Return to flip book view

QuantUniversity Gaining the Technology Edge Five innovations every quant should know I n the last decade technological advances have facilitated major innovations in the financial industry Algorithms running on super fast computers have enabled significant innovations and thus have revolutionized how algorithmic trading and high frequency trading is done Simulation and back testing in risk management are vastly scalable due to innovations in parallel and distributed computing Quants have been in the forefront of leveraging technology to enhance the performance of their models whether it is for speed performance or scalability Every aspiring and practicing quant needs to be aware how technology can enhance their model development and deployment efforts to maximize the performance of their models In the last decade I have worked with more than 25 customers helping them optimize their code to develop robust high performance quantitative models Technology can be a significant enabler and the benefits it can generate can be huge when optimally used While optimizing code is important code that is optimized to leverage the technology available reaps far more benefits Time and again I have realized that knowledge of the technologies and optimal deployment of technology within an organization helps to maximize the return on investment and to vastly improve performance in deployed models Using the appropriate technology for the right use case is important and many a times quants need to redesign their prototype implementations to fully leverage the technology they intend to use In this article I will discuss five technological innovations in the last two decades that have revolutionized quantitative model development and deployment Knowledge of these technologies and optimal deployment can help quants develop efficient models and reap significant benefits and competitive advantage that was just not possible a decade ago 18 The Five Innovations Twenty years ago the internet was in its infancy computers were excruciatingly slow and memory and processing speed in computers were scarce resources The cheapest available computer today at a local electronic store packs more power than some of the fastest available computers available to high end researchers twenty years ago The cost of computing has also dropped significantly enabling more rapid experimentation faster design to deployment cycles and a race to innovate where technology is a key enabler Gone are the days where quants had to run their models and take long coffee breaks or sometimes run models overnight to complete their wilmott magazine

Download

every aspiring and practicing quant needs to be aware how technology can enhance their model development and deployment efforts to maximize the performance of their models wilmott magazine Microsoft Excel or MATLAB can access 2GB limiting the size of datasets that can be used in a quant model By switching to a 64 bit operating system the addressable space goes up to 264 bytes up to 8TB 5 In the past many of my clients running on 32 bit systems had to redesign their code to process data in chunks to address this scalability problem By upgrading to 64 bit operating systems and by adding memory to computers which is very cheap quants can benefit significantly by allowing processing of large datasets However at times even if the operating system is 64 bit the application drivers may still be designed for 32 bit machines These may lead to a restriction of how large addressable memory which in turn means how large the datasets can be despite having significant RAM on the machine Even worse driver compatibility issues 32 bit drivers on a 64 bit machine can result in application errors If you use tools that are built for 32 bit machines check with your vendor IT department on 64 bit support before upgrading your machines Multi core processors Moore s law 1 states that the number of transistors on integrated circuits doubles approximately every two years In the last two decades we have seen Moore s law realized in the processors developed by Intel and AMD with significant increases in processor speed and with multi core experiments The alternative was to make significant assumptions run fewer experiments and extrapolate results which severely compromised the integrity of the models These short cuts resulted in many failed experiments severely limited quants from exploring and trying out new innovations Fast forward 20 years computational hardware is cheap and innovations in hardware and software packages have lowered the bar for technological access My smart phone has a Quad core processor 1 GB of RAM and 32 GB of memory which is much better in specifications than my expensive home computer I bought 15 64 bit systems years ago Super fast computers cheap memory In a recent conference when I polled my audifaster networks and innovative software packence Windows users on how many of them still ages have made it possible to design and develop used 32 bit operating systems around 40 percent sophisticated quantitative models enabling quants raised their hands Most Windows Excel users are to experiment more and focus on building more familiar with the 65 000 row limit or 2GB memrobust models In the last decade I have worked ory limit 4 when running on a 32 bit operating with more than 25 customers and helped them system Many of my clients who run MATLAB optimize their quantitative models In most cases on Windows 32 bit systems have seen Out ofclients approached us with scalability issues perMemory errors when dealing with large datasets formance overheads and at times running the The reason here is when running on a 32 bit opercode was just not practical In more than 60 perating system the theoretical addressable memory cent of the use cases my solution was to redesign available to the application running your code the application to leverage technology to make is 232 bytes 4 GB out of which applications like applications run faster scale better Figure 1 The Five Innovations or to handle use cases that was not feasible without technology Rather than complete redesign testing and model revalidation efforts of just the codebase leveraging technology with minimal code changes was an optimal path In many cases I evaluated the appropriate technology for the use case and illustrated how the technology can be used in the context of my client s application to optimize the performance of their models In the following sections I will discuss the five technologies Figure 1 that I have used time and again for my customers I will introduce and discuss the benefits of each technology and provide pointers and considerations using these technologies for their model development 19

QuantUniversity processors A faster processor means that quants can run their code faster Many times my clients see significant performance improvements by just upgrading to a faster computer In addition quants can leverage multiple processors on their machines to increase scalability This can be accomplished by writing multi threaded programs or by parallelizing their code I will discuss this in more detail later Writing multi threaded programs or parallel programs from scratch isn t trivial Quants must consider whether their program design can truly leverage multi core capabilities They must also consider whether the implementation implicitly leverages multi core processors or should be explicitly implemented to leverage the multiple cores on a machine In addition developers must consider Risk management and option pricing applications involving Monte Carlo simulations can be classic use cases for parallel and distributed computing where each simulation can be run independently Parallel computing enables each task or simulation to be run independently on different cores of a multi core processor This enhances scalability and enables the application to run faster Distributed computing further enables leveraging clusters of machines or grids to run applications significantly faster thus increasing scalability Depending on the type of problem a combination of parallel and distributed computing methodologies can be used 6 to scale the problem implementation Many times I have been able to significantly enhance performance in my client models by leveraging Quants have use cases where the same similar operation needs to be repeated multiple times for example Monte Carlo simulations or stress testing what else would be running on these machines and the resources available to their programs Quants typically use C libraries to write multi threaded programs Boost is an example of one such library which supports multi threading 2 These libraries offer granular control and lets developers tune their applications But developing multi threaded programs is inherently complex and programs must be designed and tested thoroughly to avoid problems like race conditions Higher level languages like MATLAB offer implicit multi threaded support for many functions 3 can also be used to leverage multi core processors The cautionary note here is programs that can leverage multiple cores on a machine may significantly enhance performance and scaling but is complex to design and implement Parallel and Distributed Computing Quants have use cases where the same similar operation needs to be repeated multiple times for example Monte Carlo simulations or stress testing 20 parallel computing and with minimal changes to the program design using MATLAB and associated parallel computing tools 7 Since most computers have multi core processors nowadays it is worth leveraging the available resources when applications are designed to run large number of simulations However quants must evaluate whether their application can truly benefit from parallel and distributed computing methodologies One of the primary considerations is to determine the portion of the application that can be parallelized Since there is overhead involved in passing data and synchronization adequate design and testing must be done to evaluate if the computational benefits supersede the data transfer performance hit In addition different languages support parallel and distributed computing through various APIs Quants must be comfortable designing programs optimally using these APIs to ensure that the computational resources are used optimally Further quants must consider the redesign effort necessary to leverage parallel and distributed computing technologies Though parallel and distributed programming approaches offer huge promise in enhancing scalability a thorough understanding of the methodologies in needed to use these tools properly If you are new to these areas consult someone familiar with these technologies to ensure optimal design and implementation before significant investment of time effort and resources to build a parallel and distributed computing system General purpose computing on graphics processing units GPGPUs The last five years have significant improvements in GPGPU technology 9 and usage in the computational finance GPGPUs or simply GPUs are optimized for massively data parallel operations and have found uses in Monte Carlo simulations option pricing etc Typically applications leveraging GPGPU technology require sophisticated and dedicated GPU hardware NVIDIA and AMD are major vendors who supply specialized GPU cards and associated drivers and API Designing and developing GPU programs are usually non trivial and requires significant understanding of the hardware stack the API and the algorithm design aspects For example NVIDIA provides a CUDA toolkit 8 to leverage their GPU hardware Application design and optimizing code targeted to the hardware architecture in use is essential for well designed GPU programs For example CPUs are designed for low latency access to datasets leveraging the L2 cache GPUs on the other hand have a smaller L2 cache but have a huge number of processors optimized for data parallel tasks Typically the program design involves writing code that transfers data from a CPU context to a GPU context for parallel processing and retrieving the results back to the CPU context for post processing The design also involves leveraging the CPU for massively parallel tasks whereas the CPU is used for other computational tasks MATLAB provides GPU programming support through the Parallel Computing Toolbox 7 Properly optimized code leveraging GPUs can obtain massive scalability on a desktop sized hardware platform that in the past was only feasible through massive compute grids Considering the complexity in design of the code customized for a GPU we recommend quants to work with GPU application developers to minimize the design magazine

QuantUniversity and development lifecycle of the application and to develop optimal GPU code Cloud Computing Cloud Computing has rapidly gained popularity in the last five years Though cloud computing is more prevalent outside of financial services in areas like as e commerce the popularity in the financial industry has been increasing With financial services companies opting for leaner IT infrastructure more and more vendors are offering cloud based solutions and financial companies are experimenting would help quants leverage these massively scalable technologies to develop and try innovative applications providing them with flexibility and access to hardware software choices that were never available to them in the past Conclusion Advances in technology in the last two decades have significantly enhanced the toolsets quants have to develop test and scale innovative quantitative applications Technology has been a significant enabler for quants to develop applications that scale and Technology has been a significant enabler for quants to develop applications that scale and perform in ways that was never possible a decade or two ago with running their applications on the clouds One of the key benefits of cloud computing is its elastic nature and on demand availability of resources that enables companies to scale their problems without having to buying hardware and software resources in house In addition experimentation is easier on the cloud allowing quants to try models that require significant hardware After trying models on the cloud and validating the results organizations can procure dedicated hardware or leverage private cloud services to run their models Many vendors like Amazon 10 offer a variety of hardware and software choices including hardware with GPU cards enabling quants to try out the technologies described above In addition many financial companies are experimenting with Big Data applications for which cloud computing provides the scalable hardware and software resources to run large applications Concerns such as data and application security network overhead etc are still open problems that have hindered rapid adoption of Cloud computing in the financial services industry However with vendors like Google Microsoft Amazon and Oracle making significant investments to address these issues as grow their offerings cloud computing is destined to be the next wave of the future Familiarity with the cloud computing model wilmott magazine perform in ways that was never possible a decade or two ago In this article I have discussed five technological innovations 64 bit systems multi core processors parallel and distributed computing GPGPUs and cloud computing that are being used by quants round the world to gain competitive advantage in the market place I have used these technologies with various clients and my clients have seen significant benefits with these technologies However every opportunity comes with new challenges Each technology I discussed must be thoroughly understood and used appropriately to gain maximum benefit By optimally designing applications to leverage these technologies quants can speed up and advance their innovations and bring their research into the market place References http en wikipedia org wiki Moore 27s_law http www boost org doc libs 1_38_0 doc html thread html http www mathworks com discovery multicore matlab html http msdn microsoft com en us library office ff700514 28v office 14 29 aspx http www mathworks com products matlab preparing for64 bit windows html 6 https en wikipedia org wiki Distributed_computing http www mathworks com products parallel computing http www nvidia com object cuda_home_new html https en wikipedia org wiki GPGPU http aws amazon com ec2 About the Author Sri Krishnamurthy CFA CAP is the founder of QuantUniversity com a data and quantitative analysis company He has significant experience in designing quantitative finance applications for some of the world s largest asset management and financial companies Sri has worked at MathWorks where he worked with more than 25 customers providing asset management energy analytics risk management and trading solutions Prior to that Sri was a consultant at Endeca now Oracle in their Analytics Group and at Citigroup in their Fixed Income Group He has a MS in Computer Science and an MS in Computer Systems Engineering both from Northeastern University and an MBA from Babson College Sri is also an Adjunct Lecturer at Babson College and teaches an in their MBA program He is the author of the forthcoming book published by Wiley titled Financial Application Development with MATLAB He can be reached at sri quantuniversity com 21