1
LAYER X
IDIT LEVINE
2
INSPIRATION
3
INSPIRATION
VM
Brain
Migration Tool
Virtual machine clusters have a
scheduler that continuously monitors
the physical location of the
machines, and initiate migration if
the its benefit exceeds its cost.
4
CLUSTER MANAGEMENT
Docker Swam
kubernetes
Diego
Mesos
Frameworks
Today, none of the leading cluster management solutions implements “continuous placement”
5
MIGRATION
Migration in containers: killing the
Container on one machine and starting
a new one on the other machine.
Live migration in containers: using
Project CRIU (checkpoint restore in
userspace). Via CRIU we can write
the state of the container to the
disk (checkpoint) and restore it on
another machine.
Migration
Live
Migration
6
PLACEMENT OPTIMIZATION
Continuous placement: evaluate the placement
in the cluster and reshuffle as needed.
7
ARCHITECTURE
8
APACHE MESOS: ARCHITECTURE
Machines are registered
to Mesos master.
Mesos manages the
cluster, and offers
resources to the
frameworks.
The frameworks can
either accept or reject
offers.
9
APACHE MESOS: FRAMEWORK
Scheduling
Health
API
10
APACHE MESOS: FRAMEWORK
Scheduling
Health
API
11
APACHE MESOS: FRAMEWORK
Scheduling
Health
API
A framework makes a
decision to either accept or
to decline an offer based
only on information
pertaining to its own status,
with no information about
other frameworks.
12
APACHE MESOS: FRAMEWORK
Scheduling
Health
API
Status update is being invoke
incase of lost task. The
framework will be responsible
to spin up a new task.
13
LIMITATIONS
14
APACHE MESOS: LIMITATIONS
Scheduling
Health
API
A framework makes a
decision to either accept or
to decline an offer based
only on information
pertaining to its own status,
with no information about
other frameworks.
15
APACHE MESOS: LIMITATIONS
Can a user schedule
tasks from different
frameworks on the
same slave?
Not really.
16
APACHE MESOS: LIMITATIONS
Reservations in Mesos
are based on prioritizing
frameworks, rather than
individual tasks.
5% 80% 10% 5%
17
OSS SOLUTIONS
18
NETFLIX OOS: FENZO
This solution is hard to
integrate into existing
frameworks, and will not
work with multi
frameworks.
19
SOLUTION
20
LAYER X: ARCHITECTURE
LAYER X
21
LAYER X: MULTI BRAINS
22
LAYER X: ARCHITECTURE
LAYER X
23
LAYER X: ARCHITECTURE
LAYER X
24
LAYER X: ARCHITECTURE
LAYER X
25
USE CASES
26
LAYER X: ARCHITECTURE
LAYER X
27
LAYER X: ARCHITECTURE
LAYER X
28
LAYER X: ARCHITECTURE
LAYER X
Resource Provider Interface
Task Provider Interface
CORE
29
LAYER X: ARCHITECTURE
LAYER X
Resource Provider Interface
CORE
Task Provider Interface
30
LAYER X: ARCHITECTURE
LAYER X
Resource Provider Interface
Task Provider Interface
CORE
31
LAYER X: ARCHITECTURE
LAYER X
Task Provider Interface
CORE
Resource Provider Interface
32
LAYER X: ARCHITECTURE
US-WEST-1 US-EAST-1
LAYER X
33
LAYER X: ARCHITECTURE
LAYER X
34
SUMMARY
35
THE STACK
Resources
manager
Scheduler
Cluster
Manager
Platform as
a Service
36
EMC-Advanced-Dev
http://www.github.com/emc-advanced-dev
Idit Levine
CTO of Cloud Platform Team &
Office of the CTO, EMC
@idit_levine
Scott Weiss
Software Engineer, EMC
@ilackarms
Yuval Kohavi
Software Engineer, EMC
@uvgroovy
37
@Idit_Levine
38

Layer X