Title: IBM System Blue Gene
1IBM System Blue Gene
- Ultrascaling and Power Efficiency for
Computational Advantage in Financial Services
2IBM Systems AgendaA strategic design for
delivering innovative technology, resources and
skills
3IBM System Blue Gene
Optimized for scalability, bandwidth, and massive
data handling while consuming a fraction of the
power and floor space required by todays high
performance systems
- Highest performance supercomputer! (Top500 Nov.
2005) - Ultra scalable performance
- Ultra performance per kW of power
- Ultra floor space density
- Innovative architecture and system design
- Familiar programmer/user environments
Available on demand!
Networks
3D Torus for point-to-point communications
Global Collective Network for oneto-all
communications
Gigabit Ethernet for file I/O, host interface,
control and monitoring
- Up to 131,072 PowerPC CPUs
- Up to 360 peak TeraFLOPS
4Example Value at Risk in near real-timeUsing
Monte Carlo simulation including data
pre-staging, computation, and data post-staging
Value at Risk is a measure of the maximum
potential change in the valueof a financial
portfolio with a given probability over a given
time period
- Value at Risk Calculation
- Generate a set of simulated Market scenariosfrom
the joint distribution of market changes. - Price the portfolio for all scenarios in
thissimulated set. - Estimate the VaR from the empirical
profit-lossdistribution.
5Highly parallelized and efficient I/O-intensive
data file broadcasts
GPFS or Database
Two-phase algorithm for collective I/O
Data File for Broadcast Sequential Read
Phase I A subset of access nodes read
non-overlapping sections of the required file
segment
Access nodes - The number and location of
these compute nodes is customized to maximize
parallelism and avoid I/O contention
Phase II Collective communication distributes
data to remaining compute nodes A2A All to All
exchangeMBC Multiple broadcasts
Blue Gene Torus Network
6Example Variable Annuities Capital Reserve
calculation Using Nested Stochastics for 360 time
steps
Each Blue Gene node performs Simulation Service
for Nested Stochastics
- Receive Time t state.
- Calculate optional flows to horizon along
simulation. - Return result.
- Valuations feed Hedging model for portfolio cash
flows across time and scenarios - Value at Risk for cash flows feeds regulatory
capital model
7Services Implementation
client
scatter
- Initialize
- Broadcast VA data
- Prime each BG node w/ one simulation
- Broadcast Market data
- Calculate Greeks
- single scatter-gather per outer timestep
- simulation services by node
- gather and store greeks by outer timestep
- VA Porfolio
- Market States
gather
8Optimized Analytics Infrastructure
- - Client adapter requests Service
- - Service Mgr starts/stops/queries Service
Instance - Service Instance connects to Client App
9IBM System Blue Gene - Floor Space Efficiency
Advantage vs. Commodity Linux Cluster
Equivalent TFLOPs ¼ floor space ¼ power
consumption
Blue Gene 2 Rack System 11.4 Peak TFLOPs 76.5 sq.
ft. 56 KW Peak Power Consumption
10IBM System Blue Gene Solution - Conclusions
- Blue Gene represents an innovative way to scale
to multi-teraflops capability - Leadership performance and price/performance
- Massive scalability
- Efficient packaging drives low power, cooling and
floor space requirements - Unique in the market for its balance between
massive scale-out capacity and preservation of
familiar programming environments - Blue Gene is applicable to a wide range of
computationally intensive workloads such as
financial instruments pricing, risk analytics,
monte carlo simulations, etc. - Programs are in place to ensure Blue Gene
technology is easily accessible, including
availability in IBMs Deep Computing Capacity on
Demand center - Ongoing Blue Gene RD ensures the vitality of the
offering
11IBM Virtualization Engine Platform
- IBM portfolio of differentiating virtualization
capabilities - Unifying framework for building and managing a
heterogeneous environment - Does not require rip and replace hardware and
software upgrades - Built to be
- Comprehensive
- Open
- Heterogeneous
- Common skills
12Grid Computing Enabling an On Demand IT
Infrastructure
After Grid
Before Grid
Before Grid
- Siloed Architecture
- Higher (capital operational) costs through
limited pooling of IT assets across silos - Challenging cross organization collaboration
- Limited responsiveness due to more manual
scheduling and provisioning
- Virtualized Infrastructure
- Creates a virtual application operating, storage
collaboration environment - Virtualizes application services execution
- Dynamically fulfills requests over a virtual pool
of system resources - Offers an adaptive, self-managed, high
availability operating environment
13MetaCluster Dynamic, application-centric
workload management
Using IBM DirectorsPredictive Failure Analysis
- Optimization of resource usage
- Servers can be consolidated dynamically based on
usage criteria - Optimization of performance
- Applications can be scaled up (moved to a more
potent server) - Optimization of service availability
- Workloads can be moved pro-actively to a healthy
resource, upon detection of degradation of system
resources
PFA
monitoring
app
app
app
app
App
14Why Linux is important to customers
- Linux is an excellent path to On Demand Business
- Integration
- Linux can run alongside the existing operating
system - Linux enables the same application to run on many
different platforms - Virtualization
- Linux supports one-to-many virtualization
software and microcode - Linux supports many-to-one virtualization -
Clusters, Grids - Automation - Linux delivers a robust, manageable
distributed platform - Linux is about choice and flexibility
- Linux is secure
- Linux is reliable
- Linux drives business goals
- Reduce costs
- Simplification
- Improve application service levels
- Promotes innovation
Source IDC Directions 2005
15IBM Systems AgendaA strategic design for
delivering innovative technology, resources and
skills