Training Pipeline
Model server can provide various power models for different context and learning methods. Training pipeline is an abstract of power model training that applies a set of learning methods to a different combination of energy sources, power isolation methods, available energy-related metrics.
Pipeline
pipeline
is composed of three steps, extract
, isolate
, and train
, as shown below. Kepler exports energy-related metrics as Prometheus counter, which provides accumulated number over time.
The extract
step is to convert the counter metrics to gauge metrics, similarly to the Prometheus rate()
function, giving per-second values. The extract step also clean up the data separately for each feature group
.
The power consumption retrieved from the Prometheus query is the measured power which is composed of the power portion which is varied by the workload called dynamic power and the power portion which is consumed even if in the idling state called idle power.
The isolate
step is to calculate the idle power and isolate the dynamic power consumption of each energy source
.
The train
step is to apply each trainer
to create multiple choices of power models based on the preprocessed data.
We have a roadmap to apply a pipeline to build power models separately for each node/machine type. Find more in Node Type section.
- Learn more about
energy source
from Energy source section. - Learn more about
feature group
from Feature groups section. - Learn more about the
isolate
step and corresponding concepts ofAbsPower
, andDynPower
power models from Power isolation section. - Check available
trainer
in Trainer section.
Energy source
energy source
or source
refers to the source (power meter) that provides an energy number. Each
source provides one or more energy components
. Currently supported source are shown as below.
Energy/power source | Energy/power components |
---|---|
rapl | package, core, uncore, dram |
acpi | platform |
Feature group
feature group
is an abstraction of the available features based on the infrastructure context since
some environments might not expose some metrics. For example, on the virtual machine in private cloud
environment, hardware counter metrics are typically not available. Therefore, the models are trained
for each defined resource utilization metric group as below.
Group Name | Features | Kepler Metric Source(s) |
---|---|---|
CounterOnly | COUNTER_FEATURES | Hardware Counter |
BPFOnly | BPF_FEATURES | BPF |
IRQOnly | IRQ_FEATURES | IRQ |
AcceleratorOnly | ACCELERATOR_FEATURES | Accelerator |
CounterIRQCombined | COUNTER_FEATURES, IRQ_FEATURES | BPF and Hardware Counter |
Basic | COUNTER_FEATURES, BPF_FEATURES | All except IRQ and node information |
WorkloadOnly | COUNTER_FEATURES, BPF_FEATURES, IRQ_FEATURES, ACCELERATOR_FEATURES | All except node information |
Full | WORKLOAD_FEATURES, SYSTEM_FEATURES | All |
Node information refers to value from kepler_node_info metric.
Power isolation
The power consumption retrieved from the Prometheus query is the absolute power, which is the sum of idle and dynamic power (where idle represents the system at rest, dynamic is the incremental power with resource utilization, and absolute is idle + dynamic). Additionally, this power is also the total power consumption of all process, including the users' workload, background and OS processes.
The isolate
step applies a mechanism to separate idle power from absolute power, resulting in dynamic power It also covers an implementation to separate the dynamic power consumed by background and OS processes (referred to as system_processes
).
It's important to note that both the idle and dynamic system_processes
power are higher than zero, even when the metric utilization of the users' workload is zero.
The isolate
step applies a mechanism to separate idle power from absolute power, resulting in
dynamic power. It also covers an implementation to separate the dynamic power consumed by background
and OS processes (referred to as system_processes
).
It's important to note that both the idle and dynamic system_processes
power are higher than
zero, even when the metric utilization of the users' workload is zero.
There are two common available isolators
: ProfileIsolator and MinIdleIsolator.
We refer to models trained using the isolate step as DynPower
models. Meanwhile, models
trained without the isolate step are called AbsPower
models. Currently, the DynPower
model
does not include idle power information, but we plan to incorporate it in the future.
On the other hand, MinIdleIsolator identifies the minimum power consumption among all samples in the training data, assuming that this minimum power consumption represents both the idle power and system_processes
power consumption.
While we should also remove the minimal resource utilization from the data used to train the model, this isolation mechanism includes the resource utilization by system_processes
in the training data. However, we plan to remove it in the future.
ProfileIsolator relies on collecting data (e.g., power and resource utilization) for a
specific period without running any user workload (referred to as profile data). This
isolation mechanism also eliminates the resource utilization of system_processes
from
the data used to train the model.
On the other hand, MinIdleIsolator identifies the minimum power consumption among all
samples in the training data, assuming that this minimum power consumption represents both
the idle power and system_processes
power consumption.
While we should also remove the minimal resource utilization from the data used to train the
model, this isolation mechanism includes the resource utilization by system_processes
in the
training data. However, we plan to remove it in the future.
If the profile data
that matches a given node_type
exist, the pipeline will use the
ProfileIsolator to pre-process the training data. Otherwise, the the pipeline will applied
another isolation mechanism, such as the MinIdleIsolator.
(check how profiles are generated here)
Discussion
The choice between using the DynPower
or AbsPower
model is still under investigation. In some cases, DynPower exhibits better accuracy than AbsPower
. However, we currently utilize the AbsPower
model to estimate node power for Platform, CPU and DRAM components, as the DynPower
model lacks idle power information.
It's worth mentioning that exposing idle power on a VM in a public cloud environment is not possible. This is because the host's idle power must be distributed among all running VMs on the host, and it's impossible to determine the number of VMs running on the host in a public cloud environment.
Therefore, we can only expose idle power if there is only one VM running on the node (for a very specific scenario), or if the power model is being used in Bare Metal environments.
Trainer
trainer
is an abstraction to define the learning method applies to each feature group with each given power labeling source.
Available trainer (v0.6):
- PolynomialRegressionTrainer
- GradientBoostingRegressorTrainer
- SGDRegressorTrainer
- KNeighborsRegressorTrainer
- LinearRegressionTrainer
- SVRRegressorTrainer
Node type
Kepler forms multiple groups of machines (nodes) based on its benchmark performance and trains a model separately for each group. The identified group is exported as node type
.