1. Introduction
The architecture domain is unique in its ability to use
runtime knowledge of application behavior and the thermal
status of the chip to control execution rate, distribute
the workload, and extract instruction-level parallelism
(ILP). On-chip temperature sensors can provide information
about local hot spots and temperature gradients that
can be combined with dynamic information about ILP in
order to precisely regulate temperature while minimizing
performance loss.
The main focus of research on temperature-aware architecture
to date has been on dynamic thermal management
or DTM. DTM recognizes that if the thermal package
is designed for worst-case power dissipation, it must
be designed for the most severe hot spot that could potentially
arise. Yet these worst-case scenarios are rare and
lead to over-design. A less expensive package can be employed,
designed for the worst “typical” or “interesting”
workload. Excessive heat dissipation must then be handled
by an autonomous, runtime response in the chip itself
that guarantees to reduce power densities far enough and
fast enough to maintain temperature regulation. For example,
the Intel Pentium 4 [ 121 follows this approach, designing
the thermal package for a power density 20% less
than the theoretical worst case. If environmental or workload
characteristics cause temperatures to rise too close to
the maximum allowable, the Pentium 4 responds by stopping
the clock, and hence all dynamic power, until temperature
returns to a safe level [15]. As another example,
the Transmeta Crusoe uses dynamic voltage scaling to reduce
power density when temperatures rise too high [lo].
Yet current techniques such as these are too simplistic
and impose unnecessary performance penalties. Architectural
design can take advantage of instruction-level parallelism
to implement DTM with substantially less performance
overhead. This paper describes three techniques,
“temperature-tracking dynamic frequency scaling,” “migrating
computation” [29], and “hybrid DTM” [26], that
reduce overhead by 25-35% compared to techniques like
the Pentium 4’s and the Transmeta Crusoe’s.
Recent work in the architecture community [2, 4, 13,
14, 19, 21, 23, 27, 311 demonstrates growing interest in
thermal management and shows that architectural techniques
offers substantial benefits. Accurately characterizing
thermal behavior and evaluating architectural techniques
requires an appropriate thermal model. An effective
model for architects must be simple enough to allow
them to reason about thermal effects and tradeoffs; parameterized;
capable of modeling runtime changes in temperature
within different functional units on the die; and yet
computationally efficient and portable.
We have developed and publicly released a dynamic
compact model-HotSpot [29]-that achieves all of these
requirements. It differs from prior dynamic compact models
for computer chips in that this new model requires less
design detail and hence can be used earlier in the design
cycle. It has been validated both against an independent
finite-element model and a test chip. It also accounts for
imprecision due to sensor noise and placement.
The rest of this paper is organized as follows. The next
section describes some example architectural thermalmanagement
techniques and presents preliminary results
evaluating their potential benefits. Section 3 describes the
proposed modeling approach. Then Section 4 describes
what we see as some important research questions in this
area, and Section 5 concludes the paper.
1. IntroductionThe architecture domain is unique in its ability to useruntime knowledge of application behavior and the thermalstatus of the chip to control execution rate, distributethe workload, and extract instruction-level parallelism(ILP). On-chip temperature sensors can provide informationabout local hot spots and temperature gradients thatcan be combined with dynamic information about ILP inorder to precisely regulate temperature while minimizingperformance loss.The main focus of research on temperature-aware architectureto date has been on dynamic thermal managementor DTM. DTM recognizes that if the thermal packageis designed for worst-case power dissipation, it mustbe designed for the most severe hot spot that could potentiallyarise. Yet these worst-case scenarios are rare andlead to over-design. A less expensive package can be employed,designed for the worst “typical” or “interesting”workload. Excessive heat dissipation must then be handledby an autonomous, runtime response in the chip itselfthat guarantees to reduce power densities far enough andfast enough to maintain temperature regulation. For example,the Intel Pentium 4 [ 121 follows this approach, designingthe thermal package for a power density 20% lessthan the theoretical worst case. If environmental or workloadcharacteristics cause temperatures to rise too close tothe maximum allowable, the Pentium 4 responds by stoppingthe clock, and hence all dynamic power, until temperaturereturns to a safe level [15]. As another example,the Transmeta Crusoe uses dynamic voltage scaling to reducepower density when temperatures rise too high [lo].Yet current techniques such as these are too simplisticand impose unnecessary performance penalties. Architecturaldesign can take advantage of instruction-level parallelismto implement DTM with substantially less performanceoverhead. This paper describes three techniques,“temperature-tracking dynamic frequency scaling,” “migratingcomputation” [29], and “hybrid DTM” [26], thatreduce overhead by 25-35% compared to techniques likethe Pentium 4’s and the Transmeta Crusoe’s.Recent work in the architecture community [2, 4, 13,14, 19, 21, 23, 27, 311 demonstrates growing interest inthermal management and shows that architectural techniquesoffers substantial benefits. Accurately characterizingthermal behavior and evaluating architectural techniquesrequires an appropriate thermal model. An effectivemodel for architects must be simple enough to allowthem to reason about thermal effects and tradeoffs; parameterized;capable of modeling runtime changes in temperaturewithin different functional units on the die; and yetcomputationally efficient and portable.We have developed and publicly released a dynamiccompact model-HotSpot [29]-that achieves all of theserequirements. It differs from prior dynamic compact modelsfor computer chips in that this new model requires lessdesign detail and hence can be used earlier in the designcycle. It has been validated both against an independentfinite-element model and a test chip. It also accounts forimprecision due to sensor noise and placement.The rest of this paper is organized as follows. The nextsection describes some example architectural thermalmanagementtechniques and presents preliminary resultsevaluating their potential benefits. Section 3 describes theproposed modeling approach. Then Section 4 describeswhat we see as some important research questions in thisarea, and Section 5 concludes the paper.
การแปล กรุณารอสักครู่..

1. Introduction
The architecture domain is unique in its ability to use
runtime knowledge of application behavior and the thermal
status of the chip to control execution rate, distribute
the workload, and extract instruction-level parallelism
(ILP). On-chip temperature sensors can provide information
about local hot spots and temperature gradients that
can be combined with dynamic information about ILP in
order to precisely regulate temperature while minimizing
performance loss.
The main focus of research on temperature-aware architecture
to date has been on dynamic thermal management
or DTM. DTM recognizes that if the thermal package
is designed for worst-case power dissipation, it must
be designed for the most severe hot spot that could potentially
arise. Yet these worst-case scenarios are rare and
lead to over-design. A less expensive package can be employed,
designed for the worst “typical” or “interesting”
workload. Excessive heat dissipation must then be handled
by an autonomous, runtime response in the chip itself
that guarantees to reduce power densities far enough and
fast enough to maintain temperature regulation. For example,
the Intel Pentium 4 [ 121 follows this approach, designing
the thermal package for a power density 20% less
than the theoretical worst case. If environmental or workload
characteristics cause temperatures to rise too close to
the maximum allowable, the Pentium 4 responds by stopping
the clock, and hence all dynamic power, until temperature
returns to a safe level [15]. As another example,
the Transmeta Crusoe uses dynamic voltage scaling to reduce
power density when temperatures rise too high [lo].
Yet current techniques such as these are too simplistic
and impose unnecessary performance penalties. Architectural
design can take advantage of instruction-level parallelism
to implement DTM with substantially less performance
overhead. This paper describes three techniques,
“temperature-tracking dynamic frequency scaling,” “migrating
computation” [29], and “hybrid DTM” [26], that
reduce overhead by 25-35% compared to techniques like
the Pentium 4’s and the Transmeta Crusoe’s.
Recent work in the architecture community [2, 4, 13,
14, 19, 21, 23, 27, 311 demonstrates growing interest in
thermal management and shows that architectural techniques
offers substantial benefits. Accurately characterizing
thermal behavior and evaluating architectural techniques
requires an appropriate thermal model. An effective
model for architects must be simple enough to allow
them to reason about thermal effects and tradeoffs; parameterized;
capable of modeling runtime changes in temperature
within different functional units on the die; and yet
computationally efficient and portable.
We have developed and publicly released a dynamic
compact model-HotSpot [29]-that achieves all of these
requirements. It differs from prior dynamic compact models
for computer chips in that this new model requires less
design detail and hence can be used earlier in the design
cycle. It has been validated both against an independent
finite-element model and a test chip. It also accounts for
imprecision due to sensor noise and placement.
The rest of this paper is organized as follows. The next
section describes some example architectural thermalmanagement
techniques and presents preliminary results
evaluating their potential benefits. Section 3 describes the
proposed modeling approach. Then Section 4 describes
what we see as some important research questions in this
area, and Section 5 concludes the paper.
การแปล กรุณารอสักครู่..
