1. 1 Budapest University of Technology and Economics Department of Measurement and Information Systems Cloud Based Analytics for Cloud Based Applications András Pataricza 1, Imre Kocsis 1, Zsolt Kocsis 2 et al. 1 Dept. of Meas. and Information Systems, BME, Hungary 2 IBM CAS Budapest, Hungary {pataric,ikocsis}@mit.bme.hu ICA CON 2012, April 20, 2012

2. 2 Clouds for demanding applications? Standard infrastructure vs demanding application? Standard infrastructure vs demanding application?

3. 3 Clouds for demanding applications? Virtual Desktop Infrastructure Telecommunications Extra-functional reqs: throughput, timeliness, availability „Small problems” have high impact (soft real time) Extra-functional reqs: throughput, timeliness, availability „Small problems” have high impact (soft real time)

4. 4 Experimental setup N.B.: VMware R&D published similar (March 2012) N.B.: VMware R&D published similar (March 2012)

5. 5 IT EDA is Big Data! Hypervizor (host + VMs), OS, application,... Which determine the QoS?

6. 6 IT EDA is Big Data! High availabilty, rare faults Rare events: granularity AND long horizon Searching for outliers High availabilty, rare faults Rare events: granularity AND long horizon Searching for outliers

7. 7 Rare events: lot of sand, a few pellets Typically sand: gold mining ≠ data mining

8. 8 Visual analytics = causal insight 8 Computing power use = CPU use × CPU clock rate (const.) Should be pure proportional Correlation coefficient: 0.99998477434137 Well-visible, but numerically suppressed Origin??? Computing power use = CPU use × CPU clock rate (const.) Should be pure proportional Correlation coefficient: 0.99998477434137 Well-visible, but numerically suppressed Origin???

9. 9 Visual analytics Noisy… High frequency components dominate But they correlate (93%!) YOU DON’T SEE IT Noisy… High frequency components dominate But they correlate (93%!) YOU DON’T SEE IT

10. 10 Dangers in a standard cloud for demanding apps?

11. 11 Impacts of resource sharing? Self-induced Parasitic influence

12. 12 Short transient faults – long recovery 8 sec platform overload 30 sec service outage 120 sec SLA violation As if you unplug your desktop for a second...

13. 13 Deterministic (?!) run-time in the public cloud... Variance tolerable by overcapacity Performance outage intolerable by overcapacity

14. 14 The noisy neighbour problem Hypervisor Tenant Neighbor

15. 15 Tenant-side measurability and observability Hypervisor Tenant Neighbor

16. 16 Let’s try it at user level

17. 17 The mistery shopper concept  Basic logic as with benchmarks, but...  Metric req: o same interference-sensitivities as the service o same resource-sensitivities as the service o representative for types of services  Runtime req: o Non-intrusiveness (instead of saturation) o Long running (rare events) o (Low specific impact on service) Not trivially feasible... but everything else impossible Not trivially feasible... but everything else impossible Example: short computation bursts sampling available CPU for longer computation

18. 18 Indirect platform & QoS observability The „classic” approach: deploy, run/test, observe, analyze The „classic” approach: deploy, run/test, observe, analyze The „classic” approach: deploy, run/test, observe, analyze The „classic” approach: deploy, run/test, observe, analyze The „classic” approach: deploy, run/test, observe, analyze The „classic” approach: deploy, run/test, observe, analyze 1. Connect 2. Observe 3. Infer (qualitatively) Observability problems (if present) bypassed Works without the application!

19. 19 Mystery shopper & service QoS VM internal fault Mystery shopper Main application Fast detection Reaction time window Reaction time window Noisy neighbour fault Application failure

20. 20 Summary  Technical o SLA coverage needed for all aspects o Missing guarantees can be (somewhat) compensated Cheap computing power -> redundancy „Double” autonomic computing – Cloud level – provider – Application level – user  Methodology o Visual exploratory data analysis for insight o Algorithmic analysis for proofs and evaluation o Fault-tolerance design patterns revisited Cheap redundancy in the cloud