2017 9 May, 2017
Deal axis Date axis
Language syntax highlights
11
Standard flow operators for, while, if, else, switch/case, break, continue…
Static type inference No explicit typing, everything is detected and optimized at compile time Natively supported types : Integer, Double, String, Date, Boolean, Array, Matrix, Cubes…
Custom functions Function declaration with typeless parameters
Custom structures Class-like structure definitions, possibility to define functions of structures
Function pointers Through seamless static templating
Standard library A set of optimized standard functions and algorithms provided by default
© Misys 2017 9 May, 2017
Technical insight
12
Compute invariants only once
// Retrieve rate
def rate = data0D(“INTEREST_RATE”, [currency], fixingDate)
// Compute daycount fraction
def daycount = dayCountFraction(effectiveDate, maturityDate, currency)
// Compute discount factor
def factor = discountFactor(calculationDate(), maturityDate, basis)
// Compute the price
def price = notional * daycount * (rate - fixedRate) * factor
// return price
return price
Script Precomputation
Not depending on any market data Not depending on calculation date
daycount is precomputed
notional comes from the deal daycount is precomputed
notional * daycount is precomputed
© Misys 2017 9 May, 2017
The assets of the engine
13
The framework takes care of: • Data alignment / Coalescence
• Memory copy
• Optimized occupancy
• Multi GPU
• Single or double precision
• Separation of CPU / GPU execution through precomputation
• …
Optimizing the engine optimizes all existing scripts
© Misys 2017 9 May, 2017
Going into the Cloud
14
VaaS: Valuation as a Service
© Misys 2017 9 May, 2017
Valuation as a Service
15
Provide a cloud based Integrated development environment Users just need a web browser and internet access. Based on the open source project Eclipse Che.
Runs on AWS servers Uses K80 GPUs, easy to scale up, no need to worry about the drivers. Everything is preconfigured thanks to Docker images.
Rely on collaboration Users can write an upload their codes in order to be used by others. Possibility to get up to date pricing models easily.
© Misys 2017 9 May, 2017
Valuation as a Service
16
© Misys 2017 9 May, 2017
Dispatcher
Master 1
Master 2
Master 3
Proxy
User
User workspace
Valuation as a Service
17
© Misys 2017 9 May, 2017
User 1 Workspace
User 2 Workspace
User 3 Workspace
K80 K80
Schedulor
VMs
Physical
• User workspaces can be hosted on several servers.
• K80 could have been directly used by the VMs, but the schedulor ensures that tasks are done one by one, preventing out of memory for instance.
• The schedulor also has the ability to dispatch work on different servers.
Demonstration
Tools available Debugger, non regression framework
Cloud based IDE Easy to access, easy to use, takes leverage of the community.
Summary
19
Scripting language Easy to code, read and maintain scripts
Same script, multiple use cases Can be used for 3D simulation, Monte carlo pricing, PFE, CVA…
Performance Single or double precision and multiple GPU support
© Misys 2017 9 May, 2017
misys.com Please consider the environment before printing this PowerPoint.
@MisysFS
Misys @ LinkedIn
MisysVideoChannel
Nicolas BLANC Software Engineer
Latest FPP benchmark on AWS EC2 P2 instance
21
© Misys 2017 9 May, 2017
• We are reaching 61M PV/seconds on a p2.8x instance with 8 GPUs
• Test done a 1M Swaps portfolio with the following composition
Maturity Cashflows per trade Trade <5Y 27 3%
10Y<mat<15Y
30 <nb cf <=150 17%
150<nb cf<250 20% 336 3%
15Y<mat<20Y 30<nb cf <=100 27%
200 3%
20Y<mat<25Y 30<nb cf <=100 18%
150<nb cf<290 9%
0
10,000,000
20,000,000
30,000,000
40,000,000
50,000,000
60,000,000
70,000,000
0 1 2 3 4 5 6 7 8 9
PV/s
GPUs
average PV/s