powermetrics(1) BSD General Commands Manual powermetrics(1)
NAME
powermetrics
SYNOPSIS
powermetrics [-i sample_interval_ms] [-o order] [-t wakeup_cost]
[-u output_file] [-n sample_count]
DESCRIPTION
powermetrics gathers and display CPU usage statistics (divided into time
spent in user mode and supervisor mode), timer and interrupt wakeup fre-
quency (total and, for near-idle workloads, those that resulted in pack-
age idle exits), and on supported platforms, interrupt frequencies (cate-
gorized by CPU number), package C-state statistics (an indication of the
time the core complex + integrated graphics, if any, were in low-power
idle states), as well as the average execution frequency for each CPU
when not idle.
-h, --help
Print help message.
-u file, --output-file file
Output to file instead of stdout.
-b size, --buffer-size size
Set output buffer size (0=none, 1=line)
-i N, --sample-interval N
sample every N ms (0=disabled) [default: 5000ms]
-n N, --sample-count N
Obtain N periodic samples (0=infinite) [default: 0]
-t N, --wakeup-cost N
Assume package idle wakeups have a CPU time cost of N us when
using hybrid sort orders using idle wakeups with time-based met-
rics
-o method, --order method
Order process list using specified method [default: composite]
[pid]
process identifier
[wakeups]
total package idle wakeups (alias: -W)
[cputime]
total CPU time used (alias: -C)
[composite]
weighted hybrid of package idle wakeups and CPU time used
(alias: -O)
-f format, --format format
Display data in specified format [default: text]
[text]
human-readable text output
[plist]
machine-readable property list, NUL-separated
-a N, --poweravg N
Display poweravg every N samples (0=disabled) [default: 10]
--hide-platform-power
Hide platform power data
--hide-cpu-duty-cycle
Hide CPU duty cycle data
--hide-gpu-duty-cycle
Hide GPU duty cycle data
--show-initial-usage
Print initial sample for entire uptime
--show-usage-summary
Print final usage summary when exiting
This tool also implements special behavior upon receipt of certain sig-
nals to aid with the automated collection of data:
SIGINFO
take an immediate sample
SIGIO
flush any buffered output
SIGINT/SIGTERM
stop sampling and exit
OUTPUT
Guidelines for energy reduction
CPU time, deadlines and interrupt wakeups: Lower is better
Interrupt counts: Lower is better
C-state residency: Higher is better
Running Tasks
1. CPU time consumed by threads assigned to that process, broken down
into time spent in user space and kernel mode.
2. Counts of "short" timers (where the time-to-deadline was < 5 millisec-
onds in the future at the point of timer creation) which woke up threads
from that process. High frequency timers, which typically have short
time-to-deadlines, can result in significant energy consumption.
3. A count of total interrupt level wakeups which resulted in dispatching
a thread from the process in question. For example, if a thread were
blocked in a usleep() system call, a timer interrupt would cause that
thread to be dispatched, and would increment this counter. For workloads
with a significant idle component, this metric is useful to study in con-
junction with the package idle exit metric reported below.
4. A count of "package idle exits" induced by timers/device interrupts
which awakened threads from the process in question. This is a subset of
the interrupt wakeup count. Timers and other interrupts that trigger
"package idle exits" have a greater impact on energy consumption relative
to other interrupts. With the exception of some Mac Pro systems, Mac and
iOS systems are typically single package systems, wherein all CPUs are
part of a single processor complex (typically a single IC die) with
shared logic that can include (depending on system specifics) shared last
level caches, an integrated memory controller etc. When all CPUs in the
package are idle, the hardware can power-gate significant portions of the
shared logic in addition to each individual processor's logic, as well as
take measures such as placing DRAM in to self-refresh (also referred to
as auto-refresh), place interconnects into lower-power states etc. Hence
a timer or interrupt that triggers an exit from this package idle state
results in a a greater increase in power than a timer that occurred when
the CPU in question was already executing. The process initiating a pack-
age idle wakeup may also be the "prime mover", i.e. it may be the trigger
for further activity in its own or other processes. This metric is most
useful when the system is relatively idle, as with typical light work-
loads such as web browsing and movie playback; with heavier workloads,
the CPU activity can be high enough such that package idle entry is rela-
tively rare, thus masking package idle exits due to the process/thread in
question.
5. If any processes arrived and vanished during the inter-sample inter-
val, or a previously sampled process vanished, their statistics are
reflected in the row labeled "DEAD_TASKS". This can identify issues
involving transient processes which may be spawned too frequently. dtrace
("execsnoop") or other tools can then be used to identify the transient
processes in question.
Interrupt Distribution
Powermetrics also reports interrupt frequencies, classified by interrupt
vector and associated device, on a per-CPU basis.Mac OS currently assigns
all device interrupts to CPU0, but timers and interprocessor interrupts
can occur on other CPUs. Interrupt frequencies can be useful in identify-
ing misconfigured devices or areas of improvement in interrupt load, and
can serve as a proxy for identifying device activity across the sample
interval. For example, during a network-heavy workload, an increase in
interrupts associated with Airport wireless ("ARPT"), or wired ethernet
("ETH0" "ETH1" etc.) is not unexpected. However, if the interrupt fre-
quency for a given device is non-zero when the device is not active (e.g.
if "HDAU" interrupts, for High Definition Audio, occur even when no audio
is playing), that may be a driver error.
Battery Statistics
Powermetrics reports battery discharge rates, current and maximum charge
levels, cycle counts and degradation from design capacity across the
interval in question, if a delta was reported by the battery management
unit. Note that the battery controller data may arrive out-of-phase with
respect to powermetrics samples, which can cause aliasing issues across
short sample intervals. Discharge rates across discontinuities such as
sleep/wake may also be inaccurate on some systems; however, the rate of
change of the total charge level across longer intervals is a useful
indicator of total system load. Powermetrics does not filter discharge
rates for A/C connect/disconnect events, system sleep residency etc. Bat-
tery discharge rates are typically not comparable across machine models.
Processor Energy Usage
Powermetrics next reports data derived from the Intel energy models; as
of the Sandy Bridge intel microarchitecture, the Intel power control unit
internally maintains an energy consumption model whose details are pro-
prietary, but are likely based on duty cycles for individual execution
units, current voltage/frequency etc. These numbers are not strictly
accurate but are correlated with actual energy consumption. This section
lists: power dissipated by the processor package which includes the CPU
cores, the integrated GPU and the system agent (integrated memory con-
troller, last level cache), and separately, CPU core power and GT (inte-
grated GPU) power (the latter two in a forthcoming version). The energy
model data is generally not comparable across machine models.
Powermetrics next reports, on processors with Nehalem and newer microar-
chitectures, hardware derived processor frequency and idle residency
information, labeled "P-states" and "C-states" respectively in Intel ter-
minology.
C-states are further classified in to "package c-states" and per-core C-
states. The processor enters a "c-state" in the scheduler's idle loop,
which results in clock-gating or power-gating CPU core and, potentially,
package logic, considerably reducing power dissipation. High package c-
state residency is a goal to strive for, as energy consumption of the CPU
complex, integrated memory controller if any and DRAM is significantly
reduced when in a package c-state. Package c-states occur when all CPU
cores within the package are idle, and the on-die integrated GPU if any
(SandyBridge mobile and beyond), on the system is also idle. Powermetrics
reports package c-state residency as a fraction of the time sampled. This
is available on Nehalem microarchitecture and newer processors. Note that
some systems, such as Mac Pros, do not enable "package" c-states.
Powermetrics also reports per-core c-state residencies, signifying when
the core in question (which can include multiple SMTs or "hyperthreads")
is idle, as well as active/inactive duty cycle histograms for each logi-
cal processor within the core. This is available on Nehalem microarchi-
tecture and newer processors.
This section also lists the average clock frequency at which the given
logical processor executed when not idle within the sampled interval,
expressed as both an absolute frequency in MHz and as a percentage of the
nominal rated frequency. These average frequencies can vary due to the
operating system's demand based dynamic voltage and frequency scaling.
Some systems can execute at frequencies greater than the nominal or "P1"
frequency, which is termed "turbo mode" on Intel systems. Such operation
will manifest as > 100% of nominal frequency. Lengthy execution in turbo
mode is typically energy inefficient, as those frequencies have high
voltage requirements, resulting in a correspondingly quadratic increase
in power insufficient to outweigh the reduction in execution time. Cur-
rent systems typically have a single voltage/frequency domain per-pack-
age, but as the processors can execute out-of-phase, they may display
different average execution frequencies.
Disk Usage and Network Activity
Powermetrics reports deltas in disk and network activity that occured
during the sample.
Backlight level
Powermetrics reports the instantaneous value of the backlight luminosity
level. This value is likely not comparable across systems and machine
models, but can be useful when comparing scenarios on a given system.
KNOWN ISSUES
Changes in system time and sleep/wake can cause minor inaccuracies in
reported cpu time.
Darwin October 15, 2013 Darwin
Mac OS X 10.9 - Generated Tue Oct 15 07:56:15 CDT 2013