Escalonamento e Migração de Recursos e Balanceamento de carga 27-04-20151 Carlos Ferrão Lopes nº M6935 Bruno Simões nº M6082 Celina Alexandre nº M6807.

Escalonamento e Migração de Recursos e Balanceamento de

carga

27-04-2015 1

Carlos Ferrão Lopes nº M6935

Bruno Simões nº M6082Celina Alexandre nº M6807

Paulo Cabral nº M6765 Quoc nº M6831

Tecnologias de Cloud e Datacenter

Professor Doutor Nuno M. Garcia

Overview

1. Stagger 2. Migration3. Load Balancing

27-04-2015 2

Overview

1. Stagger2. Scheduling Tasks

2.1. Algorithms Hetergeneous Earliest Finish Time (HEFT)

2.2. Algorithms Path Clustering Heuristic (PCH)

2.3 Scheduling algorithms for the comparison study

27-04-2015 3

Overview (Cont.)

3. Scheduling Procedure3.1. Process Scheduling in Distributed System

27-04-2015 4

What is Scheduling?

A system that uses multi-programming processes compete for processing, as they require dividing the execution time of each.

27-04-2015 5

Scheduling problem

Problem workflow scheduling a set of dependent services: Order of precedence of services; Costs of performance of the services; Communication costs between services; Resource processing capabilities; Data transmission capabilities of network connections,

which interconnect these resources;

27-04-2015 6

Scheduling algorithm Rating

27-04-2015 7

Rating scheduling methods

Scheduling tasks

dependent independent

27-04-2015 8

Heterogeneous Earliest Finish Time Algorithm (HEFT)

Algorithm Workflow escalation Using an acyclic digraph (DAG1) for a limited

number of heterogeneous computers. Task scheduling problem is NP-hard

27-04-2015 1DAG - Directed acyclic graph

9

Heterogeneous Earliest Finish Time Algorithm (HEFT) (Cont.)

The HEFT consists of two main phases: Prioritization phase Selection phase

27-04-2015 10

Prioritization phase

Each task should be prioritized considering the length of the critical path;

The list of tasks to be performed is then sorted in descending order of the critical path length;

A topological sorting is produced tasks, preserving the precedence constraints of the DAG;

27-04-2015 11

Selection Phase

To select which processor will perform a given task, calculating the earliest time of completion of a given task. With HEFT algorithm, the search for a vacant time slot on a processor P starts from the moment that the processor P becomes vacant.

27-04-2015 12

Path Heuristic Clustering Algorithm (PCH)

Clustering uses the technique to create groups (clusters) tasks;

Groups the ways of the DAG, thus reducing communication costs between tasks;

The tasks of the same cluster are scheduled on the same resource;

27-04-2015 13

Path Heuristic Clustering Algorithm (PCH) (Cont.)

Uses some attributes calculated for each task, estimating the start times (DS = Earliest Start Time) and end (EFT - Estimated Finish Time) of tasks and process;

EST and EFT are calculated using information that is provided by the middleware and the programming model;

27-04-2015 14

Path Heuristic Clustering Algorithm (PCH) (Cont.)

27-04-2015 15

Example scheduling using the PCH

Scheduling algorithms for the comparison study

Applications used: Application Montage1

Application Epigenomics2

27-04-2015

1 is used to generate personalized sky mosaics using multiple points of input images.

2 is used to mapping of epigenetic state of human cells

on a large scale genomics.16

Analysis Applications

27-04-2015 17

Representation of epigenomics application used in the simulation.

Representation of montage application used in the simulation.

Architecture

Were also specified two sets: homogeneous and heterogeneous;

Have been implemented in addition to the scheduling algorithms HEFT, CPOP and PCH, a FIFO type of scheduling;

27-04-2015 18

Architecture (Cont.)

27-04-2015 19

Architecture

Architecture (Cont.)

27-04-2015 20

Computing power table

Experimental Results

Applications to the Montage   We observed that the FIFO strategy only performs

well in homogeneous architecture.   The PCH algorithm does not perform well in both

architectures.   The algorithms HEFT and CPOP perform well in both

architectures with an almost negligible difference.

27-04-2015 21

Analysis Applications

Workload simulation result Montage with a set Homogeneous

Workload simulation result Montage with a set Heterogeneous

27-04-2015

Makespan – The time difference between the beginning and end

of a sequence of jobs or tasks22

Experimental Results

Applications to Epigenomics   The PCH algorithm improves performance in this type

of application.   The CPOP algorithm preserves the good performance

in both architectures.   The HEFT algorithm performs well in heterogeneous

architecture, but in the homogeneous architecture not.

27-04-2015 23

Analysis Applications

Workload simulation result Epigenomics

with a set Homogeneous

Workload simulation result Epigenomics

with a set Heterogeneous

27-04-2015

Makespan – The time difference between the beginning and end

of a sequence of jobs or tasks24

Conclusions

A FIFO strategy is appropriate only in homogeneous architectures.

In heterogeneous architectures it takes a study of both computational resources and tasks.

The HEFT and CPOP algorithms showed good performance and at the same time like in Montage application, while SHP not.

In applications with equipment that prevents the performance in a continuous process, such as montage application, it is not convenient grouping of tasks.

27-04-2015 25

Scheduling Procedure

A multi-programming computer has multiple processes running simultaneously, however, that all processes can not access the CPU at a time and so there is a fair way of implementation of these processes, an algorithm was created to organize tasks.

27-04-2015 26

Process Scheduling in Distributed System

Heterogeneity of the nodes prevents equal distribution;

  In addition to the tasks to be performed it is necessary to: Processing power; Unemployment level;

  If you are in possession of information: Weight of tasks; Estimated processing time; 

27-04-2015 27

Information nodes

Physical characteristics: speed, memory, processor type

Degree of idleness of each node: the processing power is being used;

27-04-2015 28

Measure the level of unemployment

Average waiting time in the run queue Problem: Priority difference of tasks;

    Generate CPU utilization rate 

Problem: to test specific process from time to time;

27-04-2015 29

Measure the level of unemployment (Cont.)

Problem General: update frequency  Excessive Update: scheduling algorithm becomes

main consumer; Insufficient Update: wrong choice of a node; 

Algorithms: Static scheduling; Scheduling dynamic Adaptive scheduling;

27-04-2015 30

Scheduling static

Assessment of conditions of the nodes is done once;

Data are used during the entire process;

27-04-2015 31

Dynamic Scheduling

Constant analysis of the nodes characteristics: Waiting time in the CPU queue: little processing; CPU Utilization: Small background processes making

continuous testing;  

27-04-2015 32

Adaptive Scheduling:

Dynamic Scheduling special case; You can adjust the intensity of CPU consumption:

May fail to perform some tests

27-04-2015 33

Overhead

Distributed memory: high cost in exchange of information between nodes;

Can occur over communications and performance problems: Nodes are communicating longer than processing;

27-04-2015 34

Algorithms of classification according to the policy

You can sort scheduling algorithms under policies: Transfer policy: determining a suitable node to receive

tasks; Selection Policy: determine the most appropriate task

to be sent to a node; Location Policy: determine the most appropriate node

to receive a given task; Information Policy: determine storage location of

information and frequency of update.

27-04-2015 35

Transfer Policy

Determine a receiver node: node is working below its processing capacity;

Is defined maximum load supported by each node;

27-04-2015 36

Transfer Policy (Cont.)

Problem: donor sends a task to a receiver that immediately becomes a donor: Task switching is done indefinitely.

Is defined maximum load supported by each node; Solution: Consider a donor node whose load is less

than the limit and will continue to be less than or equal after receiving the task;

27-04-2015 37

Selection Policy

Determine a task to be sent when a node is overloaded or eligible to receive;

In an overloaded node: Select the task that caused the overload;

27-04-2015 38

Selection Policy (Cont.)

Preemptive scheduling: A task already running can be transferred to another

node; High transfer costs of the task context (ex .: data in

memory); Usually picks up recently started tasks;

Non-Preemptive Scheduling: Only tasks that have not yet implemented can be

transferred;

27-04-2015 39

Location Policy

Select node to receive a task; Random method (most commonly used):

Node randomly chooses another node; Problem: chances of choosing a node in the same

situation; Immediate solution: test if selected node is receiver; Solution "better": vote between selected nodes.

Generates excessive communication.

27-04-2015 40

Information Policy

Corresponds to the exchange of information on their status across multiple nodes: Local storage: a specific node receives all the status

messages of others. Broadcast: each node sends to others; Sub-demand: each node has information about you.

Information about others only when it becomes donor or recipient.

27-04-2015 41

Scheduling algorithms

Started by the donor: The node that has a donor in accordance with the

transfer policy utilized, looking for a receiver; Algorithm Random (most common), making the choice

of a single receptor twice. Sub-demand: each node has information about you.

Information about others only when it becomes donor or recipient.

27-04-2015 42

Scheduling algorithms (Cont.)

Started by the receiver: The receiver node calls tasks to process; Most advantageous approach: unoccupied node has

more time to choose a suitable donor;

27-04-2015 43