Alluxio + Eckerson Webinar | Simplifying and Accelerating Data Access for AI/ML Model Training

Alluxio, Inc.
Alluxio, Inc.Alluxio, Inc.
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
Simplifying and Accelerating Data Access
for AI/ML Model Training
Kevin Petrie
Vice President of Research
Sridhar Venkatesh
SVP of Product
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
The Rise of Generative AI
After lots of training, language models generate strings of words
that become logical sentences and paragraphs
• A neural network whose nodes
share inputs and outputs
• Learns, summarizes, and
generates content
• Creates textual answers to
natural language questions
Source: The Economist
WHAT IS A LANGUAGE MODEL?
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
The Risk: Robots Get Things Wrong Too
Data teams must inspect, validate, and govern language model outputs
RISKS OF LANGUAGE MODELS
DATA QUALITY
Inaccuracies due to inaccurate/insufficient inputs, lack of context
EXPLAINABILITY
Vague/unknown sources or reasoning
PRIVACY
Exposure or theft due to user tracking
INTELLECTUAL PROPERTY
Liability for mishandled trademarks, copyrights, etc.
FAIRNESS
Perpetuation of bias in training data
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
Time to Get Domain Specific
Domain-specific, “small” language models reduce risk and boost productivity
by providing more governed and specialized outputs
• Enriched, detailed user
prompts
• Fine-tuned training on
enterprise data
• Augmented outputs; e.g.,
from multiple models
Small
Language
Model (SLM)
Large
Language
Model (LLM)
More
Governed
Generic Specialty
Less
Governed
ENTER THE SMALL
LANGUAGE MODEL
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
Small Language Models Will Drive the GenAI Boom
30% of data practitioners are building or training their own language models now. 20% more plan to do so*
*Source: Active LinkedIn survey of 55 respondents to date
“We believe in a world where
everyone is empowered to build
and train their own models,
imbued with their own opinions
and viewpoints.”
- Naveen Rao, Co-Founder and CEO,
MosaicML
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
“We hold these truths to be self-evident…”
TEXT
TOKENS We hold these …
VECTORS
truths to
[.45, 6.3, .99]
[7.6, .04, 19] [84, .13, 1.6]
VECTOR DB [.45, 6.3, .99] [7.6, .04, 19] [84, .13, 1.6]
LANGUAGE
MODEL
QUERY ONE QUERY TWO
1
2
3
4
5
Data teams must design and build new pipelines to feed their domain-specific data into language models
Data Processing for Language Models
Assemble unstructured text from various files
Convert words and punctuation marks to tokens
Use embeddings to convert tokens into numerical
vectors that describe their semantics
Load, organize, and index these vectors in a vector
database
Use a language model to search and query the
vectors while responding to real-time user prompts
NEW DATA PIPELINE
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
ON PREMISES | HYBRID | CLOUD | MULTI CLOUD
GENERATIVE AI EMBEDDED
CATALOG GOVERN
OBSERVE
INTEGRATE MASTER
STRUCTURED DATA
(DB TABLES)
SEMI STRUCTURED
(LOGS, CLICKSTREAMS, SENSORS…)
UNSTRUCTURED
(TEXT, IMAGES…)
CATALOG
INTEGRATE MASTER
ANALYTICS OPERATIONS
As companies embed generative AI into their workflows, they must manage
and process multi-structured data in a more holistic and efficient manner
The New Generative AI Data Stack
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
AI/ML Initiatives Need Fast and Simple Data Access
AI/ML initiatives require companies to balance, optimize, and secure workloads
across distributed datasets and compute resources
• Data access. View and process data wherever
it resides
• Performance. Retrieve data with low
latency/high throughput
• Portability. Run applications wherever suitable
compute resides
• Cost visibility. Monitor and control compute
cycles
• Multi tenancy. Isolate application compute to
safeguard performance
• Security. Restrict data access to minimize risk
of breaches
REQUIREMENTS
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
The Data Access Layer: Architecture
The data access layer continuously adjusts workloads, storage, and compute
• Namespace. Unified interface for
all data access
• APIs. Dynamic communication
between applications and storage
• Caching. Tier data by priority:
memory, SSDs, object store
• Metadata. Centralize descriptions
of data objects and resources
• Security. Authenticate users,
authorize access, log actions
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
Use Cases
AI initiatives have a range of use cases that require fast and simple data access
DATA CENTER CLOUD 1 CLOUD 2
ANALYTICS & AI IN A HYBRID ENVIRONMENT
ANALYTICS & AI ACROSS CLOUDS
WORKLOAD BURSTS
PROJECT EXPANSIONS
MIGRATIONS
COST OPTIMIZATION
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
Guiding Principles
Data leaders should evaluate where and how to implement a data access layer
to support generative AI initiatives
FIND THE BOTTLENECK
DECIDE WHETHER TO BUILD OR BUY
PLAN FOR GROWTH
© Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com
Questions?
I’m listening!
13
Retooling the enterprise data infrastructure
Legacy data centers can’t keep up
High Performance
Computing
Specialized
Hardware
Varied
Workloads
We're seeing incredible orders to retool
the world's data centers… a 10-year
transition to basically recycle or reclaim
the world's data centers and build it out
as accelerated computing.
Jensen Huang
Nvidia CEO
“
Challenges as you try to scale
14
GPUs are this year’s
toilet paper.
Wall Street Journal
“
GPUs are
scarce
GPUs are
expensive
Low GPU
Utilization
Business Pressures Complex & Costly Solutions
GPUs are
scarce
GPUs are
expensive
Low GPU
Utilization
Faster model
development times
Increased
freshness
Higher accuracy
and traceability
Rapidly growing
datasets
Extensive data engineering
managing data copies
Specialized storage
Out of control cloud and
infra costs
15
16
Alluxio Data Platform
High Performance data access, unified global view
1.Faster Time-to-Market
50%
Hundreds of thousands of dollars saved annually
compared to previous deployment.
2-3X Model Training Performance Cost Reduction, Performance Boost
International B2C with a multi-cloud, cross-region AI platform, serving LLMs and training
models from object storage. They optimized their AI platform with Alluxio to speed data
delivery to training clusters and facilitate faster model deployment in latency sensitive
production use cases.
Models Deployed in Minutes vs Days
Faster model deployment times
2. Higher GPU Utilization
“In a cloud environment, where GPU hardware is paid for as a function of time, you need
fast, performant, reliable, and cost effective data for your model training pipelines to keep
your GPU utilization close to 99%.”
20-30%
Average reported GPU utilization
based on direct access from remote
storage
GPU Utilization accessing commodity storage
GPU Utilization accessing Alluxio
Alluxio serves high throughput data to K8s training
workloads.
90
%
GPU utilization from Alluxio serving
data pulled from object storage. In
increase from 50% utilization via s3fs-
fuse.
3. Reduction in Personnel
Increase in Productivity
Pre-Processed
Data
Data
Management
Pre-
Processed
Data
Training
Clusters
Data scientists
send requests to
AI platform
teams. Platform
teams set up
individual data
pipelines.
With Alluxio, data
scientists just
access their data.
Alluxio
consolidates many
pipelines into an
access layer.
Pipeline or
Scheduler
Training
Clusters
20
4. Reduction in Infrastructure Spend
Alluxio optimizes data platforms to increase efficiency
Data Engineering
Pipelines
Data workflows improved by on-
demand access from Alluxio cache
S3 Egress and API
Fees
Fees significantly reduced via
granular caching and data
reuse
High Performance
Computing
Replaceable with low-cost hardware
at comparable performance
Reduced or Eliminated
Network Congestion Network congestion reduced by
serving files locally
5. Cloud Vendor Leverage
Multi-cloud strategies with cost-effective benefits
Respond to Limited GPU Availability
Demand for GPUs has exploded
Organizations use Alluxio to supply high performance data access
to remote GPU clusters wherever they find capacity.
Increase Cloud Agility
Competing CSPs may provide attractive discounts
Alluxio empowers organizations to capitalize on hardware discounts
or cost-effective storage in real-time. Users access data wherever it
resides.
Avoid Vendor Lock-In
Negotiate with CSPs from a stronger position
Single cloud deployments are convenient, but that may become an
obstacle in negotiations. Alluxio facilitates hybrid and multi-cloud.
Twitter.com/alluxio
Linkedin.com/alluxio
Website
www.alluxio.io
Slack
http://slackin.alluxio.io/
@
Social Media
Q&A
1 sur 22

Recommandé

2022 Trends in Enterprise Analytics par
2022 Trends in Enterprise Analytics2022 Trends in Enterprise Analytics
2022 Trends in Enterprise AnalyticsDATAVERSITY
511 vues36 diapositives
Curiosity and fourTheorem present: From Coverage Guesswork to Targeted Test G... par
Curiosity and fourTheorem present: From Coverage Guesswork to Targeted Test G...Curiosity and fourTheorem present: From Coverage Guesswork to Targeted Test G...
Curiosity and fourTheorem present: From Coverage Guesswork to Targeted Test G...Curiosity Software Ireland
29 vues33 diapositives
Data Science Salon: Applying Machine Learning to Modernize Business Processes par
Data Science Salon: Applying Machine Learning to Modernize Business ProcessesData Science Salon: Applying Machine Learning to Modernize Business Processes
Data Science Salon: Applying Machine Learning to Modernize Business ProcessesFormulatedby
783 vues22 diapositives
Analytics and Lakehouse Integration Options for Oracle Applications par
Analytics and Lakehouse Integration Options for Oracle ApplicationsAnalytics and Lakehouse Integration Options for Oracle Applications
Analytics and Lakehouse Integration Options for Oracle ApplicationsRay Février
423 vues45 diapositives
Data science lab enabling flexibility par
Data science lab   enabling flexibilityData science lab   enabling flexibility
Data science lab enabling flexibilityKognitio
362 vues18 diapositives
Estimating the Total Costs of Your Cloud Analytics Platform par
Estimating the Total Costs of Your Cloud Analytics PlatformEstimating the Total Costs of Your Cloud Analytics Platform
Estimating the Total Costs of Your Cloud Analytics PlatformDATAVERSITY
335 vues39 diapositives

Contenu connexe

Similaire à Alluxio + Eckerson Webinar | Simplifying and Accelerating Data Access for AI/ML Model Training

Privacy preserving public auditing for secured cloud storage par
Privacy preserving public auditing for secured cloud storagePrivacy preserving public auditing for secured cloud storage
Privacy preserving public auditing for secured cloud storagedbpublications
32 vues9 diapositives
Modernizing the Legacy Data Warehouse – What, Why, and How 1.23.19 par
Modernizing the Legacy Data Warehouse – What, Why, and How 1.23.19Modernizing the Legacy Data Warehouse – What, Why, and How 1.23.19
Modernizing the Legacy Data Warehouse – What, Why, and How 1.23.19Cloudera, Inc.
2.5K vues35 diapositives
Building a Logical Data Fabric using Data Virtualization (ASEAN) par
Building a Logical Data Fabric using Data Virtualization (ASEAN)Building a Logical Data Fabric using Data Virtualization (ASEAN)
Building a Logical Data Fabric using Data Virtualization (ASEAN)Denodo
272 vues40 diapositives
Data Architecture Best Practices for Advanced Analytics par
Data Architecture Best Practices for Advanced AnalyticsData Architecture Best Practices for Advanced Analytics
Data Architecture Best Practices for Advanced AnalyticsDATAVERSITY
921 vues41 diapositives
Big Data Fabric: A Necessity For Any Successful Big Data Initiative par
Big Data Fabric: A Necessity For Any Successful Big Data InitiativeBig Data Fabric: A Necessity For Any Successful Big Data Initiative
Big Data Fabric: A Necessity For Any Successful Big Data InitiativeDenodo
588 vues24 diapositives
(ENT211) Migrating the US Government to the Cloud | AWS re:Invent 2014 par
(ENT211) Migrating the US Government to the Cloud | AWS re:Invent 2014(ENT211) Migrating the US Government to the Cloud | AWS re:Invent 2014
(ENT211) Migrating the US Government to the Cloud | AWS re:Invent 2014Amazon Web Services
1.3K vues17 diapositives

Similaire à Alluxio + Eckerson Webinar | Simplifying and Accelerating Data Access for AI/ML Model Training(20)

Privacy preserving public auditing for secured cloud storage par dbpublications
Privacy preserving public auditing for secured cloud storagePrivacy preserving public auditing for secured cloud storage
Privacy preserving public auditing for secured cloud storage
dbpublications32 vues
Modernizing the Legacy Data Warehouse – What, Why, and How 1.23.19 par Cloudera, Inc.
Modernizing the Legacy Data Warehouse – What, Why, and How 1.23.19Modernizing the Legacy Data Warehouse – What, Why, and How 1.23.19
Modernizing the Legacy Data Warehouse – What, Why, and How 1.23.19
Cloudera, Inc.2.5K vues
Building a Logical Data Fabric using Data Virtualization (ASEAN) par Denodo
Building a Logical Data Fabric using Data Virtualization (ASEAN)Building a Logical Data Fabric using Data Virtualization (ASEAN)
Building a Logical Data Fabric using Data Virtualization (ASEAN)
Denodo 272 vues
Data Architecture Best Practices for Advanced Analytics par DATAVERSITY
Data Architecture Best Practices for Advanced AnalyticsData Architecture Best Practices for Advanced Analytics
Data Architecture Best Practices for Advanced Analytics
DATAVERSITY921 vues
Big Data Fabric: A Necessity For Any Successful Big Data Initiative par Denodo
Big Data Fabric: A Necessity For Any Successful Big Data InitiativeBig Data Fabric: A Necessity For Any Successful Big Data Initiative
Big Data Fabric: A Necessity For Any Successful Big Data Initiative
Denodo 588 vues
(ENT211) Migrating the US Government to the Cloud | AWS re:Invent 2014 par Amazon Web Services
(ENT211) Migrating the US Government to the Cloud | AWS re:Invent 2014(ENT211) Migrating the US Government to the Cloud | AWS re:Invent 2014
(ENT211) Migrating the US Government to the Cloud | AWS re:Invent 2014
Insurtech, Cloud and Cybersecurity - Chartered Insurance Institute par Henrique Centieiro
Insurtech, Cloud and Cybersecurity -  Chartered Insurance InstituteInsurtech, Cloud and Cybersecurity -  Chartered Insurance Institute
Insurtech, Cloud and Cybersecurity - Chartered Insurance Institute
THE SURVEY ON REFERENCE MODEL FOR OPEN STORAGE SYSTEMS INTERCONNECTION MASS S... par IRJET Journal
THE SURVEY ON REFERENCE MODEL FOR OPEN STORAGE SYSTEMS INTERCONNECTION MASS S...THE SURVEY ON REFERENCE MODEL FOR OPEN STORAGE SYSTEMS INTERCONNECTION MASS S...
THE SURVEY ON REFERENCE MODEL FOR OPEN STORAGE SYSTEMS INTERCONNECTION MASS S...
IRJET Journal2 vues
IRJET - Efficient and Verifiable Queries over Encrypted Data in Cloud par IRJET Journal
 IRJET - Efficient and Verifiable Queries over Encrypted Data in Cloud IRJET - Efficient and Verifiable Queries over Encrypted Data in Cloud
IRJET - Efficient and Verifiable Queries over Encrypted Data in Cloud
IRJET Journal12 vues
Infrastructure Solutions for Deploying AI/ML/DL Workloads at Scale par Robb Boyd
Infrastructure Solutions for Deploying AI/ML/DL Workloads at ScaleInfrastructure Solutions for Deploying AI/ML/DL Workloads at Scale
Infrastructure Solutions for Deploying AI/ML/DL Workloads at Scale
Robb Boyd357 vues
Advanced Analytics and Machine Learning with Data Virtualization (India) par Denodo
Advanced Analytics and Machine Learning with Data Virtualization (India)Advanced Analytics and Machine Learning with Data Virtualization (India)
Advanced Analytics and Machine Learning with Data Virtualization (India)
Denodo 109 vues
Building ML Pipelines with DCOS par QAware GmbH
Building ML Pipelines with DCOSBuilding ML Pipelines with DCOS
Building ML Pipelines with DCOS
QAware GmbH277 vues
Databricks on AWS.pptx par Wasm1953
Databricks on AWS.pptxDatabricks on AWS.pptx
Databricks on AWS.pptx
Wasm1953164 vues
Building Data Ecosystems for Accelerated Discovery par adamkraut
Building Data Ecosystems for Accelerated DiscoveryBuilding Data Ecosystems for Accelerated Discovery
Building Data Ecosystems for Accelerated Discovery
adamkraut33 vues
ADV Slides: The Evolution of the Data Platform and What It Means to Enterpris... par DATAVERSITY
ADV Slides: The Evolution of the Data Platform and What It Means to Enterpris...ADV Slides: The Evolution of the Data Platform and What It Means to Enterpris...
ADV Slides: The Evolution of the Data Platform and What It Means to Enterpris...
DATAVERSITY829 vues
Breed data scientists_ A Presentation.pptx par GautamPopli1
Breed data scientists_ A Presentation.pptxBreed data scientists_ A Presentation.pptx
Breed data scientists_ A Presentation.pptx
GautamPopli16 vues
BEST FINAL YEAR PROJECT IEEE 2015 BY SPECTRUM SOLUTIONS PONDICHERRY par Raushan Kumar Singh
BEST FINAL YEAR PROJECT IEEE 2015 BY SPECTRUM SOLUTIONS PONDICHERRYBEST FINAL YEAR PROJECT IEEE 2015 BY SPECTRUM SOLUTIONS PONDICHERRY
BEST FINAL YEAR PROJECT IEEE 2015 BY SPECTRUM SOLUTIONS PONDICHERRY
ADV Slides: Building and Growing Organizational Analytics with Data Lakes par DATAVERSITY
ADV Slides: Building and Growing Organizational Analytics with Data LakesADV Slides: Building and Growing Organizational Analytics with Data Lakes
ADV Slides: Building and Growing Organizational Analytics with Data Lakes
DATAVERSITY567 vues
How Data Virtualization Puts Enterprise Machine Learning Programs into Produc... par Denodo
How Data Virtualization Puts Enterprise Machine Learning Programs into Produc...How Data Virtualization Puts Enterprise Machine Learning Programs into Produc...
How Data Virtualization Puts Enterprise Machine Learning Programs into Produc...
Denodo 73 vues

Plus de Alluxio, Inc.

Alluxio Monthly Webinar | Why NFS/NAS on Object Storage May Not Solve Your AI... par
Alluxio Monthly Webinar | Why NFS/NAS on Object Storage May Not Solve Your AI...Alluxio Monthly Webinar | Why NFS/NAS on Object Storage May Not Solve Your AI...
Alluxio Monthly Webinar | Why NFS/NAS on Object Storage May Not Solve Your AI...Alluxio, Inc.
13 vues24 diapositives
AI Infra Day | Accelerate Your Model Training and Serving with Distributed Ca... par
AI Infra Day | Accelerate Your Model Training and Serving with Distributed Ca...AI Infra Day | Accelerate Your Model Training and Serving with Distributed Ca...
AI Infra Day | Accelerate Your Model Training and Serving with Distributed Ca...Alluxio, Inc.
21 vues22 diapositives
AI Infra Day | The Generative AI Market And Intel AI Strategy and Product Up... par
AI Infra Day | The Generative AI Market  And Intel AI Strategy and Product Up...AI Infra Day | The Generative AI Market  And Intel AI Strategy and Product Up...
AI Infra Day | The Generative AI Market And Intel AI Strategy and Product Up...Alluxio, Inc.
37 vues9 diapositives
Alluxio Monthly Webinar | Efficient Data Loading for Model Training on AWS par
Alluxio Monthly Webinar | Efficient Data Loading for Model Training on AWSAlluxio Monthly Webinar | Efficient Data Loading for Model Training on AWS
Alluxio Monthly Webinar | Efficient Data Loading for Model Training on AWSAlluxio, Inc.
74 vues28 diapositives
Alluxio Product school Webinar - Distributed Caching for Generative AI par
Alluxio Product school Webinar - Distributed Caching for Generative AIAlluxio Product school Webinar - Distributed Caching for Generative AI
Alluxio Product school Webinar - Distributed Caching for Generative AIAlluxio, Inc.
115 vues15 diapositives
Alluxio Product School Webinar - Get Started with Alluxio on Kubernetes par
Alluxio Product School Webinar - Get Started with Alluxio on KubernetesAlluxio Product School Webinar - Get Started with Alluxio on Kubernetes
Alluxio Product School Webinar - Get Started with Alluxio on KubernetesAlluxio, Inc.
70 vues14 diapositives

Plus de Alluxio, Inc.(20)

Alluxio Monthly Webinar | Why NFS/NAS on Object Storage May Not Solve Your AI... par Alluxio, Inc.
Alluxio Monthly Webinar | Why NFS/NAS on Object Storage May Not Solve Your AI...Alluxio Monthly Webinar | Why NFS/NAS on Object Storage May Not Solve Your AI...
Alluxio Monthly Webinar | Why NFS/NAS on Object Storage May Not Solve Your AI...
Alluxio, Inc.13 vues
AI Infra Day | Accelerate Your Model Training and Serving with Distributed Ca... par Alluxio, Inc.
AI Infra Day | Accelerate Your Model Training and Serving with Distributed Ca...AI Infra Day | Accelerate Your Model Training and Serving with Distributed Ca...
AI Infra Day | Accelerate Your Model Training and Serving with Distributed Ca...
Alluxio, Inc.21 vues
AI Infra Day | The Generative AI Market And Intel AI Strategy and Product Up... par Alluxio, Inc.
AI Infra Day | The Generative AI Market  And Intel AI Strategy and Product Up...AI Infra Day | The Generative AI Market  And Intel AI Strategy and Product Up...
AI Infra Day | The Generative AI Market And Intel AI Strategy and Product Up...
Alluxio, Inc.37 vues
Alluxio Monthly Webinar | Efficient Data Loading for Model Training on AWS par Alluxio, Inc.
Alluxio Monthly Webinar | Efficient Data Loading for Model Training on AWSAlluxio Monthly Webinar | Efficient Data Loading for Model Training on AWS
Alluxio Monthly Webinar | Efficient Data Loading for Model Training on AWS
Alluxio, Inc.74 vues
Alluxio Product school Webinar - Distributed Caching for Generative AI par Alluxio, Inc.
Alluxio Product school Webinar - Distributed Caching for Generative AIAlluxio Product school Webinar - Distributed Caching for Generative AI
Alluxio Product school Webinar - Distributed Caching for Generative AI
Alluxio, Inc.115 vues
Alluxio Product School Webinar - Get Started with Alluxio on Kubernetes par Alluxio, Inc.
Alluxio Product School Webinar - Get Started with Alluxio on KubernetesAlluxio Product School Webinar - Get Started with Alluxio on Kubernetes
Alluxio Product School Webinar - Get Started with Alluxio on Kubernetes
Alluxio, Inc.70 vues
Alluxio Product School Webinar - Boosting Trino Performance. par Alluxio, Inc.
Alluxio Product School Webinar - Boosting Trino Performance.Alluxio Product School Webinar - Boosting Trino Performance.
Alluxio Product School Webinar - Boosting Trino Performance.
Alluxio, Inc.132 vues
Alluxio Product School Webinar - Transparent URI par Alluxio, Inc.
Alluxio Product School Webinar - Transparent URIAlluxio Product School Webinar - Transparent URI
Alluxio Product School Webinar - Transparent URI
Alluxio, Inc.73 vues
Building a Distributed File System for the Cloud-Native Era par Alluxio, Inc.
Building a Distributed File System for the Cloud-Native EraBuilding a Distributed File System for the Cloud-Native Era
Building a Distributed File System for the Cloud-Native Era
Alluxio, Inc.136 vues
Zookeeper vs Raft: Stateful distributed coordination with HA and Fault Tolerance par Alluxio, Inc.
Zookeeper vs Raft: Stateful distributed coordination with HA and Fault ToleranceZookeeper vs Raft: Stateful distributed coordination with HA and Fault Tolerance
Zookeeper vs Raft: Stateful distributed coordination with HA and Fault Tolerance
Alluxio, Inc.518 vues
Modernize Your Data Platform for Analytics and AI par Alluxio, Inc.
Modernize Your Data Platform for Analytics and AIModernize Your Data Platform for Analytics and AI
Modernize Your Data Platform for Analytics and AI
Alluxio, Inc.149 vues
The Power of Data Orchestration: Storage Acceleration and Servitization at Sh... par Alluxio, Inc.
The Power of Data Orchestration: Storage Acceleration and Servitization at Sh...The Power of Data Orchestration: Storage Acceleration and Servitization at Sh...
The Power of Data Orchestration: Storage Acceleration and Servitization at Sh...
Alluxio, Inc.45 vues
Apache Hudi: Community-Driven Development par Alluxio, Inc.
Apache Hudi: Community-Driven DevelopmentApache Hudi: Community-Driven Development
Apache Hudi: Community-Driven Development
Alluxio, Inc.43 vues
Unified Data API for Distributed Cloud Analytics and AI par Alluxio, Inc.
Unified Data API for Distributed Cloud Analytics and AIUnified Data API for Distributed Cloud Analytics and AI
Unified Data API for Distributed Cloud Analytics and AI
Alluxio, Inc.31 vues
Real-Time Analytics: Going Beyond Stream Processing With Apache Pinot par Alluxio, Inc.
Real-Time Analytics: Going Beyond Stream Processing With Apache PinotReal-Time Analytics: Going Beyond Stream Processing With Apache Pinot
Real-Time Analytics: Going Beyond Stream Processing With Apache Pinot
Alluxio, Inc.32 vues
ML-Based SQL Query Resource Usage Prediction par Alluxio, Inc.
ML-Based SQL Query Resource Usage PredictionML-Based SQL Query Resource Usage Prediction
ML-Based SQL Query Resource Usage Prediction
Alluxio, Inc.47 vues
Accelerating Cloud Training With Alluxio par Alluxio, Inc.
Accelerating Cloud Training With AlluxioAccelerating Cloud Training With Alluxio
Accelerating Cloud Training With Alluxio
Alluxio, Inc.38 vues
Deconstructing a Machine Learning Pipeline with Virtual Data Lake par Alluxio, Inc.
Deconstructing a Machine Learning Pipeline with Virtual Data LakeDeconstructing a Machine Learning Pipeline with Virtual Data Lake
Deconstructing a Machine Learning Pipeline with Virtual Data Lake
Alluxio, Inc.82 vues

Dernier

Generic or specific? Making sensible software design decisions par
Generic or specific? Making sensible software design decisionsGeneric or specific? Making sensible software design decisions
Generic or specific? Making sensible software design decisionsBert Jan Schrijver
6 vues60 diapositives
FIMA 2023 Neo4j & FS - Entity Resolution.pptx par
FIMA 2023 Neo4j & FS - Entity Resolution.pptxFIMA 2023 Neo4j & FS - Entity Resolution.pptx
FIMA 2023 Neo4j & FS - Entity Resolution.pptxNeo4j
17 vues26 diapositives
Airline Booking Software par
Airline Booking SoftwareAirline Booking Software
Airline Booking SoftwareSharmiMehta
7 vues26 diapositives
Programming Field par
Programming FieldProgramming Field
Programming Fieldthehardtechnology
6 vues9 diapositives
BushraDBR: An Automatic Approach to Retrieving Duplicate Bug Reports par
BushraDBR: An Automatic Approach to Retrieving Duplicate Bug ReportsBushraDBR: An Automatic Approach to Retrieving Duplicate Bug Reports
BushraDBR: An Automatic Approach to Retrieving Duplicate Bug ReportsRa'Fat Al-Msie'deen
8 vues49 diapositives
Page Object Model par
Page Object ModelPage Object Model
Page Object Modelartembondar5
6 vues5 diapositives

Dernier(20)

Generic or specific? Making sensible software design decisions par Bert Jan Schrijver
Generic or specific? Making sensible software design decisionsGeneric or specific? Making sensible software design decisions
Generic or specific? Making sensible software design decisions
FIMA 2023 Neo4j & FS - Entity Resolution.pptx par Neo4j
FIMA 2023 Neo4j & FS - Entity Resolution.pptxFIMA 2023 Neo4j & FS - Entity Resolution.pptx
FIMA 2023 Neo4j & FS - Entity Resolution.pptx
Neo4j17 vues
BushraDBR: An Automatic Approach to Retrieving Duplicate Bug Reports par Ra'Fat Al-Msie'deen
BushraDBR: An Automatic Approach to Retrieving Duplicate Bug ReportsBushraDBR: An Automatic Approach to Retrieving Duplicate Bug Reports
BushraDBR: An Automatic Approach to Retrieving Duplicate Bug Reports
AI and Ml presentation .pptx par FayazAli87
AI and Ml presentation .pptxAI and Ml presentation .pptx
AI and Ml presentation .pptx
FayazAli8713 vues
20231129 - Platform @ localhost 2023 - Application-driven infrastructure with... par sparkfabrik
20231129 - Platform @ localhost 2023 - Application-driven infrastructure with...20231129 - Platform @ localhost 2023 - Application-driven infrastructure with...
20231129 - Platform @ localhost 2023 - Application-driven infrastructure with...
sparkfabrik8 vues
Top-5-production-devconMunich-2023.pptx par Tier1 app
Top-5-production-devconMunich-2023.pptxTop-5-production-devconMunich-2023.pptx
Top-5-production-devconMunich-2023.pptx
Tier1 app8 vues
DRYiCE™ iAutomate: AI-enhanced Intelligent Runbook Automation par HCLSoftware
DRYiCE™ iAutomate: AI-enhanced Intelligent Runbook AutomationDRYiCE™ iAutomate: AI-enhanced Intelligent Runbook Automation
DRYiCE™ iAutomate: AI-enhanced Intelligent Runbook Automation
HCLSoftware6 vues
Unlocking the Power of AI in Product Management - A Comprehensive Guide for P... par NimaTorabi2
Unlocking the Power of AI in Product Management - A Comprehensive Guide for P...Unlocking the Power of AI in Product Management - A Comprehensive Guide for P...
Unlocking the Power of AI in Product Management - A Comprehensive Guide for P...
NimaTorabi215 vues
2023-November-Schneider Electric-Meetup-BCN Admin Group.pptx par animuscrm
2023-November-Schneider Electric-Meetup-BCN Admin Group.pptx2023-November-Schneider Electric-Meetup-BCN Admin Group.pptx
2023-November-Schneider Electric-Meetup-BCN Admin Group.pptx
animuscrm15 vues
Bootstrapping vs Venture Capital.pptx par Zeljko Svedic
Bootstrapping vs Venture Capital.pptxBootstrapping vs Venture Capital.pptx
Bootstrapping vs Venture Capital.pptx
Zeljko Svedic14 vues
Copilot Prompting Toolkit_All Resources.pdf par Riccardo Zamana
Copilot Prompting Toolkit_All Resources.pdfCopilot Prompting Toolkit_All Resources.pdf
Copilot Prompting Toolkit_All Resources.pdf
Riccardo Zamana16 vues

Alluxio + Eckerson Webinar | Simplifying and Accelerating Data Access for AI/ML Model Training

  • 1. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com Simplifying and Accelerating Data Access for AI/ML Model Training Kevin Petrie Vice President of Research Sridhar Venkatesh SVP of Product
  • 2. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com The Rise of Generative AI After lots of training, language models generate strings of words that become logical sentences and paragraphs • A neural network whose nodes share inputs and outputs • Learns, summarizes, and generates content • Creates textual answers to natural language questions Source: The Economist WHAT IS A LANGUAGE MODEL?
  • 3. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com The Risk: Robots Get Things Wrong Too Data teams must inspect, validate, and govern language model outputs RISKS OF LANGUAGE MODELS DATA QUALITY Inaccuracies due to inaccurate/insufficient inputs, lack of context EXPLAINABILITY Vague/unknown sources or reasoning PRIVACY Exposure or theft due to user tracking INTELLECTUAL PROPERTY Liability for mishandled trademarks, copyrights, etc. FAIRNESS Perpetuation of bias in training data
  • 4. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com Time to Get Domain Specific Domain-specific, “small” language models reduce risk and boost productivity by providing more governed and specialized outputs • Enriched, detailed user prompts • Fine-tuned training on enterprise data • Augmented outputs; e.g., from multiple models Small Language Model (SLM) Large Language Model (LLM) More Governed Generic Specialty Less Governed ENTER THE SMALL LANGUAGE MODEL
  • 5. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com Small Language Models Will Drive the GenAI Boom 30% of data practitioners are building or training their own language models now. 20% more plan to do so* *Source: Active LinkedIn survey of 55 respondents to date “We believe in a world where everyone is empowered to build and train their own models, imbued with their own opinions and viewpoints.” - Naveen Rao, Co-Founder and CEO, MosaicML
  • 6. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com “We hold these truths to be self-evident…” TEXT TOKENS We hold these … VECTORS truths to [.45, 6.3, .99] [7.6, .04, 19] [84, .13, 1.6] VECTOR DB [.45, 6.3, .99] [7.6, .04, 19] [84, .13, 1.6] LANGUAGE MODEL QUERY ONE QUERY TWO 1 2 3 4 5 Data teams must design and build new pipelines to feed their domain-specific data into language models Data Processing for Language Models Assemble unstructured text from various files Convert words and punctuation marks to tokens Use embeddings to convert tokens into numerical vectors that describe their semantics Load, organize, and index these vectors in a vector database Use a language model to search and query the vectors while responding to real-time user prompts NEW DATA PIPELINE
  • 7. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com ON PREMISES | HYBRID | CLOUD | MULTI CLOUD GENERATIVE AI EMBEDDED CATALOG GOVERN OBSERVE INTEGRATE MASTER STRUCTURED DATA (DB TABLES) SEMI STRUCTURED (LOGS, CLICKSTREAMS, SENSORS…) UNSTRUCTURED (TEXT, IMAGES…) CATALOG INTEGRATE MASTER ANALYTICS OPERATIONS As companies embed generative AI into their workflows, they must manage and process multi-structured data in a more holistic and efficient manner The New Generative AI Data Stack
  • 8. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com AI/ML Initiatives Need Fast and Simple Data Access AI/ML initiatives require companies to balance, optimize, and secure workloads across distributed datasets and compute resources • Data access. View and process data wherever it resides • Performance. Retrieve data with low latency/high throughput • Portability. Run applications wherever suitable compute resides • Cost visibility. Monitor and control compute cycles • Multi tenancy. Isolate application compute to safeguard performance • Security. Restrict data access to minimize risk of breaches REQUIREMENTS
  • 9. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com The Data Access Layer: Architecture The data access layer continuously adjusts workloads, storage, and compute • Namespace. Unified interface for all data access • APIs. Dynamic communication between applications and storage • Caching. Tier data by priority: memory, SSDs, object store • Metadata. Centralize descriptions of data objects and resources • Security. Authenticate users, authorize access, log actions
  • 10. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com Use Cases AI initiatives have a range of use cases that require fast and simple data access DATA CENTER CLOUD 1 CLOUD 2 ANALYTICS & AI IN A HYBRID ENVIRONMENT ANALYTICS & AI ACROSS CLOUDS WORKLOAD BURSTS PROJECT EXPANSIONS MIGRATIONS COST OPTIMIZATION
  • 11. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com Guiding Principles Data leaders should evaluate where and how to implement a data access layer to support generative AI initiatives FIND THE BOTTLENECK DECIDE WHETHER TO BUILD OR BUY PLAN FOR GROWTH
  • 12. © Eckerson Group 2021 Twitter: @eckersongroup www.eckerson.com Questions? I’m listening!
  • 13. 13 Retooling the enterprise data infrastructure Legacy data centers can’t keep up High Performance Computing Specialized Hardware Varied Workloads We're seeing incredible orders to retool the world's data centers… a 10-year transition to basically recycle or reclaim the world's data centers and build it out as accelerated computing. Jensen Huang Nvidia CEO “
  • 14. Challenges as you try to scale 14 GPUs are this year’s toilet paper. Wall Street Journal “ GPUs are scarce GPUs are expensive Low GPU Utilization
  • 15. Business Pressures Complex & Costly Solutions GPUs are scarce GPUs are expensive Low GPU Utilization Faster model development times Increased freshness Higher accuracy and traceability Rapidly growing datasets Extensive data engineering managing data copies Specialized storage Out of control cloud and infra costs 15
  • 16. 16 Alluxio Data Platform High Performance data access, unified global view
  • 17. 1.Faster Time-to-Market 50% Hundreds of thousands of dollars saved annually compared to previous deployment. 2-3X Model Training Performance Cost Reduction, Performance Boost International B2C with a multi-cloud, cross-region AI platform, serving LLMs and training models from object storage. They optimized their AI platform with Alluxio to speed data delivery to training clusters and facilitate faster model deployment in latency sensitive production use cases. Models Deployed in Minutes vs Days Faster model deployment times
  • 18. 2. Higher GPU Utilization “In a cloud environment, where GPU hardware is paid for as a function of time, you need fast, performant, reliable, and cost effective data for your model training pipelines to keep your GPU utilization close to 99%.” 20-30% Average reported GPU utilization based on direct access from remote storage GPU Utilization accessing commodity storage GPU Utilization accessing Alluxio Alluxio serves high throughput data to K8s training workloads. 90 % GPU utilization from Alluxio serving data pulled from object storage. In increase from 50% utilization via s3fs- fuse.
  • 19. 3. Reduction in Personnel Increase in Productivity Pre-Processed Data Data Management Pre- Processed Data Training Clusters Data scientists send requests to AI platform teams. Platform teams set up individual data pipelines. With Alluxio, data scientists just access their data. Alluxio consolidates many pipelines into an access layer. Pipeline or Scheduler Training Clusters
  • 20. 20 4. Reduction in Infrastructure Spend Alluxio optimizes data platforms to increase efficiency Data Engineering Pipelines Data workflows improved by on- demand access from Alluxio cache S3 Egress and API Fees Fees significantly reduced via granular caching and data reuse High Performance Computing Replaceable with low-cost hardware at comparable performance Reduced or Eliminated Network Congestion Network congestion reduced by serving files locally
  • 21. 5. Cloud Vendor Leverage Multi-cloud strategies with cost-effective benefits Respond to Limited GPU Availability Demand for GPUs has exploded Organizations use Alluxio to supply high performance data access to remote GPU clusters wherever they find capacity. Increase Cloud Agility Competing CSPs may provide attractive discounts Alluxio empowers organizations to capitalize on hardware discounts or cost-effective storage in real-time. Users access data wherever it resides. Avoid Vendor Lock-In Negotiate with CSPs from a stronger position Single cloud deployments are convenient, but that may become an obstacle in negotiations. Alluxio facilitates hybrid and multi-cloud.