Investigating the Cloud Computing Business Framework (CCBF) - Modelling and Benchmarking of Financial Assets and job submissions in Clouds

Transcription

1 Investigating the Cloud Computing Business Framework (CCBF) - Modelling and Benchmarking of Financial Assets and job submissions in Clouds Victor Chang and Gary Wills, University of Southampton, UK. David De Roure, University of Oxford and University of Southampton, UK. Clinton Chee, Commonwealth Bank of Australia, Sydney. 15 September 2010, Cardiff

2 Overview Literature reviews. Finance application portability for this demo. Introduce CCBF - Financial Cloud Framework (FCF) and Middleware Framework. Explain Monte Carlo Methods (MCM) and Black Scholes Models (BSM). Codes and experiments used in different clouds. Discussions Conclusion with Q & A.

3 Technical and Business Challenges Literature reviews, technical and business challenges are identified. Technical: Vendors lock-in; Interoperability; Security; Business: Little linkage between quantitative and qualitative cloud business frameworks working in the same domain. Most are either quantitative or qualitative. Not many structured frameworks to measure cloud business performance. Some are in preliminary stage. Application portability from desktops to clouds, and later on between clouds offered by different vendors, is challenging. Why bother portability? Need not rewrite API or recode/redesign software. Save $ & time. Recessions triggered by finance they cannot longer just hold data without sharing. Clouds offer a good platform for joint analysis.

4 Some issues related to Finance Global recessions triggered by Finance is an interdisciplinary research challenge. Not many/few Universities are dealing with Financial cloud. Many financial modelling calculations are on desktops. Grid/HPC approaches: tend to deal with large scale calculations, & are less flexible to deal with fast-paced finance challenges/issues. Finance applications use VBA, VB and Visual C++ etc, and need not use HPC benchmark or HPC languages (other than C++), at least not all the times. Some firms/banks just use Excel or desktop software. Despite there are interests in clouds, they don t know how to do, or have impression it costs a lot of money/effort, or just outsource. Portability for them is a business challenge.

5 Cloud Computing Business Framework (CCBF): Our motivation and approaches CCBF: contains 4 frameworks. Use strategic/top-down approaches to identify problems, and use bottom-up methods to verify (experiments, modelling, simulations, visualisation etc). CCBF includes Middleware Framework (MF) & Financial Cloud Framework (FCF). MF deals with IaaS and PaaS portability. FCF deals with SaaS portability issue. This paper focuses more on FCF, and presents existing problems in Finance; and our proposal, analysis, experimental results, discussions and recommendation.

6 Monte Carlo Methods (MCM) Part 1 MCM, Capital Asset Models and Binomial Model etc are studied. Most commonly used method for finance is MCM, that s why it s chosen. We use MCM for pricing calculation. MATLAB is used due to its ease of use with relatively good speed. While the volatility is known and provided, prices for buy and sale can be calculated. The following code demonstrates calculation of prices. Call prices are for buy and put prices are for sale. > fareastmc [LowerLimit MCPrice UpperLimit] Call Prices: [ ] Put Prices: [ ] MCM application is in the form of a VBA Excel program to calculate the best call and put prices. For example, Spot Price = 100; Strike Price = 105; Volatility = 0.1; Risk free rate = 0.05, Option Maturity = 1; Time steps = 10 and Number of simulations are The VBA Excel program will calculate the best call price as 4.009, and best put price as

7 Monte Carlo Methods (MCM) Part 2 Key advantages: A large number of simulations can be calculated in one go; eg 100,000. MCM have been widely used and proven in many cases. Calculations are close to reality if key data such as volatility, risk free rate etc are known. Speed, accuracy and portability can be demonstrated. See core part of a code. MCM are used in Operational Risk and models are used to simulate the operational risks in Commonwealth Bank, Australia (CBA). MCM simulations are written in Fortran and C#. Such simulations may take several hours or over a day. Results are useful for banks. For our work, we use Variance-Gamma (VG) Processes (a specific technique in MCM) to be used for experiments and benchmark in the clouds. VG processes are suitable in reducing errors apart from fast calculations. Details for VG processes and its experimental results in Public and Private Clouds are in slide 8-10.

8 Black Scholes Model (BSM) Methods such as Fourier series, stochastic volatility &BSM are used for volatility. As a main stream option, BSM is selected for risk analysis. BSM has finite difference equations to approximate derivatives. We write MATLAB code to calculate call price and risk analysis based on BSM, and contain key values such as strike price: the price targeted for sale. upper boundary: the highest possible range a price or risk can reach. risk free rate: interest investors would expect from an absolutely risk-free investment over a period of time. maturity: the loan is due to be repaid on a fixed date. volatility: used to quantify the risk of assets. dividend yield: the return on investment for an asset. asset steps: a specific BSM method called explicit time steps. The more steps, the more accurate the analysis. This allows us to calculate and track call prices if variations for maturity, risk free rate and volatility change. Easy to track volatility for risk analysis if other variables are changed.

9 Experiments in Clouds Part 1 1 Desktop, 1 Public Cloud (Amazon) and 2 Private Clouds (hardware spec refer to our paper) are used for portability run the same program, run simulation/calculation and record down time taken. Pictures of private clouds. We are interested in SaaS (finance application) running on private clouds.

10 Experiments in Clouds Part 2 VG Processes (a technique in MCM) is used. MATLAB 2007 and Octave are used to run 5,000, 10,000 and 15,000 simulations on Clouds. Timing is important to run a large No of simulations! MATLAB runs nearly 5 times faster than Octave, but more expensive and difficult to set up. Table below summarises the timing benchmark result while running the modelling of assets (MoA) code. Number of simulations and time taken (sec) 5,000 10,000 15,000 Desktop Public cloud (large instance) Private cloud (virtual server) Private cloud (rack server)

11 Experiments in Clouds Part 3 time taken (sec) Desktop Private Cloud (virtual server) Private cloud (rack server) This Figure shows the time taken for 5,000, 10,000 and 15,000 simulations with MATLAB 2007 on desktop and 2 private clouds. Public cloud is not available at the time for testing. Private cloud (rack server) needs the least amount of time. MCM simulations (1 unit = 5,000 simulations) time taken (sec) No of steps taken (1 unit = 500 steps) Desktop Public cloud (large instance( Private cloud (virtual server) Private cloud (rack server) This Figure shows the time taken for running 500, 1,000 and 1,500 steps in BSM application on 4 different clouds and 1 desktop. Time series used in BSM can take accommodate up to 1,500 simulations. Private cloud (rack server) has the best hardware configuration with the fastest network speed & significantly high bandwidth, so it runs the fastest.

12 Middleware Framework GridSAM 2.3 GridSAM is widely used in the UK community and is very easy to modify the Job Submission Description Language (JSDL) for job submission. The GridSAM 2.3 contains bespoke application that allows multiple submissions for around 20,000 jobs submitted at each instance. A Java program, Primes, is written to list prime numbers, and is able to submit jobs to Private and Public Clouds. For example: Two JSDL files, primes- 0to10000.jsdl and prime-10001to20000.jsdl are written to send 20,000 jobs, and a primes_file.pl is modified to lists prime numbers between 0 and 20, Time taken (sec) Public cloud Private cloud (virtual server) Private cloud (rack server) Job submission and its status check is used to record time taken No of jobs subm itted (1 unit =5,000 jobs)

13 Three implications for Banking Three implications: Security; Financial Regulations and Portability. Security: Evolving challenges in location of clouds and data privacy. Mechanisms such as X509 vs. single sign-on are discussed. Regulations: Tighter risk management controls. Cloud computation and resources help financial institutions in running more analytical simulations to calculate risks to the client organisations. Portability: letting clients to install their own libraries. Users who run MATLAB only need to install the MATLAB Runtime to run MATLAB executable on Clouds. For financial simulations written in Fortran or C++, users may need Math libraries to be installed in the Clouds. Clouds must facilitate an easy way to install and configure user required libraries, without the need to write additional APIs like several practices do.

14 Conclusion and Future Work MCM, BSM and GridSAM are used to demonstrate how portability, speed, accuracy and reliability can be achieved while moving financial applications from desktops to clouds. This well fits-in an objective in the CCBF to allow portability on top of, secure, fast, accurate and reliable clouds for IaaS, PaaS and SaaS. Our research is not intended to test pure computing performance of Cloud Computing. The intention is to port and test financial applications to run on the Cloud. It is not yet critical to use Dongarra's netlib functions for benchmark, but there are plans to use HPC languages such as (Visual) C++ & test on 6-core processing private clouds for next stage. We have Health Cloud Storage Framework (HCSF) that can demonstrate portability in details. We hope to strengthen our results and extents of collaboration (such as financial services and non-uk governments).

15 An example in Private Clod: 6-core & 4 terabytes Upgradable to 16 GB DDR3, 8-12 TB per server. Cost effective with high performance capabilities. Better performances in MCM and BSM financial applications.

16 Collaboration and Questions? Other collaborators: 2 anonymous government department branches, 2 anonymous investment banks and 3 anonymous small/medium enterprises. Thank you!

17 Appendix: An example in Monte Carlo Methods (MCM) %%%%% Input parameters %%%%%%%% S=100; %spot price K=105; %strike T=1; %maturity r=0.05; %interest rate sigma=0.25; %volatility NSimulations=100000; %no of monte carlo simulations NSteps=100; %no of time steps %%%%%%%%%%%%%%%%% %%%%%%%%%%%%%% dt=t/(nsteps-1); vsqrdt=sigma*dt^0.5; drift=(r-(sigma^2)/2)*dt; x=randn(nsimulations,nsteps); Smat=zeros(NSimulations,NStep s); Smat(:,1)=S; for i=2:nsteps, Smat(:,i)=Smat(:,i- 1).*exp(drift+vsqrdt*x(:,i)); end %************* calculate call price *********** callpayoffvec=max(averagesvec-k,0); mccallprice=exp(-r*t)*mean(callpayoffvec); std_dev_call=(var(callpayoffvec))^0.5; %calculate bounds of expected value assuming 95% confidence level mclimit_call=1.96*std_dev_call/nsimulations^0.5; call_lowerlimit=mccallprice-mclimit_call; call_upperlimit=mccallprice+mclimit_call; printf("call Prices : [%f %f %f]\n",call_lowerlimit,mccallprice,call_upperlimit); %************* calculate put price *********** putpayoffvec=max(k-averagesvec,0); mcputprice=exp(-r*t)*mean(putpayoffvec); std_dev_put=(var(putpayoffvec))^0.5; %calculate bounds of expected value assuming 95% confidence level mclimit_put=1.96*std_dev_put/nsimulations^0.5; put_lowerlimit=mcputprice-mclimit_put; put_upperlimit=mcputprice+mclimit_put; printf("put Prices : [%f %f %f]\n",put_lowerlimit,mcputprice,put_upperlimit);

18 Appendix: An example in Black Scholes Model (BSM) % Enter input Parameters S=100; %Spot price K=100; %Strike UB=115; %upper boundary r=0.15; %risk free rate T=.5; %maturity sigma=0.2; %volatility D=0.05; %dividend yield assetsteps=150; %Explicit scheme time steps % dt=1/(sigma*assetsteps)^2; timesteps=ceil(t/dt)+1; Su=UB; ds=su/assetsteps; dt=t/timesteps; for k=1:timesteps, V(2:assetsteps,k+1)=Ai.*V(1:assetsteps- 1,k)+(1+Bi).*V(2:assetsteps,k)+Ci.*V(3:assetsteps +1,k); %Apply lower boundary condition V(1,k+1)=0; %apply upper boundary condition V(assetsteps+1,k+1)=0; end V2=fliplr(V); OptionPriceVec=V2(:,1); option_price=interp1(svec,optionpricevec,s) V=zeros(assetsteps+1,timesteps+1); Svec=0:ds:Su; Svec=Svec'; V(:,1)=max((Svec-K).*(Svec<UB),0); ivec=1:assetsteps-1; ivec=ivec'; Ai=0.5*(sigma*sigma*ivec.*ivec-(r-D)*ivec)*dt; Bi=-(sigma*sigma*ivec.*ivec+ r)*dt; Ci=0.5*(sigma*sigma*ivec.*ivec+(r-D)*ivec)*dt;