Releasing Cloud Databases from the Chains of Prediction Models Ryan - - PowerPoint PPT Presentation

▶

Apr 14, 2023 258 likes •456 views

Releasing Cloud Databases from the Chains of Prediction Models Ryan Marcus and Olga Papaemmanouil Brandeis University Cloud Databases Landscape Cloud Infrastructure as a Service (IaaS) Deployment Challenges Q Q Q Q Data Management

SLIDE 1

Releasing Cloud Databases from the Chains of Prediction Models

Ryan Marcus and Olga Papaemmanouil Brandeis University

SLIDE 2

Infrastructure as a Service (IaaS)

Cloud Databases Landscape

Cloud

SLIDE 3

IaaS Provider

Cost Management Performance Management Resource Provisioning Workload Scheduling NP-hard problem

Deployment Challenges

Data Management Application

Q Q Q Q

VM VM VM VM

SLIDE 4

Placement Provisioning Scheduling PMAX (Liu et al.) Auto (Rogers et al.) SmartSLA (Xiong et al.) Shepherd (Chi et al.) SLATree (Chi et al.) Multi-tenant SLOs (Lang et al.) iCBS (Chi et al.) Delphi / Pythia (Elmore et al.) Hypergraph (Çatalyürek et al.) SCOPE (Chaiken et al.) Bazaar (Jalaparti et al.) many traditional methods ...

State-of-the-art

SLIDE 5

Placement Provisioning Scheduling PMAX (Liu et al.) Auto (Rogers et al.) SmartSLA (Xiong et al.) Shepherd (Chi et al.) SLATree (Chi et al.) Multi-tenant SLOs (Lang et al.) iCBS (Chi et al.) Delphi / Pythia (Elmore et al.) Hypergraph (Çatalyürek et al.) SCOPE (Chaiken et al.) Bazaar (Jalaparti et al.) many traditional methods ...

State-of-the-art

Query deadline Workload deadline Piecewise linear Average latency Percentile deadline

SLIDE 6

Performance Prediction Models

q DBMS-related challenges

q isolated vs. concurrent query execution q known vs unseen query types (“templates”) q extensive off-line training q state-of-the-art: 15-20% prediction error

q Cloud-related challenges

q numerous resource configurations q dynamic environment: “noisy neighbors”

SLIDE 7

Wish List

Challenges

complex interactions arbitrary workloads arbitrary goals End-to-end cost-aware service

(resource provisioning, workload scheduling)

Agnostic to workload characteristics

(templates, arrival rates, execution times)

Application-defined performance goals

(per query deadline, percentile, average latency, max latency )

ML approach: model dynamic, complex decisions

Dynamic resource availability arbitrary resources

SLIDE 8

Bandit: ML-Based Cost Management

IaaS Provider

Data Management Application

Cost Management SLA Management Resource Provisioning Workload Scheduling

VM VM VM VM

Q Q Q Q

SLIDE 9

Reinforcement Learning

VM VM IaaS Provider VM action reward

Environment

internal state (past experiences)

bservation

agent

SLIDE 10

internal state (past experiences)

CMABs

(Contextual Multi-Armed Bandits)

VM VM IaaS Provider VM

Environment

action reward

bservation

agent

SLIDE 11

CMABs in Bandit

(Contextual Multi-Armed Bandits) VM VM IaaS Provider

Data Management Application

Environment

action cost $$

bservation

agent

internal state (past experiences)

Q Q Q Q

SLIDE 12

CMABs in Bandit

(Contextual Multi-Armed Bandits)

IaaS Provider

Data Management Application

action cost $$

bservation

VM VM VM VM VM Tier 1 VM Tier 2

SLA Q Q Q Q

internal state (past experiences)

SLIDE 13

CMABs in Bandit

(Contextual Multi-Armed Bandits)

IaaS Provider

Data Management Application

action cost $$

bservation

VM VM VM VM VM Tier 1 VM Tier 2

Q SLA Q Q Q

internal state (past experiences)

pass down accept

SLIDE 14

CMABs in Bandit

(Contextual Multi-Armed Bandits)

IaaS Provider

Data Management Application

action cost $$

bservation

VM VM VM VM VM Tier 1 VM Tier 2

SLA Q Q Q

internal state (past experiences)

SLIDE 15

CMABs in Bandit

(Contextual Multi-Armed Bandits)

IaaS Provider

Data Management Application

action cost $$

bservation

VM VM VM VM VM Tier 1 VM Tier 2

Q SLA Q Q Q

(pass, context, $$) (down, context, $$) (accept, context, $)

SLIDE 16

Feature Selection

Data Management Application

Model Generator Context Collector Experience Collector Q Q Q Q IaaS Provider

VM VM VM VM

SLIDE 17

Probabilistic Action Selection

Data Management Application

Model Generator Context Collector Experience Collector action Q Q Q Q IaaS Provider

VM VM VM VM

SLIDE 18

Evaluation

100 200 300 400 500 1000 2000 3000 4000 5000 6000 7000

Average cost per query (1/10 cent) Queries processed

Bandit, one query at a time Bandit, one query per vCPU Bandit, two queries per vCPU Clairvoyant, one query at a time Clairvoyant, one query per vCPU Clairvoyant, two queries per vCPU 100 200 300 400 500 600 700 800 500 1000 1500 2000 2500 3000 3500 4000

Average cost per query (1/10 cent) Queries processed

All new templates at once New templates over time

200 400 600 800 1000 500 1000 1500 2000 2500 3000 3500

Average cost per query (1/10 cent) Queries processed

8 templates 80 templates 800 templates 50 100 150 200 Value-based Hash-based

Converged cost (1/10 cent) Segmentation Type

Round-robin Clairvoyant PO2 Bandit

4% cost from solutions with perfect prediction model Adapts quickly to new unseen queries templates Converges after few 1000s queries of 100s templates Learns best execution site for partitioned data

SLIDE 19

Conclusions

q Cost vs performance trade-offs are complex

q human ability to derive insight is not improving

q Benefits of ML-drive approach

q discover customized solutions q automate decision making q adapt to dynamic environments

q Future Steps

q alternative learning techniques q more advanced tasks: scheduling, data movement q learning-based database as a service (DaaS) systems