COST IC804 – IC805 Joint meeting, February 7-8 2013
Jorge G. Barbosa, Altino M. Sampaio, Hamid Harabnejad
Universidade do Porto, Faculdade de Engenharia, LIACCPorto, Portugal, [email protected]
Experiments on cost/power and failure aware scheduling for clouds and grids
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 2
Outline Dynamic Power- and Failure-aware Cloud Resources
Allocation for Sets of Independent Tasks
A Budget Constrained Scheduling Algorithm for Workflow Applications on Heterogeneous Clusters
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 3
Outline Dynamic Power- and Failure-aware Cloud Resources
Allocation for Sets of Independent Tasks
A Budget Constrained Scheduling Algorithm for Workflow Applications on Heterogeneous Clusters
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 4
Dynamic Power- and Failure-aware Cloud Resources Allocation for Sets of Independent Tasks
Cloud computing paradigm
Image source: http://www.commputation.kit.edu/92.php
Dynamic provisioning of computing services.
Employs Virtual Machine (VM) technologies for consolidation and environment isolation purposes.
Node failure can occur due to hardware or software problems.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 5
Characteristics Dependability of the infrastructure
Distributed systems continue to grow in scale and in complexity Failures become norms, which can lead to violation of the negotiated SLAs Mean Time Between Failures (MTBF) would be 1.25h on a petaflop system(1)
Energy consumption The main part of energy consumption is determined by the CPU Energy consumption dominates the operational costs
(1) S. Fu, "Failure-aware resource management for high-availability computing clusters with distributed virtual machines," Journal of Parallel and Distributed Computing, vol. 70, April 2010, pp. 384-393, doi: 10.1016/j.jpdc.2010.01.002.
VMM VMM VMMVMM
VM1 VM 4VM 2
PM 1 PM 2 PM 3 PM m
...
Task 1 Task 2 Task n
VM n
Task 3
PM – Physical Machine
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 6
Related Work
(1) Optimistic Best-Fit (OBFIT) algorithm- Selects the PM with minimum weighted available capacity and reliability.
(2) Pessimistic Best-Fit (PBFIT) algorithm - Selects also unreliable PMs in order to increase the job completion rate. - Selects the unreliable PM p with capacity Cp such that Cavg + Cp results in the minimum required capacity
Cavg average capacity from reliable PMs.
Dynamic allocation of VMs, considering PMs’ reliability Based in a failure predictor tool with 76.5% of accuracy
Proposed architecture for reconfigurable distributed VM (1)
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 7
Approach The goal
It is a best-effort approach, not a SLA based approach; Virtual-to-physical resources mapping decisions must consider both the
power-efficiency and reliability levels of compute nodes; Dynamic update of virtual-to-physical configurations (CPU usage and
migration).
Construct power- and failure-aware computing environments, in order to maximize the rate of completed jobs by their deadline
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 8
Approach Multi-objective scheduling algorithms are addressed in three ways:
1- Finding the pareto optimal solutions, and let the user select the best solution.
2- Combination of the two functions in a single objective function.
3- Bicriteria scheduling which the user specifies a limitation for one criterion (power or budget constraints), and the algorithm tries to optimize the other criterion under this constraint.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 9
Approach Leverage virtualization tools
Xen credit scheduler Dynamically update cap parameter But enforcing work-conserving
Stop & copy migration Faster VM migrations, preferable for proactive failure management
CPU
CPU% Powerconsumption
100
0
VM
VMVM
VMVM
VM
timePM3
PM2
PM1
– Failure – Stop & copy migration
Incr
easin
g
– Failure prediction accuracy
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 10
System Overview Cloud architecture
Private cloud Homogenous PMs Cluster coordinator manages
user’ jobs VMs are created and destroyed
dynamically
Users’ jobs A job is a set of independent tasks A task runs in a single VM, which CPU-intensive workload is known Number of tasks per job and tasks deadlines are defined by user
Private cloud management architecture
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 11
Power Model Linear power model
P = p1 + p2.CPU%
Power Efficiency of P
Completion rate of users’ jobs
Working Efficiency
%( ) 1 2
1 2 %
CPUEff P p p
p p CPU
Example of power efficiency curve (p1 = 175w, p2 = 75w)
1( )
1
lJccEff J kJss
( )
11( ) ,
uEff Pf ii
s uEff Eff J u hf
Measures the quantity of useful work done (i.e. completed users’ jobs) by the consumed power.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 12
Proposed algorithms Minimum Time Task Execution (MTTE) algorithm
Selects a PM if:It guarantees maximum processing power
required by the VM (task);It has higher reliability;And if It increases CPU Power Efficiency.
PM i capacity constraints 1
qr Ct it
Slack time to accomplish task t
max
Wtdt rt
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 13
Proposed algorithms Relaxed Time Task Execution (RTTE) algorithm
Unlike MTTE, the RTTE algorithm always reserves to VM the minimum amount of resources necessary to accomplish the task within its deadline
Host
CPU
100%
0%
VM
Cap set in Xen credit scheduler
CAP
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 14
Performance Analysis Simulation setup
50 PMs, each modeled with one CPU core with the performance equivalent to 800 MFLOPS;
VMs stop & copy migration overhead takes 12 secs; 30 synthetic jobs, each being constituted of 5 CPU-intensive
workload tasks; Failed PMs stay unavailable during 60 secs; Predicted occurrence time of failure precedes the actual occurrence
time; Failures instants, jobs arriving time, and tasks workload sizes follow
an uniform distribution;
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 15
Performance Analysis Implementation considerations
Stabilization to avoid multiple migrations Concurrence among cluster coordinators
Algorithms compared to ours Common Best-Fit (CBFIT)
Selects the PM with the maximum power-efficiency and do not consider resources reliability
Optimistic Best-Fit (OBFIT) Pessimistic Best-Fit (PBFIT)
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 16
Performance Analysis Migrations occurring due to
proactive failure management only:
Failure predictor tool has 76.5% of accuracy; RTTE algorithm presents the best results;
Working efficiency, as well as the jobs completion rate, decreases with failure prediction inaccuracy.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 17
Performance Analysis Migrations occurring due to proactive
failure management and power efficiency:
Sliding window of 36 seconds, with threshold of 65% (a migration starts if CPU usage below 65%);
RTTE returns the best results for 76.5% failure prediction accuracy;
Comparing to earlier results, the rate of completed jobs diminishes, since the number of VMs migrations increases.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 18
Performance Analysis Number of migrations occurring due
to failure management and power efficiency
RTTE and MTTE have stable number of migrations and respawns along failure accuracy variation
Migrations occurring due to proactive failure management only (75% accuracy)
RTTE and MTTE return the best working efficiency as the number of failures in the cloud infrastructure rises
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 19
Conclusions (1) Conclusion remarks:
Power- and failure-aware dynamic allocations improve the jobs completion rate; Dynamically adjusting cap parameter of Xen credit scheduler prove to be
capable of obtaining better jobs completion rate (RTTE); Excessive number of VM migrations to optimizing power efficiency reduces job
completion rate.
Future directions: Dynamic allocation considering workload characteristics; Data locality; Scalability; Compare/integrate DVFS feature; Improve PM consolidation (why 65% threshold?); Heterogeneous CPUs.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 20
Outline Dynamic Power- and Failure-aware Cloud Resources
Allocation for Sets of Independent Tasks
A Budget Constrained Scheduling Algorithm for Workflow Applications on Heterogeneous Clusters
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 21
A Budget Constrained Scheduling Algorithm for Workflow Applications on Heterogeneous Clusters
A Job is represented by a workflowA workflow is a Directed Acyclic Graph (DAG)
a node is an individual task
an edge represents the inter-job dependency
CPU1
CPU2
CPU3
Workflow schedulingMapping Tasks to ResourcesMain goal is to have a lower finish time of the exit task
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 22
IntroductionTarget platform: - Utility Grids that are maintained and managed by a service provider. - Based on user requirements, the provider finds a scheduling that meets user constrains.
In utility Grids, other QoS attributes than execution time, like economical cost or deadline, may be considered. It is a multi-objective problem.
Multi-objective scheduling algorithms are addressed in three ways:1- Finding the pareto optimal solutions, and let the user select the best solution;2- Combination of the two functions in a single objective function;3- Bicriteria scheduling which the user specifies a limitation for one criterion (power or budget constraints), and the algorithm tries to optimize the othercriterion under this constraint.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 23
Proposed AlgorithmHeterogeneous Budget Constraint Scheduling Algorithm (HBCS)
HBCS has two phases:
Task Selection Phase :
We use Upward rank to assign the priority to tasks in the DAG
Processor Selection Phase :
We combine both objective functions (cost and time) in a single
function; the processor that maximizes that function for the
current task is selected.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 24
Proposed AlgorithmHeterogeneous Budget Constraint Scheduling Algorithm (HBCS)
0<=k<= 1
(Objective function)
vtimevtip TimeCCostCworthiness ..cos)(
25
Experimental Result
0<=k<= 1
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013
Workflow Structure:
Synthetic DAG generation
(www.loria.fr/~suter/dags.html)
Applications have between 30 and 50 tasks, generated randomly.
Total number of DAGs in our simulation is 1000.
Workflow Budget: BUDGET = C cheapest + k (CHEFT – Ccheapest)Lower budget (k=0) Cheapest scheduling, higher makespanHighest budget (k=1) shortest makespan (HEFT scheduling)
Performance Metric:
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 26
Experimental ResultSimulation Platform :
We use SIMGRID that allows a realistic description of the infrastructure parameters.
We consider a bandwidth sharing policy; only one processor can send data over one network link at a time.
We consider nodes of clusters from the GRID’5000 platform.
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 27
ResultsShopia Rennes Grenoble
HBCS Time complexity
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013 28
Conclusions (2) Conclusion remarks
We considered a realistic model of the infrastructure; The HBCS algorithm achieves better performances, in particular for
lower budget values (makespan and time complexity);
Future directions Compare other combinations of cost and time factors in the
objective function; Data locality; Multiple DAG scheduling.
29
THANK YOU!
COST IC804 – IC805 Joint meeting, Tenerife, February 7-8 2013
Top Related