|
Dr. Phil Heidelberger
IBM Watson Research Center
https://researcher.ibm.com/researcher/view.php?person=us-philiph
Applying Advanced Simulation Methodologies to Supercomputer Design:
A “Pure” Operations Researcher’s Downward Path
 This
talk gives an overview of the IBM Blue Gene family of supercomputers
and describes the role that several advanced simulation methodologies
played in their design. Blue Gene supercomputers are low power and
massively parallel, with up to 100,000 nodes. Each node is a system-on-a-chip
consisting of multiple low power processor cores, multiple levels
of cache memory, interconnection networks and network interface
logic, all integrated onto a single chip. This level of integration
enhances reliability and permits a high level of compute density
with 1024 nodes in a rack.
To architect the network, a near
cycle accurate parallel discrete event simulation model of the network
was developed and used in a production manner. For a full-sized
system, the model can be viewed as a queuing network with over six
million resources. Because of the scale of the network, and the
large memory footprint implied by such a model, parallel simulation
is the only practical approach for conducting meaningful performance
studies. The talk describes the simulator, gives several performance
tradeoff examples and describes validation against measurements
on the real hardware.
Concepts from rare event simulation
estimation were also used in the logic verification of the network,
i.e., in verifying that the hardware logic always performs correctly.
In rare event simulation an importance sampling distribution is
selected to move the simulation towards the rare event of interest
and the output is multiplied by a likelihood ratio in order to obtain
an unbiased estimate of the probability of the rare event. In logic
simulation there are a multitude of rare events or "corners",
many of which are unknown a priori, and the goal is to reach all
of them. No likelihood ratio is required, but the analog of appropriate
importance sampling distributions need to be selected so as to reach
all the corners. The challenges and approaches to designing effective
logic verification simulations will be described.
Philip Heidelberger received his
B.A. in Mathematics from Oberlin College in 1974 and his Ph.D. in
Operations Research from Stanford University in 1978. He has been
a Research Staff Member at the IBM T.J. Watson Research Center in
Yorktown Heights, New York since 1978. Prior to 2000, his research
was primarily focused on developing highly efficient algorithms
for rare event and parallel discrete event simulations. Since 2000,
he has been a member of IBM Research's Blue Gene supercomputer hardware
team where he was initially responsible for developing a parallel,
cycle-accurate simulation model of the first generation Blue Gene/L
interconnection network that was used as the basis for network architectural
decisions. He has been involved in many aspects of Blue Gene, including
logic verification, hardware bringup, network software interfaces
and efficient communications algorithms. He was a lead architect
of the second generation Blue Gene/P Direct Memory Access device
that interfaces to the network and currently leads the group focused
on future Blue Gene networks and their memory interfaces.
Dr. Heidelberger has co-authored
over 115 papers, seven of which have won outstanding paper awards
including the 2006 ACM/IEEE Gordon Bell Prize. He was an Associate
Editor of Operations Research (1983-1990), Editor-in-Chief of the
ACM Transactions on Modeling and Computer Simulation (1996-1997),
Program Chair of the 1988 Winter Simulation Conference, Program
Co-Chair of the 1992 ACM Sigmetrics Conference, General Chair of
the Sigmetrics/Performance 2001 Conference, and Vice-President of
ACM Sigmetrics (2004-2007). He is a co-inventor on 30 U.S. Patents
and has received seven IBM Outstanding Technical Achievement Awards.
He is a Fellow of the IEEE and ACM.
|
Prof. Paul Kleindorfer,
Paul Dubrule Professor of Sustainable Development, INSEAD
Anheuser-Busch Professor Emeritus of Management Science, Wharton
School, University of Pennsylvania
From Back of the Envelope to Large-scale Simulation: Public
Policy Evaluation and Support in Complex Domains
 This
talk is concerned with the use of large-scale simulation models
in support of public policy evaluation. The use of large-scale simulation
typically occurs in contexts where fundamental changes are regarded
as required, but where the irreversibility of policy moves or the
complexity of the interacting organizational and economic drivers
make ex ante evaluation of alternatives a prudent necessity. I will
first discuss three case examples I have been involved with in the
past decade to illustrate the challenges. These case studies come
from the market opening of postal and delivery markets in the European
Union (EU), designing risk transfer instruments for catastrophe
risks from natural hazards in global risk markets, and conversion
of commercial fleets to electric and hybrid vehicles as part of
current moves to develop sustainable transportation models with
low-carbon emissions.
I will use these examples to highlight
several crucial themes now emerging in the design and validation
of large-scale simulation models when these models are used in support
of public policy. The crucial themes include integrating validation
across multiple stakeholders (the EU postal market required integration
across representatives from all 27 EU Member States); legitimation
in the face of epistemic/knowledge risks (typical in catastrophe
models for climate change risks); and enabling collaborative risk
sharing in complex environments (for conversion to low-carbon fleet
operations, the required collaboration is between fleet operators,
the government, electricity suppliers and auto makers).
These case studies reflect
the important contributions simulation has been making to the policy
arena over the past several decades. They also highlight the ever
present problems of legitimation and validation that are important
for all simulation studies, but that take on a special character
when these intersect with public policy choices. Research challenges
for the simulation community involved in public policy analysis
and support will conclude my talk.
Dr. Kleindorfer is Paul Dubrule
Professor of Sustainable Development & Distinguished Research
Professor in Technology and Operations Management at INSEAD. He
is also the Anheuser-Busch Professor Emeritus of Management Science
at the Wharton School of the University of Pennsylvania. Dr. Kleindorfer
was a faculty member at Wharton from 1973 to 2006. He graduated
with distinction (B.S.) from the U. S. Naval Academy in 1961. He
studied on a Fulbright Fellowship in Mathematics at the University
of Tübingen, Germany (1964/65), followed by doctoral studies
at Carnegie Mellon University, from which he received his Ph.D.
in 1970 in Systems and Communication Sciences at the Graduate School
of Industrial Administration. Before joining INSEAD in 2006, Dr.
Kleindorfer held university appointments at Carnegie Mellon University
(l968/9), Massachusetts Institute of Technology (1969/72), The Wharton
School (1973 - 2006), and several universities and international
research institutes, including the University of Frankfurt, INSEAD,
Ulm University, IIASA and The Science Center (Berlin). Dr. Kleindorfer
has published over 30 books and many research papers in the areas
of risk management, managerial economics and management science.
His recent book, The Network Challenge, co-authored with Jerry Wind
(Wharton Publishing, 2009) examines global interdependencies arising
from the network economy. Dr. Kleindorfer's current research is
on risk management, with applications to supply management, pricing
and investment in network industries, and sustainability issues
related to climate change.
|
