__ __ ____ _ _____
| \/ | ___| _ \ / \|_ _|
| |\/| |/ __| |_) / _ \ | |
| | | | (__| __/ ___ \| |
|_| |_|\___|_| /_/ \_\_|
McPAT: (M)ulti(c)ore (P)ower, (A)rea, and (T)iming
Current version 1.3
==================================================
* What McPAT is:
--Architectural integrated power, area, and timing modeling framework, focuses on power and area modeling, with a target clock rate as a design constraint.
-Consider power, area, and timing simultaneously
-Complete power envelope
-Power management techniques
---Manycore processor modeling framework
-Different cores, uncore, and system (I/O) components
-Holistically modeling across stacks: Technology models from ITRS projections (also supports user defined vdd), processor modeling based on modern processors
---Flexible, extensible, and high (i.e., architecture) level framework
-A framework for architecture research
-Flexible to make researchers's life easier
Pre-populated micro-architecture configurations (can be changed by experienced users too!)
Multilevel automatic optimization
-Hierarchical modeling framework for easy extension and porting
Standalone for TDP
Paired up with performance simulators (or machine profiling statistics) for fine-grained study
* What McPAT is NOT
---a hardware design EDA platform; nor a performance simulator
-Use RTL/SPICE/...(not McPAT) if focusing on details of complex logic or analog components
Empirical and curve-fitting based modeling for complex logic and analog building blocks (the most practical modeling methodology for high level framework).
Solution1: Users replace those models with in-house models obtained from EDA tools
Solution2: Users contribute their EDA based detailed models back to the community for sharing
-Use performance simulators for performance (McPAT cannot do performance simulations)
---a restrictive environment
-It is a framework (rather than a black-box tool)
-Its built-in models are for references and for providing methodological examples.
McPAT's built-in model includes simplified assumptions (e.g. unified instruction window for all instruction types)
McPAT provides building blocks so that it is composable
Users should always understand the methodology when using the built-in models or compose their own models.
---finished!
-There is always room for improvement . . .
-Thanks for the continueous contributing from the user community!
====================
For complete documentation of the McPAT, please refer to the following paper,
"McPAT: An Integrated Power, Area, and Timing Modeling
Framework for Multicore and Manycore Architectures",
that appears in MICRO 2009. Please cite the paper, if you use
McPAT in your work. The bibtex entry is provided below for your convenience.
@inproceedings{mcpat:micro,
author = {Sheng Li and Jung Ho Ahn and Richard D. Strong and Jay B. Brockman and Dean M. Tullsen and Norman P. Jouppi},
title = "{McPAT: An Integrated Power, Area, and Timing Modeling Framework for Multicore and Manycore Architectures}",
booktitle = {MICRO 42: Proceedings of the 42nd Annual IEEE/ACM International Symposium on Microarchitecture},
year = {2009},
pages = {469--480},
}
How to use the tool?
====================
McPAT takes input parameters from an XML-based interface,
then it computes area and peak power of the
Please note that the peak power is the absolute worst case power,
which could be even higher than TDP.
1. Steps to run McPAT:
-> define the target processor using inorder.xml or OOO.xml
-> run the "mcpat" binary:
./mcpat -infile <*.xml> -print_level < level of detailed output>
./mcpat -h (or mcpat --help) will show the quick help message.
2. Optimization:
McPAT will try its best to satisfy the target clock rate.
When it cannot find a valid solution, it gives out warnings,
while still giving a solution that is closest to the timing
constraints and calculate power based on it. The optimization
will lead to larger power/area numbers for target higher clock
rate. McPAT also provides the option "-opt_for_clk" to turn on
("-opt_for_clk 1") and off this strict optimization for the
timing constraint. When it is off, McPAT always optimize
component for ED^2P without worrying about meeting the
target clock frequency. By turning it off, the computation time
can be reduced, which suites for situations where target clock rate
is conservative.
3. Outputs:
McPAT outputs results in a hierarchical manner. Increasing
the "-print_level" will show detailed results inside each
component. For each component, major parts are shown, and associated
pipeline registers/control logic are added up in total area/power of each
components. In general, McPAT does not model the area/overhead of the pad
frame used in a processor die.
4. How to use the XML interface for McPAT
4.1 Set up the parameters
Parameters of target designs need to be set in the *.xml file for
entries tagged as "param". McPAT have very detailed parameter settings.
please remove the structure parameter from the file if you want
to use the default values. Otherwise, the parameters in the xml file
will override the default values.
4.2 Pass the statistics
There are two options to get the correct stats: a) the performance
simulator can capture all the stats in detail and pass them to McPAT;
b). Performance simulator can only capture partial stats and pass
them to McPAT, while McPAT can reason about the complete stats using
the partial information and the configuration. Therefore, there are
some overlap for the stats.
4.3 Interface XML file structures (PLEASE READ!)
The XML is hierarchical from processor level to micro-architecture
level. McPAT support both heterogeneous and homogeneous manycore processors.
1). For heterogeneous processor setup, each component (core, NoC, cache,
and etc) must have its own instantiations (core0, core1, ..., coreN).
Each instantiation will have different parameters as well as its stats.
Thus, the XML file must have multiple "instantiation" of each type of
heterogeneous components and the corresponding hetero flags must be set
in the XML file. Then state in the XML should be the stats of "a" instantiation
(e.g. "a" cores). The reported runtime dynamic is of a single instantiation
(e.g. "a" cores). Since the stats for each (e.g. "a" cores) may be different,
we will see a whole list of (e.g. "a" cores) with different dynamic power,
and total power is just a sum of them.
2). For homogeneous processors, the same method for heterogeneous can
also be used by treating all homogeneous instantiations as heterogeneous.
However, a preferred approach is to use a single representative for all
the same components (e.g. core0 to represent all cores) and set the
processor to have homogeneous components (e.g. <param name="homogeneous_cores
" value="1"/> ). Thus, the XML file only has one instantiation to represent
all others with the same architectural parameters. The corresponding homo
flags must be set in the XML file. Then, the stats in the XML should be
the aggregated stats of the sum of all instantiations (e.g. aggregated stats
of all cores). In the final results, McPAT will only report a single
instantiation of each type of component, and the reported runtime dynamic power
is the sum of all instantiations of the same type. This approach can run fast
and use much less memory.
5. Guide for integrating McPAT into performance simulators and bypassing the XML interface
The detailed work flow of McPAT has two phases: the initialization phase and
the computation phase. Specifically, in order to start the initialization phase a
user specifies static configurations, including parameters at all three levels,
namely, architectural, circuit, and technology levels. During the initialization
phase, McPAT will generate the internal chip representation using the configurations
set by the user.
The computation phase of McPAT is called by McPAT or the performance simulator
during simulation to generate runtime power numbers. Before calling McPAT to
compute runtime power numbers, the performance simulator needs to pass the
statistics, namely, the activity factors of each individual components to McPAT
via the XML interface.
The initialization phase is very time-consuming, since it will repeat many
times until valid configurations are found or the possible configurations are
exhausted. To reduce the overhead, a user can let the simulator to call McPAT
directly for computation phase and only call initialization phase once at the
beginning of simulation. In this case, the XML interface file is bypassed,
please refer to processor.cc to see how the two phases are called.
6. Sample input files:
This package provide sample XML files for validating target processors. Please find the
enclosed Niagara1.xml (for the Sun Niagara1 processor), Niagara2.xml (for the Sun Niagara2
processor), Alpha21364.xml (for the Alpha21364 processor), Xeon.xml (for the Intel
Xeon Tulsa processor), and ARM_A9_2GHz.xml (for ARM Cortex A9 hard core 2GHz implementation from
ARM)
7. Modeling of power management techniques:
McPAT supports both DVS and power-gating. For DVS, users can use default ITRS projected vdd
at each technology node as supply voltage at DVS level 0 (DVS0) or define voltage at DVS0.
For power-gating, McPAT supports both default power-saving virtual supply voltage computed
automatically using technology parameters. Default means using technology (ITRS based)
lowest value for state-retaining power-gating User can also defined voltage for Power-saving states,
as shown in example file of Xeon.xml (search for power_gating_vcc). When using user-defined power-saving
virtual supply voltage, please understand the implications when setting up voltage for different sleep states.
For example, when deep sleep state is used (voltage lower than the technology allowed state retaining supply voltage),
the effects of losing data and cold start effects (beyond the scope of McPAT) must be considered when waking up the architecture.
Power-gating and DVS cannot happen at the same time. Because power-gating happens when circuit is idle, while DVS happens when
circuit blocks are active.
====================
McPAT includes its special version of Cacti (called Cacti-P) based on Cacti6.5 release. The major changes of
the special Cacti, called Cacti-P in this distro, (compared to cacti6.5) include the following new features.
The inclosed Cacti-P can run stand-alone if users want to use these features.
* CAM and fully associative cache modeling
* Improved leakage power modeling with consideration of device/gate topology
* long channel device for reduce sub-threshold leakage power
* Sleep transistor based power-gating modeling
* gate leakage power
* Support user defined voltage supply (Vdd)
* Dynamic voltage scaling (DVS)
For complete documentation of Cacti-P, please refer to the following paper,
"CACTI-P: Architecture-Level Modeling for SRAM-based Structures with Advanced Leakage Reduction Techniques",
that appeared in ICCAD2011. Please cite the paper, if you use
Cacti-P in your work. The bibtex entry is provided below for your convenience.
@inproceedings{cacti-p:iccad,
author = {Sheng Li and Ke Chen and Jung Ho Ahn and Jay B. Brockman and Norman P. Jouppi},
title = {CACTI-P: Architecture-level modeling for SRAM-based structures with advanced leakage reduction techniques},
booktitle = {ICCAD: International Conference on Computer-Aided Design},
year = {2011},
pages = {694-701},
}
====================
McPAT uses an opensource XML parser written and kindly specially licensed by Mr. Frank Vanden Berghen.
The detailed information about this XML parser can be found at the license information in xmlParse.cc/xmlParse.h
====================
McPAT is in its beginning stage. We are still improving the tool.
Please come back to its website for newer versions.
McPAT has been constantly and rapidly improved with new models and latest technology.
Please always refer to its code for most up-to-date and most accurate information.
If you have any comments, questions, or suggestions, please write to us:
Sheng Li
Sheng.sli@gmail.com