跳转至

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.

Pipeline Plot

  • 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 of AbsPower, and DynPower 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
CgroupOnly CGROUP_FEATURES cGroups
BPFOnly BPF_FEATURES BPF
IRQOnly IRQ_FEATURES IRQ
AcceleratorOnly ACCELERATOR_FEATURES Accelerator
CounterIRQCombined COUNTER_FEATURES, IRQ_FEATURES BPF and Hardware Counter
Basic COUNTER_FEATURES, CGROUP_FEATURES, BPF_FEATURES All except IRQ and node information
WorkloadOnly COUNTER_FEATURES, CGROUP_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.

Copyright Contributors to the Kepler's project.

The Linux Foundation® (TLF) has registered trademarks and uses trademarks. For a list of TLF trademarks, see Trademark Usage.