Decoupling
compute and storage
for data workloads
Carlos Queiroz
Data processing workloads at DBS
Hadoop introduced in 2015 Business and Regulatory
Reporting
DataWareHouse replacement? Analytics
datanode
JVM
DataNode1
datanode
JVM
DataNode2
datanode
JVM
DataNodeX
…
namenode
JVM
NameNode
namenode
JVM
NameNode
ETL Batch
Bare-Metal
Data Locality
HDFS on
local disks
Enterprise transactions
Logs
mainframe
H
D
F
S
ETL Processing
Data Science
H
D
F
S
User
ETL
ETL
Why decoupling?
Current model
• Hard to scale

• Scale Compute AND Storage

• It is not flexible

• Costs
Bare-Metal
Data Locality
HDFS on
local disks
Also in 2015
EMC and Adobe bringing HDaaS
https://www.brighttalk.com/webcast/1744/156173
Decoupling compute and storage
Bare-Metal
Data Locality
HDFS on
local disks
Containers and VMs
Separate Compute
and Storage
Shared Storage
Data as a Service
Agility and cost
savings
Faster time to
foresights
Traditional Assumptions A New Approach Benefits and Value
https://www.bluedata.com/blog/2015/12/separating-hadoop-compute-and-storage/Adapted from
Fast Forward to 2017
Re-engineering the data platform
StorageCompute
DataIngestion
Decisionsupport
Object store
In-memory
Filesystem
Compute engine I
Compute Engine II
Compute Engine III
Compute Engine IV
…
Fast Forward to 2017
Storage
Object store
In-memory
Filesystem
Compute
Compute engine I Compute Engine II
Compute
Compute engine I
Compute
Compute engine I Compute Engine II Compute Engine X
Multi-tenancy Different SLAs Different Engines Different Cluster sizes
Implementing it
In-memory filesystem
• Apps only talk to Alluxio
• Simple Add/Remove
• No App Changes
• Highest performance in
Memory
Alluxio - Server side API translation
HDFS Interface
S3A Interface S3 compatible
S3 compatible software
Reference implementation
StorageCompute
Current implementation status
• Development environment with 50 VMs 

• Running benchmarks for performance
evaluation 

• Cloudera 5.13.x

• S3 compatible object store
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
vCPU: 12
RAM: 128
Disk: 400GB
Questions???

Decoupling Compute and Storage for Data Workloads

  • 1.
    Decoupling compute and storage fordata workloads Carlos Queiroz
  • 2.
    Data processing workloadsat DBS Hadoop introduced in 2015 Business and Regulatory Reporting DataWareHouse replacement? Analytics datanode JVM DataNode1 datanode JVM DataNode2 datanode JVM DataNodeX … namenode JVM NameNode namenode JVM NameNode ETL Batch Bare-Metal Data Locality HDFS on local disks Enterprise transactions Logs mainframe H D F S ETL Processing Data Science H D F S User ETL ETL
  • 3.
  • 4.
    Current model • Hardto scale • Scale Compute AND Storage • It is not flexible • Costs Bare-Metal Data Locality HDFS on local disks
  • 5.
    Also in 2015 EMCand Adobe bringing HDaaS https://www.brighttalk.com/webcast/1744/156173
  • 6.
    Decoupling compute andstorage Bare-Metal Data Locality HDFS on local disks Containers and VMs Separate Compute and Storage Shared Storage Data as a Service Agility and cost savings Faster time to foresights Traditional Assumptions A New Approach Benefits and Value https://www.bluedata.com/blog/2015/12/separating-hadoop-compute-and-storage/Adapted from
  • 7.
    Fast Forward to2017 Re-engineering the data platform StorageCompute DataIngestion Decisionsupport Object store In-memory Filesystem Compute engine I Compute Engine II Compute Engine III Compute Engine IV …
  • 8.
    Fast Forward to2017 Storage Object store In-memory Filesystem Compute Compute engine I Compute Engine II Compute Compute engine I Compute Compute engine I Compute Engine II Compute Engine X Multi-tenancy Different SLAs Different Engines Different Cluster sizes
  • 9.
  • 10.
    In-memory filesystem • Appsonly talk to Alluxio • Simple Add/Remove • No App Changes • Highest performance in Memory
  • 11.
    Alluxio - Serverside API translation HDFS Interface S3A Interface S3 compatible S3 compatible software
  • 12.
  • 13.
    Current implementation status •Development environment with 50 VMs • Running benchmarks for performance evaluation • Cloudera 5.13.x • S3 compatible object store vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB vCPU: 12 RAM: 128 Disk: 400GB
  • 14.