spdf/elektron
Archived
1
0
Fork 0
Code Issues 4 Pull requests 1 Projects Releases Packages Wiki Activity

moved documentation of scheduling policies and power-capping strategies into a documentation folder called docs/

Browse source
This commit is contained in:
Pradyumna Kaushik 2017-09-28 15:53:36 -04:00
parent 4051aa1fb3
commit ad9f67a2a0
2 changed files with 0 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

13
docs/PowerCappingStrategies.md Normal file
View file

@ -0,0 +1,13 @@
Elektron: Power-Capping Strategies
==================================
##__Power-Capping Policies__
* **Extrema** - A dynamic power-capping strategy that is able to make smart trade-offs
between makespan and power consumption. *Extrema* reacts to power trends in the
cluster and restrains the power consumption of the cluster to a power
envelope defined by a high threshold and low threshold.
* **Progressive-Extrema** - A modified version *Extrema* that performs
power-capping in phases. Unlike in *Extrema*, where picking a previously
capped node as a victim resulted in a NO-OP, *Progressive-Extrema* applies
a harsher capping value for that victim.

24
docs/SchedulingPolicies.md Normal file
View file

@ -0,0 +1,24 @@
Elektron: Pluggable Scheduling Policies
=======================================
##__Scheduling Policies__
* **First Fit** - *Find the first task in the job queue whose
resource constraints are satisfied by the resources available in
a resource offer. If a match is made between an offer and a task,
the offer is consumed in order to schedule the matching task. Else,
move onto a new mesos resource offer and repeat the process.*
* **Bin-packing** - *For each resource offer received, tasks are matched
from the priority queue keyed by the Watts resource requirement. If a task's
resource requirements are not satisfied by the remaining resources in the
offer, the next task in the queue is evaluated for fitness. This way, tasks
are packed into a resource offer.*
* **Max-Min** - *Pack a resource offer by picking a mixed set of tasks
from the job queue (used as a double-ended queue) that is sorted in
non-decreasing order of tasks' Watts requirement. Max-Min alternates
between picking tasks from the front and back of the queue.*
* **Max-GreedyMins** - *Similar to Max-Min, a double-ended queue,
sorted in non-decreasing order of tasks' Watts requirement, to store
the tasks. Max-GreedyMins aims to pack tasks into an offer by picking
one task from the end of the queue, and as many from the beginning,
until no more task can fit the offer.*