CS 744: MESOS Shivaram Venkataraman Fall 2020 ADMINISTRIVIA lie - - PowerPoint PPT Presentation

CS 744: MESOS

Shivaram Venkataraman Fall 2020

good

morning

!

ADMINISTRIVIA

• Assignment 1: How did it go?
• Assignment 2 out tonight
• Project details
• Create project groups
• Bid for projects/Propose your own
• Work on Introduction

→ fill out

lie poll

!

→

distributed

• N 3 students

next

week

↳

1

• 2

page

• check
• in
• Final

report

and

poster

session

COURSE FORMAT

Paper reviews “Compare, contrast and evaluate research papers” Discussion

Scalable Storage Systems Datacenter Architecture Resource Management Computational Engines Machine Learning SQL Streaming Graph Applications

Assignment

T

→

d park

→ GFS

• →

MapReduce GFS Spark

=

BACKGROUND: OS SCHEDULING

code, static data heap stack code, static data heap stack code, static data heap stack CPU

How do we share CPU between processes ?

pt

P2

B

Evin

time sharing

rim

for

• - 10ms

go

for

;

• as

an¥777

• time

CLUSTER SCHEDULING

Scale

large

number of machines

↳

• ne

scheduler

?

Fairness

• '

naff:h

space

searing

multi

• WT

fault

"""1M€

time

• r

/

preferences (placement

)

pump

scheduling

↳

TARGET ENVIRONMENT

Multiple MapReduce versions Mix of frameworks: MPI, Spark, MR Data sharing across frameworks Avoid per-framework clusters

utilization

used

g→

{

Different

kinds

• f

applications

• n

cluster

• Faris

hankering .

L

t !

• in

:

F

¥

DESIGN

ars

Two

• level scheduling

awww

stoked

• ¥÷m5onYs

I

↳ scheduling

framework

Single

master-

scheduler

fi¥

ME

wide

scheduler

fret

^

↳ Add

www.oibi

"

↳

Scalability

RESOURCE OFFERS

¥m!:#

ant

Dared

7

reply

• ffering

===

zcpuisgb

c-

ri:

he:*

====

←

CONSTRAINTS

Examples of constraints Constraints in Mesos:

• Dita

locality

soft

Gpu

machines

hard

↳

frameworks

reject

• ffer

↳

fitters

Boolean

functions

DESIGN DETAILS

Allocation: Guaranteed allocation, revocation Isolation Containers (Docker)

Dai

T

• ,kfd f To Hers

↳

short . lived

tasks ! L

running task

• '

be

pre

• empted ,

podcefya.me

" 4¥

Other

frameworks

express

interest

FAULT TOLERANCE

¥:&

.

son

l

ft

master failure

+adf.qt.ws ¥

doesnt affect

jobs

# heartbeat

PLACEMENT PREFERENCES

What is the problem? How do we do allocations?

↳ If

you

frameworks

than machines

available

cluster

↳

weighted

lottery

scheme

make

• ffers

µportioned

in size to

• verall

needs

CENTRALIZED VS DECENTRALIZED

Centralized

Decentralised

• f frameworks

each

rloos

• f apps
• ptimal

solution

handle

1

Complexity

for framework developer

CENTRALIZED VS DECENTRALIZED

Framework complexity Fragmentation, Starvation Inter-dependent framework ✓

→ If

resource

• ffers

too

small

COMPARISON: YARN

Per-job scheduler AM asks for resource RM replies

→ Apache Hadoop

ng

matter

¥g

per

framework

⇐ -

• Fer - job

scheduler

COMPARISON: BORG

Single centralized scheduler Requests mem, cpu in cfg Priority per user / service Support for quotas / reservations

→ Google

• I Better

packing

SUMMARY

• Mesos: Scheduler to share cluster between Spark, MR, etc.
• Two-level scheduling with app-specific schedulers
• Provides scalable, decentralized scheduling
• Pluggable Policy ? Next class!

framework

Go

DISCUSSION

https://forms.gle/urHSeukfyipCKjue6

What are some problems that could come up if we scale from 10 frameworks to 1000 frameworks in Mesos?

Fragmentation / starvation

• dds

go

up

→

Master

bottleneck

↳

it

takes

time

to

wait

for

frameworks to

reply

Mems master

→

Yes

has

soft state

→

failure

recovery

takes longer ?

why ? /

penni

:

":

pongee

.

O

O O

Rigid

framework

terror

:

y

Ihle!fain

MPI 's

List any one difference between an OS scheduler and Mesos

↳

Motivation

part

• f

lecture

• spark
• n
• versubscribed

clusters

Data locality

↳ a

:÷÷÷÷:*

. . .. ..

÷÷÷÷÷

.

• empted

→

cache

blown

away

long

lived

Layard

gamanteed

share

• ffers

perform

better

I

e

• up

C-

• ptimal

c-dis

bonparisons

Borg

NEXT STEPS

Next class: Scheduling Policy Further reading

• https://www.umbrant.com/2015/05/27/mesos-omega-borg-a-survey/
• https://queue.acm.org/detail.cfm?id=3173558

µ

"in:*

Athgnmentz

will

be

released

• Delay

scheduling

• ffer

wait for

• r

D

after

• ffer

Holik

← offer me