1Budapest University of Technology and EconomicsDepartment of Measurement and Information Systems

Cloud Based Analytics forCloud Based Applications

András Pataricza1, Imre Kocsis1, Zsolt Kocsis2 et al.

1Dept. of Meas. and Information Systems, BME, Hungary2IBM CAS Budapest, Hungary

{pataric,ikocsis}@mit.bme.huICA CON 2012, April 20, 2012

2

Clouds for demanding applications?

Standard infrastructure vs

demanding application?

Standard infrastructure vs

demanding application?

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

Test automation

Hypervisor

Interference

Lab

OS and hypervisor

metrics

OS and hypervisor

metrics

LOLO

HIHI

Experimental setupN.B.: VMware R&D published

similar(March 2012)

N.B.: VMware R&D published

similar(March 2012)

5

IT EDA is Big Data!

Hypervizor (host + VMs), OS, application, ...Which determine the QoS?Which determine the QoS?

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

Rare events: lot of sand, a few pellets

Typically sand: gold mining ≠ data miningTypically sand: gold mining ≠ data mining

8

Visual analytics = causal insight

8

Computing power use = CPU use ×CPU clock rate (const.)Should be pure proportional

Correlation coefficient:0.99998477434137Well-visible, but numerically suppressed

Origin???

Computing power use = CPU use ×CPU clock rate (const.)Should be pure proportional

Correlation coefficient:0.99998477434137Well-visible, but numerically suppressed

Origin???

9

Visual analytics

Noisy…High frequency components dominateBut they correlate (93%!)YOU DON’T SEE IT

Noisy…High frequency components dominateBut they correlate (93%!)YOU DON’T SEE IT

10

Dangers in a standard cloud for demanding apps?

11

Impacts of resource sharing?

Self-inducedSelf-induced Parasitic influenceParasitic influence

12

Short transient faults – long recovery

8 sec platform overload

8 sec platform overload

30 sec service outage

30 sec service outage

120 sec SLA violation120 sec SLA violation

As if you unplug your

desktop for a second...

As if you unplug your

desktop for a second...

13

Deterministic (?!) run-time in the public cloud...

Variance tolerable by overcapacity

Variance tolerable by overcapacity

Performance outage

intolerable by overcapacity

Performance outage

intolerable by overcapacity

14

The noisy neighbour problem

HypervisorHypervisor

TenantTenant NeighborNeighbor

15

Tenant-side measurability and observability

HypervisorHypervisor

TenantTenant NeighborNeighbor

16

Let’s try it at user level

17

The mistery shopper concept Basic logic as with benchmarks, but...

Metric req:o same interference-sensitivities as the serviceo same resource-sensitivities as the serviceo 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

Example: short computation bursts sampling available CPU for longer

computation

18

Indirect platform & QoS observability

VM

metric

Mistery shopperApplication

QoS

Resource view

Interferences

Load

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. Connect1. Connect

2. Observe2. Observe

3. Infer 3. Infer (qualitatively)(qualitatively)

3. Infer 3. Infer (qualitatively)(qualitatively)

Observability problems (if present) bypassedObservability problems (if present) bypassedObservability problems (if present) bypassedObservability problems (if present) bypassed

Works Works withoutwithout the application! the application!Works Works withoutwithout the application! the application!

19

Mystery shopper & service QoS

VM internal fault

Mystery shopper

Main applicationFast detection

Reaction timewindow

Reaction timewindow

Noisy neighbour fault

Application failure

20

Summary Technical

o SLA coverage needed for all aspectsoMissing guarantees can be (somewhat) compensated• Cheap computing power -> redundancy• „Double” autonomic computing

– Cloud level – provider– Application level – user

Methodologyo Visual exploratory data analysis for insighto Algorithmic analysis for proofs and evaluationo Fault-tolerance design patterns revisited• Cheap redundancy in the cloud