Elektron is a Mesos framework that behaves as a playground to experiment with different scheduling policies to schedule ad-hoc jobs in docker containers. It is designed as a lightweight, configurable framework, which can be used in conjunction with built-in power-capping policies to reduce the peak power and/or energy usage of co-scheduled tasks.
However, in addition to being a scheduler, Elektron also takes advantage of tools such as Performance Co-Pilot and RAPL to help contain the power envelope within defined thresholds, reduce peak power consumption, and also reduce total energy consumption. Elektron is able to leverage the Mesos-provided resource abstraction to allow different algorithms to decide how to consume resource offers made by a Mesos Master.
Elektron is comprised of three main components: Task Queue, Scheduler and Power Capper.
- Task Queue - Maintains tasks that are yet to be scheduled.
- Scheduler - Matches tasks' resource requirements with Mesos resource offers. Tasks that matched offers are then launched on the corresponding nodes.
- Power Capper - The Power Capper monitors the power consumption of the nodes in the cluster through the use of Performance Co-Pilot. A power capping policy uses this information and decides to power cap or power uncap one or more nodes in the cluster using RAPL.
- Pradyumna Kaushik, Madhusudhan Govindaraju, Srinidhi Raghavendra, Devesh Tiwari, "Exploring the Potential of using Power as a First Class Parameter for Resource Allocation in Apache Mesos Managed Clouds", in the 13th IEEE/ACM International Conference on Utility and Cloud Computing (UCC 2020), 2020. doi, pdf
- Pradyumna Kaushik, Akash Kothawale, Renan DelValle, Abhishek Jain, Madhusudhan Govindaraju, “Analysis of Dynamically Switching Energy-Aware Scheduling Policies for Varying Workloads”, in the 11th IEEE International Conference on Cloud Computing (IEEE Cloud), 2018. [pdf]
- Renan Delvalle, Pradyumna Kaushik, Abhishek Jain, Jessica Hartog, Madhusudhan Govindaraju, “Exploiting Efficiency Opportunities Based on Workloads with Electron on Heterogeneous Clusters”, in the The 10th IEEE/ACM International Conference on Utility and Cloud Computing (UCC 2017), 2017. [pdf]
- Renan DelValle, Abhishek Jain, Pradyumna Kaushik, Jessica Hartog, Madhusudhan Govindaraju, “Electron: Towards Efficient Resource Management on Heterogeneous Clusters with Apache Mesos”, in the IEEE International Conference on Cloud Computing (CLOUD), Applications Track, 2017. [pdf]
Note: Elektron was previously known as Electron. We decided to change the name of the framework to avoid confusion with other projects named Electron.
Please go through the log info to get information on different data that are logged.
Requires Performance Co-Pilot tool pmdumptext to be installed on the machine on which electron is launched for logging to work and PCP collector agents installed on the Mesos Agents
Compatible with the following versions:
- Mesos 1.5.0
- Go 1.9.7 (if using go vendor for dependency management)
- Go 1.11+ (if using go modules for dependency management)
Go Modules can now be used for dependency management. To download the dependencies, run the below command.
go mod download
Note that you would require Go version 1.11+ to be able to use go modules.
If vendoring dependencies, then use the below commands after cloning elektron.
git submodule initgit submodule update
An alternative is to clone elektron using the command git clone --recurse-submodules [email protected]:spdfg/elektron.git.
Compile the source code using the go build tool as shown below.
go build -o elektron
Use the -h option to get information about other command-line options.
Use the -workload option to specify the location of the workload json file. Below is an example workload.
[
{
"name": "minife",
"cpu": 3.0,
"ram": 4096,
"watts": 63.141,
"image": "rdelvalle/minife:electron1",
"cmd": "cd src && mpirun -np 3 miniFE.x -nx 100 -ny 100 -nz 100",
"inst": 10
},
{
"name": "dgemm",
"cpu": 3.0,
"ram": 32,
"watts": 85.903,
"image": "rdelvalle/dgemm:electron1",
"cmd": "/./mt-dgemm 1024",
"inst": 10
}
]./elektron -master <host:port> -workload <workload json>
Use the -logPrefix option to provide the prefix for the log file names.
Elektron is also capable of running power capping policies along with scheduling policies.
Use the -powercap option with the name of the power capping policy to be run.
./elektron -master <host:port> -workload <workload json> -powercap <powercap policy name>
If the power capping policy is Extrema or Progressive Extrema, then the following options must also be specified.
-hiThreshold- If the average historical power consumption of the cluster exceeds this value, then one or more nodes would be power capped.-loThreshold- If the average historical power consumption of the cluster is lesser than this value, then one or more nodes would be uncapped.
Use the -schedPolicy option with the name of the scheduling policy to be deployed.
The default scheduling policy is First Fit.
./elektron -master <host:port> -workload <workload json> -schedPolicy <sched policy name>
Note: To obtain the list of possible scheduling policy names, use the -listSchedPolicies option.
Use the -switchSchedPolicy option to enable scheduling policy switching.
One needs to also provide a scheduling policy configuration file (see schedPolConfig for reference).
Use the -schedPolConfig option to specify the path of the scheduling policy configuration file.
./elektron -master <host:port> -workload <workload json> -switchSchedPolicy -schedPolConfig <config file>
The following options can be used when scheduling policy switching is enabled.
-fixFirstSchedPol- Fix the first scheduling policy that is deployed.-fixSchedWindow- Allow the size of the scheduling window to be fixed.-schedWindowSize- Specify the size of the scheduling window. If no scheduling window size specified andfixSchedWindowoption is enabled, the default size of 200 is used.-schedPolSwitchCriteria- Criteria to be used when deciding the next scheduling policy to switch to. Default criteria is task distribution (taskDist) based. However, one can either switch based on a Round Robin (round-robin) or Reverse Round Robin (rev-round-robin) order.