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!
