1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
|
// A Beta README file to allow folks to use GPUWattch
1. INTRODUCTION:
General-purpose GPU architectures are becoming increasingly prevalent in
mainstream computing, and as such they require judicious optimization for
energy efficiency. To enable such research, we propose a new GPU power model
that offers flexibility, adaptability, and stability. Flexibility is achieved
by using a bottom-up methodology and abstracting parameters from the
microarchitectural components as model inputs. Adaptability ensures that both
program and microarchitectural level interactions are captured during
execution, thereby enabling new power-management techniques specifically
targeted at GPUs. Stability is examined by validating the power model against
measurements of two commercial GPUs comprehensively through the leakage power,
average dynamic power, and dynamic power trace. The measured error is within
9.7% and 13.6% across our evaluated benchmark suite for the two target GPUs
(GTX 480 and Quadro FX 5600 respectively) and the model accurately tracks the
relative power consumption trend over time.
The power model modifies and extends the McPAT CPU power model simulator to
model the power of contemporary GPGPUs and drive the modified McPAT version
with a cycle-accurate simulator, GPGPU-Sim.
2. Using the Power Model
2.1 Compiling the Power Model
By default, GPGPU-Sim bypasses the McPAT compilation. Hence, it is not
necessary to install our modified McPAT along with GPGPU-Sim. However, if the
power model is required, our modified version of McPAT will be compiled
alongside GPGPU-Sim into libcudart.so.
The location of McPAT is specified by the GPGPUSIM_POWER_MODEL environment
variable and is used in GPGPU-Sim's Makefile to compile McPAT. This can either
be set manually or automatically set by the setup_evironment.sh file in
<GPGPU-Sim>/distribution if McPAT is found in the default directory
(<GPGPU-Sim>/mcpat/).
2.2. Configuration Options
This sections lists the most relevant configuration options either to describe
model a specific architecture, model a certain improvement, or to produce extra
data from the simulation. Some of these configurations are set in the GPGPU-Sim
config files and others will be set in the McPAT XML configuration files.
GPGPU-Sim configurations mainly force specific changes to the performance
modeling itself or the output format of data, while McPAT configurations mainly
determine specific parameters for the different architectural components that
are reflected in the power modeling of these components. Here, we metion only
GPGPUsim cofigurations:
1- power_simulation_enabled: Enablle the power model simulator; if enabled, an
output file is genarated to include the detailed Power coefficients for the
simulated configuration and the Average/Maximum/Minimum total power breakdowns
for each kernel.
2- mcpat_xml_file: The McPAT XML configuration file name; by default it is
mcpat.xml.
3- gpu_stat_sample_frequency: Determines the sampling frequency used in the
power calculations, the performance counters are reset before each samples and
accumulated during the sampling period, and finaly passed to the power model
(McPAT) at the end of each sample.
4- power_trace_enabled: If enabled, it produces two output files that details
the power breakdown values, and the accumulative performance counters values
for each sample.
5- power_per_cycle_dump: Dump detailed power data each sample
6- steady_power_levels_enabled: If enabled, it tracks the steady state
power level throughout the execution and report the start/end values with the
average power recorded for each componenet. The steady state is determined by
(-steady_state_definition) option.
7- steady_state_definition: Takes two values. First value detemines the allowed
deviation within the steady state and the second value determines minimum
number of samples required to assume this is a steady state power level.
3. Understanding Simulation Output
In this section, we detail the fromat of the power simulator output. By
default, if the power simulation is enabled at least one output file that
reports the average/maximum/minimum power values for each kernel is produced.
More outputs can be enabled if the corresponding configurations is enabled.
1- gpgpusim_power_report_(date&time).log- Requires (-power_simulation_enabled
1): Includes the detailed power coefficients for this configuration and the
Average/Maximum/Minimum total power and their breakdowns for the different
components for each kernel
2- gpgpusim_power_trace_(date&time).log.gz- Requires (-power_trace_enabled 1):
A compressed file that has a detailed average power breakdown trace in a comma
separated format
3- gpgpusim_metric_trace_(date&time).log.gz- Requires (-power_trace_enabled 1):
A compressed file that has a detailed performance counters trace in a comma
separated format
4- gpgpusim_steady_state_tracking_report_(date&time).log.gz- Requires
(-steady_power_levels_enabled 1): It rports the steady state power level
throughout the execution with the start/end values of each interval and the
average power recorded for each componenet during this interval in a comma
separated format
5. Software Design of the Power Model
This section presents the software design of the power model.
5.1. File list and brief description
In this section, we briefly describe the files that were added to interface
GPGPU-Sim with McPAT.
1- power_stat.cc/h @ <GPGPU_SIM>/distribution/src/gpgpu-sim/: These files
contain the main structures used for recording GPGPU-Sim performance counters:
power_core_stat_t (for all core related counters) and power_mem_stat_t (for all
memory related counters), which are contained in the wrapper power_stat_t
object. The core and mem stat structures contain multiple counter pointer
arrays with 2 locations per counter (e.g. unsigned *m_counter[2]): [0] ->
pointer to counter with the current value, [1] -> previous sampled value. The
difference, [0]-[1], is used to get the per-sample estimated power in McPAT.
2- gpgpu_sim_wrapper.cc/h @ <GPGPU_SIM>/mcpat/: These files contain the
gpgpu_sim_wrapper class that contains all of the McPAT structures (such as
Processor, ParseXML, etc), manages the power output files, and passes the
GPGPU-Sim performance counters (described in power_stat.cc/h (1)) into McPAT.
The gpgpu_sim_wrapper structure is used in power_interface.cc/h (3) to separate
the McPAT structures and interface from GPGPU-Sim.
3- power_interface.cc/h @ <GPGPU_SIM>/distribution/src/gpgpu-sim/: These files
are used to interface GPGPU-Sim with McPAT via two main functions: init_mcpat()
and mcpat_cycle(). init_mcpat() is called from gpgpu_sim::init() in gpu-sim.cc
and through the gpgpu_sim_wrapper object, initializes all of the power related
structures in GPGPU-Sim and McPAT. Similarly, mcpat_cycle() is called from
gpgpu_sim::cycle() in gpu-sim.cc, which passes all of the performance counters
to McPAT (through the gpgpu_sim_wrapper object).
4- gpgpu_sim.verify @ <GPGPU_SIM>/mcpat/: This file is distributed with our
modified version of McPAT to ensure the correct McPAT version is used with
GPGPU-Sim
|
