Maxeler Introduction for Telenor Jan 1, 2017
Maxeler Technologies Roots at Stanford University, Bell Labs, and Imperial College London Founded in 2003, incorporated in Delaware and England 2006: starting Finite Difference project with Chevron in San Ramon CA, developing first Dataflow Engine 2009: First Low Latency Trading Appliance deployed with trading customer on the CME Exchange 2011: 20% of Maxeler bought by JP Morgan’s strategic investments group (AFTA award for risk analytics) 2014 Series B: received investment from CME Group, Juniper, Galaxis Capital 2016: Major Partnerships with Amazon AWS, Hitachi Data Systems, see Press Dec 2016.
Industry Consortium: OpenSPL Open Spatial Programming Language: The Concept of Computing in Space! In computing in space, computation takes space, while moving data takes time University Course at Imperial College London: cc.doc.ic.ac.uk/openspl14 Founding Corporations: Founding Academic Partners: http://www.OpenSPL.org launched on Dec 9, 2013
Maxeler Technology and Markets Government ExaScale / PetaScale Face and Object Recognition Machine Learning and Security Science & Engineering Imaging for Oil&Gas Video Encoding, CFD Physics, Chemistry, Engineering Telecom Cloud Computing Deep Packet Inspection Intrusion Detection (e.g. SNORT) Networking (e.g. NFV, packet capture) Storage (e.g. SQL optimization) Finance Exchange Gateways and Matching Risk/Pricing Analytics High Speed Transactions Processing *20 Patents submitted/granted
UK National Scientific Computing Cloud at Daresbury Labs Industry Government Engagement type Hardware and Software Platform Sale Main Contacts Ministry of Science STFC Date Dec 2013 Critical Client Issues ► Competitive advantage in international race ► Transitioning to Big Data Analytics while conventional solutions do not manage to keep up ► High Energy Physics keeps pushing computational demands Approach The approach: Selecting Multiscale Dataflow Computing architecture Selecting Maxeler Software Platforms Training of staff to work with Maxeler technology Support and maintenance of the installation Client benefits ► 20-50x increased compute capability per cubic-foot of data center space => single Maxeler rack brings compute capability of over 20 conventional racks ► Enabling the evaluation of portable Petascale computing systems ► Green computing: chance to beat the top machines in the Green500 supercomputer list ► European Petascale Supercomputing Competition, now in Round III of EU PCP PRACE!
Dataflow Engines as a Cloud Service DFEs Big-Data Analytics Web Browser CPU Network Analysis Python Ethernet, InfiniBand, PCI Express CPU Geophysics Engine Matlab CPU Financial Risk Engine R CPU C/C++ Fully Custom DFE Algo-Trading Platform CPU NEW EC2 F1 INSTANCE COMPATIBLE with MAX5 DFEs 6
Maxeler Intrusion Detection (SNORT) vs Other Top Tier Security Vendor One Maxeler DFE (QFX PFA) Size: 11 Rack Units IPS Throughput: up to 100G PCRE Latency: >50µs Cost: >$1M Size: 1 Rack-Units IPS Throughput: 160G Full Line Rate Analysis PCRE Latency: <10µs Cost: <$100K Juniper QFX Switch with DFE inside
Joint Switch Product with JUNIPER ideal for Deep Packet Inspection Juniper QFX Switch w/ DFE inside Maxeler MPC-N/C Series Four DFEs with up to 12x40G Exchange-in-a-Box Finance Gateways Credit Controls Fast Market Making CyberSecurity Inline Risk Enriched Market Data Packet Capture (MiFID II) Load Balancing Publish/Subscribe Data Bus Fast Web Server
http://appgallery.maxeler.com/
Maxeler Dataflow Engines (DFEs) for Financial Exchanges DFE-based Matching Engines Algos: Price-time Priority, FIFO Features: Straightforward Orders, Volatility Quotes, Spread Quotes Throughput > 1MTPS Latency < 5µs Order Routing DFE Gateways FIFO message ordering Up to 16,384 concurrent TCP connections Order Validation and Pre-Trade Risk Checking Throughput: 10G, Latency < 4µs Market Data DFE Gateways UDP Multicast Feed Generation Standard Feeds - Market By Order, Market by Price Enriched Market Data Feeds Throughput: 10G, Latency < 10µs Real-time Live Risk Pre Trade: Credit Controls DMA Post Trade Clearing House Live Greeks for Trading Desk
2013: Chicago Exchange Gateways
Convolutional Neural Networks Maxeler Dataflow Advantage Xenon E5-2643v3 (6 physical cores) Tensorflow
Multiscale Dataflow Computing Technology
Data-centric Computing in Space Program operations are fixed per stream of data Results flow through the computational graph / network Dataflow Op Dataflow Op Dataflow Op Dataflow Op Input Data Dataflow Op Dataflow Op Perfect Cache Results Dataflow Op Dataflow Op Dataflow Op
Maxeler Dataflow Programming MaxJ – Dataflow programming from Java MaxIDE / WebIDE – Graphical development environments MaxCompilerSim – Fast and accurate application simulation, debug and test environment
Low Latency Application with MaxJ Packet Input 382 parallel operations Simple algorithm Calculates average price over a time-window Software controlled decision parameters Order decision and construction 390.4ns Latency 10Gbps sustained throughput No Jitter Every node operates independently. Like a single CPU instruction – but they all run in parallel Packet Output
The 1U Dataflow Appliance for Cloud / Datacenters Connect DFEs to any number of CPUs MPC-X1000 Ultrahigh Densite Compute 8 DataFlow Engines (DFEs) in 1U 768 GB of DFE RAM Dynamic allocation of DFEs to conventional CPU servers Zero-copy RDMA between CPUs and DFEs over Infiniband Equivalent performance to 20-60 x86 servers …
DFE Advantage: Minimizing Operational Cost of Computing Example Solution 1 50x Speed-up per 1U server node 32 Maxeler Node Solution Equivalent to 1600 CPU-only Nodes $3.2M Operational cost savings over 3 years Example Solution 3 20x Speed-up per 1U server node 50 Maxeler Node Solution Equivalent to 1000 CPU-only Nodes $1.7M Operational cost savings over 3 years Example Solution 2 30x Speed-up per 1U server node 40 Maxeler Node Solution Equivalent to 1200 CPU-only Nodes $1.8M Operational cost savings over 3 years Example Solution 4 40x Speed-up per 1U server node 32 Maxeler Node Solution Equivalent to 1280 CPU-only Nodes $2.6M Operational cost savings over 3 years “In the Cloud, Cost is Everything”
Cooperation Opportunities for Telenor High Performance Computing- Fast and cheap analysis of own in-house big data Deep packet inspection and filtering for intrusion detection Telenor cloud services to external users - distribution of processing power users Much more processing power for much less money! -- CONFIDENTIAL --
Next Steps Operational demo of all functionalities Project definition Value distribution agreement between Telenor and Maxeler, agreement about preferred form of cooperation (joint venture, per usage charge, consulting etc.) Contract and execution -- CONFIDENTIAL --