Clustered
Multi-Threading: A Platform for Integrated Dynamic Thermal and Reliability
Management
Tuesday April 5, 2005
Hamerschlag Hall D-210
4:00 pm
David
Albonesi
Cornell University
Clustered processors have long been proposed as a more scalable
alternative to monolithic single-threaded cores. In a clustered
processor, the back-end resources are partitioned into several (usually
identical) clusters. The front-end instruction steering mechanism
attempts to balance cluster utilization and inter-cluster communication
in directing instructions to the back ends. Years of tuning these
steering mechanisms has led to cycle-level performance that rivals
monolithic designs. Clustered designs also have the properties of
redundancy and flexibility that makes them highly amenable to microarchitecture-level
dynamic thermal and reliability management.
In this talk, we discuss Clustered Multi-Threading, a multi-threaded
version of a clustered processor, as a potential microarchitecture
for future highly scaled technologies. We discuss how CMT processors
can provide on-demand thermal relief and added robustness, and we
explore several CMT design alternatives, including thread steering
policies, and the partitioning and sharing of the L1 data cache.
CMT is shown to provide competitive cycle-level performance to SMT
with a greatly simplified and more energy-efficient microarchitecture.
Finally, we discuss the potential for CMOS-compatible on-chip optics
to improve interconnect performance in future large scale multi-CMT
designs.
David H. Albonesi is an Associate Professor of Electrical and
Computer Engineering at Cornell University and a member of the Computer
Systems Laboratory. He received his B.S.E.E. from the University
of Massachusetts Amherst in 1982, his M.S.E.E. from Syracuse University
in 1986, and his Ph.D. in Electrical and Computer Engineering from
the University of Massachusetts Amherst in 1996. Prior to receiving
his Ph.D., he held technical and management leadership positions
for 10 years at IBM Corporation (1982-86) and Prime Computer, Incorporated
(1986-1992). The primary focus of his industry work was on the microarchitecture
of low-latency, high-bandwidth memory hierarchies for high performance
processors, the design of shared memory multiprocessor systems,
and the development and application of architectural evaluation
and hardware emulation tools. From 1996-2004, he was a faculty member
in ECE at the University of Rochester, where he led the Complexity-Adaptive
Processing (CAP) project. His current research interests include
adaptive and power-efficient microarchitecture, multithreaded processors,
and power-efficient highly-available systems. Dr. Albonesi received
a National Science Foundation CAREER Award and IBM Faculty Partnership
Awards in 2001, 2002, and 2003. He co-founded the Workshop on Complexity-Effective
Design that has been held the last five years at the International
Symposium on Computer Architecture and will be held again at ISCA
in 2005. He holds seven U.S. patents and is a Senior Member of the
IEEE.
|