Improving Resource Utilization in Cloud using Application Placement Heuristics
-
Upload
atakan-aral -
Category
Engineering
-
view
77 -
download
3
description
Transcript of Improving Resource Utilization in Cloud using Application Placement Heuristics
Improving Resource Utilization in Cloudusing Application Placement Heuristics
Atakan Aral, Tolga Ovatman
Istanbul Technical University, Department of Computer Engineering
{aralat, ovatman}@itu.edu.tr
Motivation
• Application placement is an important concept when providing software as a service in cloud environments.
• Most of the time additional resource acquisition is preferred over migrating the applications due to the high cost of migration.
• This situation results in under-utilized resources.
2
Problem
• Inefficient distribution of cloud applica-tions to virtual machines.
3
Aim
• Optimally placing applications to VMs
• Maximize the number of applications assigned.
• Minimize the number of application migrations.
4
Candidate Solutions
• Greedy algorithms (e.g. Round-Robin)
– Easy to implement
– Does not produce good placements
• Linear Optimization (e.g. MIP)
– Guaranteed to produce optimal placement
– Costly, Migrations required
• Evenness Heuristics
– May produce good enough placements
– There is no universally accepted heuristic
5
Suggested Solution
• Assign using an intelligent heuristic when possible
• Do not make any migrations yet
• When no VM is able to hold the arriving application, use Mixed Integer Programming
• Minimize number of migrations as the objective function
6
Suggested Solution
7
Contribution
• Novel evenness heuristics
• Comparison of the heuristics with other approaches
• Evaluation of the heuristic approach on various VM counts and capacities
• MIP formulation that minimizes number of migrations
8
Heuristics
• Standard Deviation (SD)
– Natural choice but may miss outliers
• Span (SP)
– Focuses only on the outliers
• Cumulative Difference (CD)
– Similar to SD, only simpler
• CD from Minimum (DM)
– Considers only high outliers
• Skewness (SK) by Xiao et al
9
MIP Formulation
• Objective: Minimize number of migrations
• Constraints:
– Each app must be assigned to exactly one VM:
– No VM should be overloaded:
10
Experimental Results
• Experiment 1: Postponement of Migration
11
Experimental Results
• Experiment 1: Postponement of Migration
12
Experimental Results
• Experiment 2: Minimization of Migration Count
13
Experimental Results
• Experiment 3: Analysis on Constant Total Capacity
14
0%1%2%3%4%5%6%7%8%9%10%11%12%13%14%15%
50,051,052,053,054,055,056,057,058,059,060,061,062,063,064,065,0
3 4 5 6 7 8 9 10 11 12
Greedy Heuristic Improvement
Conclusion
• Optimal placement in up to 10,8% of the cases, four times better than greedy
– No migrations necessary, better service quality
• Delay MIP algorithm up to 12,1% w.r.t. greedy
– More applications accepted before migrations
• MIP results in 34,5% less application migrations
15
Future Work
• Increasing the performance of cloud software as a whole by optimizing resource allocation
• Graph and Automata based modeling of the cloud environment and RA problem
• Optimization on 3 different but related phases of the cloud software life cycle
16
Future Work
17
Future Work
• MapReduce Configuration
– «The cost of using 1000 machines for 1 hour, is the same as using 1 machine for 1000 hours in the cloud paradigm.»
– Maximizing the utilization of all nodes for a Hadoop job by calculating the optimum parameters i.e. number of map/reduce
– Higher values mean higher parallelism but may cause resource contention and coordination problems.
18
Future Work
• MapReduce Configuration
– Optimum parameters depend on the resource consumption of the software.
– Design model of the software will be statically analyzed in order to guess resource consumption pattern.
– Critical (bottleneck) resources will be identified.
– Optimization at Application Level
19
Future Work
• Resource Selection and Optimization
– «In distributed computing environments, up to 85% of computing capacity remains idle due to poor optimization of placement.»
– Optimally assigning interconnected virtual nodes to substrate network with constraints
• Constraints: Datacenter capacities and bandwidth, Incompletely known topology, Locality and Jurisdiction
• Objectives: Less Inter-DC communication, Geographical proximity to user, and Latency
20
Future Work
• Resource Selection and Optimization
– VN requests and DC network are represented as undirected weighted graphs.
– Graph similarity - subgraph matching algorithms will be employed.
– Optimization at Infrastructure Level
21
Future Work
• Work Distribution to Resources
– «Data flows between nodes only in the shuffle step of the MapReduce job, and tuning it can have a big impact on job execution time.»
– Analysis of the shuffle and scheduling algorithms of Apache Hadoop framework.
– Mapper and reducer nodes should be selected carefully to minimize network traffic.
– Dynamic load balancing should be ensured.
22
Future Work
• Work Distribution to Resources
– Cost based formal analysis of the existing schedulers
– Which scheduler is more appropriate for given costs of map, reduce, shuffle phases?
– How can DataNode selection for blocks and clones be optimized to reduce network traffic?
– Optimization at Platform Level
23
Thank you!