Azinta Gpu Cloud Services London Financial Python Ug 1.2

An Azinta Presentation for the London Financial Python User Group 17th January 2011

Azinta Systems Ltd www.azinta.com

17th January 2011 Commercial In Confidence - (c) Azinta 2011

Azinta Systems: ◦ Product and Systems Integration & Consultancy Company

Azinta Solutions: ◦ Developed the APADO Business Agility Platform:

APADO Platform integrates 30 Enterprise Open Source products covering BPM, Business Rules, SOA, CEP, Analytics, Business Constraint Optimisation into a single deployment platform – “Six Levers of Business Agility”.

◦ GPU-Accelerated Analytics Cloud Services Provide GPU acceleration services, for non GPU experts, who want the

benefits of 400x increase in performance without having to become GPU experts.

Azinta Systems – A Protean Corporation ◦ (See book “The Future Arrived Yesterday: the rise of the Protean

Corporation” by Michael Malone) ◦ Azinta are “competence aggregators” who dynamically brings

people together with IP, products, services and expertise to solve business problems


Azinta GPU Cloud Services Comprises:

◦ Dedicated GPU Hosting Service – powered by PEER 1 Hosting GPU Cloud comprising:

Nvidia Tesla M1070 GPUs

Nvidia Tesla M2050 GPUs (Fermi)

Azinta is GOLD partner of PEER 1 Hosting ◦ Python, Matlab, R and Mathematica Acceleration

Accelerate algorithms by up to 1000x. Azinta provides GPU algorithm tuning services

◦ GPU-Accelerated Analytics Services

Conduct analytical and data mining processing on large data sets using GPU acceleration

Reduce analytical data processing from days, hours to minutes, seconds

Azinta provides analytical algorithms tailored for GPUs service; removing the need for domain modellers to be GPU development experts

◦ GPU-Accelerated Derivative Pricing and Valuations

Develop near real-time pricing and risk management models optimised for GPU acceleration

◦ Custom GPU Development Services

Off-shore GPU development services (India + Eastern Europe)

GPU software development services using Python, Matlab, R, C, C** for both Windows and Linux.


One of the Fermi 16 Streaming Multiprocessors (SM)

Fermi delivers 1 teraflop single precision OR 515 double precision gigaflops per GPU

Fermi contains 16 SM: ◦ Each containing 32 cores ◦ Each core contains one integer

and one floating-point maths unit ◦ Each SM can schedule two groups

of 32 threads at once ◦ Each SM contains 4 Special

Function Unit for complex maths such as sine and cosines

◦ Total of 512 cores ◦ 64kb shared L1 cache ◦ All 16 SM can access 768kb L2

cache and can access up to 6GB of GDDR5 memory over 384 bit interface with ECC support

◦ Supports tens of thousands of concurrent threads


Notes ◦ GPU require CPUs where the sequential code is

executed on the CPU and massively parallel processing is executed on the GPU

◦ Data to be processed is transferred from the CPU memory to the GPU GDDR 5 memory

◦ A GPU requires at least one CPU core for execution ◦ There are two main GPUs families:

Nvidia CUDA GPUs – the market leaders with rich ecosystem of development tools + applications

AMD Fusion GPUs – the challengers.

Since there are as yet no AMD GPU cloud services we will focus only on Nvidia CUDA GPUs.


In 2015 Nivida Predicts CUDA GPUs will be: ◦ 20x the performance of the Fermi

◦ 5,000 cores instead of 512 cores in the Fermi

◦ 20 TFLOPS

◦ 1.2 TB/s of memory bandwidth

◦ This GPU is called Maxwell

In 2012 the Kepler GPU will arrive with 4x performance of Fermi


◦ Massive acceleration of Financial, Risk Management and Analytical Applications create high value:

Traditionally GPUs have been used for pricing and valuation of derivatives and to support front and back office operations

However there are many other applications – Namely any application that requires statistical, mathematical and analytical modelling coupled with processing of large data sets can get a significant speed up typically 400x to 1000x or more

There is a link between speed of processing and additional revenues a company can make or to mitigate losses from real-time analysis


◦ For further information on some of the analytical performance improvements that GPU can deliver see my blog post: “GPUs for Large Scale Data Mining” http://goo.gl/dH0i

◦ Also for my vision on how distributed processing across GPU clusters could be implemented see my blog post: “Scaling up GPUs for Big Data Analytics, MapReduce and Fat Nodes” http://goo.gl/jG4Q

◦ /


http://goo.gl/dH0i

http://goo.gl/dH0i

http://goo.gl/jG4Q

http://goo.gl/jG4Q

http://www.smedirector.com/2010/10/16/scaling-up-gpus-big-data-analytics%E2%80%93mapreduce-fat-nodes/

So why Python and GPUs? ◦ Python is very good for Financial and Analytical

modeling and the rapid production of prototypes which can then be deployed into production applications

◦ Developing good GPU kernels is non-trivial exercise and is normally done using C or C++ with a deep knowledge of how massively parallel processing should be implemented on GPUs

◦ What you want is the ability hide the complexity of the GPU world using efficient Python wrapper libraries that will generate the required C code for the GPUs and handle data transfers between the GPU and the CPU.


Many Python CUDA Software options: ◦ PyCUDA http://mathema.tician.de/software/pycuda ◦ Tutorial Slides

http://mathema.tician.de/news.tiker.net/files/main.pdf

◦ PyCUDA documentation http://documen.tician.de/pycuda/

◦ Theano http://deeplearning.net/software/theano/tutorial/using_gpu.html

◦ Ian Ozsvald gave an excellent talk on Python and GPU programming at a recent London Financial Python User Group meeting


http://mathema.tician.de/software/pycuda

http://mathema.tician.de/software/pycuda




http://documen.tician.de/pycuda/

http://documen.tician.de/pycuda/

http://deeplearning.net/software/theano/tutorial/using_gpu.html



Development & Prototyping ◦ Question:

I have a PC that has an Nvidia GPU card is this ok for development and prototyping?

◦ Answer: No because it will be based on earlier versions of the GPU and will not

have all the new hardware and software features of the Fermi. Plus it is not likely to have the on GPU memory you need.

◦ Using a GPU Hosted Service gives you: No need to use out of date GPU development environment. No need to get IT approval to add GPU clusters into your data centre for

your prototyping and development. Get full proactive support so that you do not have to manage and

support the GPU infrastructure. Jump-Start Service Kit to get you up and running on your development

environment. No up-front capital investment… Pay as you go.


Production Deployment: ◦ Question:

I have finished my Python/GPU application and I want to deploy it, but no space in data centre to support new GPU clusters. What are my options?

How can I quickly deploy my production application securely and keeping to EU/Corporate regulations on data storage location?

Also the application requires Windows so CPU attached to the GPUs must support Windows, is this supported?

◦ Answer: ◦ Use a secure fully hosted dedicated GPU clusters, located

in the UK with full proactive support and network security monitoring. Depending on your cloud provider Windows is supported.


NVIDIA Tesla S1070 GPU Cloud

Sever Specs: Prices start at:

Dell R710 Server 4 x Tesla 1070 GPUs 2 x Intel 5520 Quad Core Processors 16GB RAM

$18/GPU/Day $365/GPU/Month

NVIDA Tesla M2050 GPU Cloud

Server Specs Prices start at:

SuperMicro 1U Server 2 Fermi 2050 GPUs 2 x Intel 5520 Quad Core Processors 16 GB RAM


Notes: • Both Windows and Linux Operating Systems supported (Amazon only support Linux) • Supports all Windows CUDA debugging tools (not supported on Amazon) • 2TB data transfer per month provided for free (Amazon charges extra for this) • GPU networks comprising between 100 and 200 GPUs have been configured for clients in

the banking and insurance sectors. • Technical support manager provided (Amazon has a utility support not personalised) • Peer 1 has data centres based in UK and Canada (Amazon only in the USA) • Major customers such as Microsoft, MathWorks, Autodesk, Mental Images, many clients

in banking and insurance. New 55,000 sq. ft. data centre in London opening in 2011 • Many GPU app require full access to CPU power (Amazon only provides VMs)


Cluster GPU Instances Sever Specs: Prices start at:

22 GB of memory 33.5 EC2 Compute Units (2 x Intel Xeon X5570, quad-core “Nehalem” architecture) 2 x NVIDIA Tesla “Fermi” M2050 GPUs


Notes: • Amazon GPU Clusters only support Linux operating system • Amazon GPU Clusters are only available within the USA • Data transfers and inter-node cluster transfers are chargeable at extra costs • No CUDA third-party development and debugging tools such as NVIDIA

Parallel Nsight for Microsoft Visual Studio • No dedicated proactive support manager


Amazon GPU Cluster Offerings: ◦ If your GPU applications are Linux only and you have experience

of using Amazon cloud services ◦ If you have no requirement for EU based GPU clusters EU based

data storage ◦ If you do not need data centre in UK/London to eliminate trans-

Atlantic network latencies.

PEER 1 GPU Cluster Offerings ◦ If your GPU Application require Windows and or Linux ◦ If you require a secure dedicated hosting with firewalls and

network monitoring ◦ If you require a UK based GPU data centre (London data centre

opening 2011) to meet your EU storage requirements and to eliminate trans-Atlantic network latencies

◦ If you want to use development tools such as Nvidia Parallel Nsight for Microsoft Visual Studio and other Windows applications

◦ If you want an assigned support manager for proactive help and support


PEER 1 Hosting: ◦ One of the world’s leading IT hosting providers ◦ Over 10, 000 customers world-wide ◦ 10GB SuperNetwork ™ connected to 17 data

centres, 21 point-of-presence and 10 collocation facilities through-out North America and Europe.

◦ 100% uptime guarantee 24x7x365 ◦ Portfolio includes: Managed Hosting, Dedicated

Servers, Colocation and Cloud Services ◦ Headquartered in Vancouver Canada ◦ Traded on the TSX under the symbol PIX ◦ Website: http://www.peer1hosting.co.uk/


To find out how GPU’s can provide substantial revenue generation or real-time risk mitigation: ◦ Consider implementing a Proof-of-Concept (PoC).

◦ To discuss how Azinta can help you in your project email me at [email protected] or

call me +44 (0) 845 658 6909


mailto:[email protected]

Millennium House

3 Humber Trading Estate

Humber Road

London

NW2 6DW

Tel: +44 (0) 845 658 6909

www.azinta.com

[email protected]


http://www.azinta.com/

Azinta Gpu Cloud Services London Financial Python Ug 1.2

Documents

atlantic network

large data

microsoft

full proactive

nvidia cuda

data centre

prices start

eliminate