Kantar AI Summit- Under Embargo till Wednesday, 24th April 2024, 4 PM, IST.pdf
Vitali@SCC2018
1. Virtual Machine Profiling for Analyzing
Resource Usage of Applications
Xuesong Peng, Barbara Pernici, and Monica Vitali
monica.vitali@polimi.it
vitali.faculty.polimi.it
2018 International Conference on Services Computing
2. CHALLENGE
VMs and containers are black box for the data center administrator
Knowing the behaviour of applications in a data center is fundamental for
driving successful deployment and adaptation decisions
Prediction is even more important in Edge Computing due to resource
shortage
Monitoring information is the only source of knowledge
GOAL
Provide a methodology for building an application profile reflecting relevant
behavioral features of a VM using only monitoring information
3. APPROACH
Building an application profile from monitoring data generated during the
application execution
Enables…
● Supporting deployment decisions
● Detecting anomalies
● Classifying homogeneous VMs in terms of resource usage and patterns of usage in
time
The profile captures the dynamic behavior and considers: intensiveness in
resource usage and periodicity of the VM behavior
5. RESOURCES INTENSIVENESS
GOAL classify VMs in three groups (invensive, medium-intensive, non
intensive)
Metrics used for classification
● the average resource usage of all the samples of a metric in the dataset;
● the percentage of samples of a specific metric which exceed a warning threshold
● the percentage of samples of a specific metric which exceed a critical threshold
To get appropriate thresholds we refer to the literature (DCMM and VMWare)
E.g., CPU > 75% for 5 minutes -> warning MEM > 85% for 10 minutes -> warning
CPU > 90% for 5 minutes -> critical MEM > 95% for 10 minutes -> critical
12. PERIODICITY
Step 1: detect relevant periods (for each metric of each VM)
Extract candidate
periodicities using DFT
Refine periods using
auto-correlation
13. PERIODICITY
Step 2: extract typical shape for the selected period (average of all the
instances for the period)
15. VM PROFILE - EXAMPLES
VM 1
INTENSIVENESS
CPU MEM BW IO
Medium no no medium
PERIODICITIES
CPU MEM BW IO
1 day 1 day no 0.3 days
7 days
VM 2
INTENSIVENESS
CPU MEM BW IO
Medium no medium no
PERIODICITIES
CPU MEM BW IO
1 day 1 day 1 day 1 day
7 days 3.5 days
7 days
NB: Each period is associated with an average pattern
16. RESULTS - Intensiveness analysis
Applications are mainly either CPU
or memory intensive (rarely both)
Applications that are both CPU and
memory intensive are the most
critical
17. RESULTS - Periodicity analysis
30% of applications are both
memory and CPU periodic
Only 16% of applications are
not-periodic
M
EM
CPU
20. FINAL REMARKS
RESULTS
A methodology for extracting profiles (resource intensiveness + periodic
behaviour) of VMs and containers from monitoring data
Validation with real monitoring data
Validated relations between intensiveness and periodicity and impact of
intensiveness on migrations
FUTURE WORK
Exploit profiles for VMs and containers placement, anomaly detection,
resource planning
21. Virtual Machine Profiling for Analyzing
Resource Usage of Applications
Xuesong Peng, Barbara Pernici, and Monica Vitali
monica.vitali@polimi.it
vitali.faculty.polimi.it
2018 International Conference on Services Computing