Initial Results on Provisioning Variation in Cloud Services M. - - PowerPoint PPT Presentation

▶
initial results on provisioning variation in cloud
SMART_READER_LITE
LIVE PREVIEW

Initial Results on Provisioning Variation in Cloud Services M. - - PowerPoint PPT Presentation

Mar 24, 2024 259 likes •507 views

Initial Results on Provisioning Variation in Cloud Services M. Suhail Rehman Research Analyst Cloud Computing Lab Carnegie Mellon University in Qatar Collaborators: Prof. Majd F. Sakr, Jim Gargani Carnegie Mellon University Supported By: 1

slide-1
SLIDE 1

Initial Results on Provisioning Variation in Cloud Services

  • M. Suhail Rehman

Research Analyst Cloud Computing Lab Carnegie Mellon University in Qatar

Supported By:

Collaborators: Prof. Majd F. Sakr, Jim Gargani Carnegie Mellon University

1

slide-2
SLIDE 2

Cloud Computing / IaaS

2

What about other Application Domains? Scientific Applications High Performance Computing

slide-3
SLIDE 3

Cloud Computing / IaaS

3

What about other Application Domains? Scientific Applications High Performance Computing Application Performance on the Cloud

slide-4
SLIDE 4

What could affect performance?

Virtualization

4

Virtualized and Multiplexed Hardware Multitenancy

Abstraction

Identical Requests are not guaranteed to give you Identical Hardware Simplified, Abstracted Hardware

slide-5
SLIDE 5

Related Work

Many Studies on Virtualization and Application Performance Application Performance on EC2

  • Detailed studies on performance variance:
  • Service Oriented Applications : upto 4x

[Dejun 2009]

  • ~ 10- 25% Variation observed for

benchmarks on EC2 [Schad 2010]

5

slide-6
SLIDE 6

A closer look…

Physical Host L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

Physical Host

L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

Physical Host

L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

6

slide-7
SLIDE 7

A closer look…

Physical Host L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

Physical Host

L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

Physical Host

L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

7

VM VM VM VM VM VM VM VM

slide-8
SLIDE 8

A closer look…

Physical Host L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

Physical Host

L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

Physical Host

L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

8

VM VM VM VM VM VM VM VM

In Cloud Computing These Details are Abstracted from the User

slide-9
SLIDE 9

3 Potential Reasons for Performance Issues on the Cloud

9

Loads from other VMs on the same machine

1

slide-10
SLIDE 10

3 Potential Reasons for Performance Issues on the Cloud

10

Loads from other VMs on the same machine Variation in the physical resources being assigned to identical instances

1 2

slide-11
SLIDE 11

3 Potential Reasons for Performance Issues on the Cloud

11

Loads from other VMs on the same machine Variation in the physical resources being assigned to identical instances Configuration of the VM layout (where the VMs are placed during provisioning)

1 2

3

slide-12
SLIDE 12

12

Why does layout matter?

I want 4 VMs each with 1 vCPU, 1 GB RAM and 80 GB Disk

Client

Cloud Provider

Physical Hardware Virtual Machine

Resource Request

slide-13
SLIDE 13

Provisioning Variation

“variation due to ambiguity in the mapping of virtual resources to physical resources in a cloud computing environment”

Application

VMs from the Cloud Application Performance Variation

13 13

VM VM VM VM

slide-14
SLIDE 14

Experimental Methodology

  • Create Identical VM cluster instances in different physical

layouts manually

  • Evaluate the effect on performance for various applications.

Controlled Experimentation on a private cloud

V M V M V M V M

Client request for 4 VMs

Physical Host Physical Host Physical Host Physical Host

V M V M V M V M

Physical Host Physical Host

V M V M

Physical Host

V M V M V M V M V M V M

Layout 1: 4 VMs across 4 blades Layout 2: 4 VMs across 2 blades Layout 3: 4 VMs across 1 blade

Provisioned

  • n a private

cloud 14

slide-15
SLIDE 15

Testbed Configuration

Physical Host L3 RAM

Disk

4-core CPU

L2

L1 4-core CPU

L2

L1

CPU: 2 x Quad Xeon E5420 2.5 GHz w/ 12MB L2 Cache RAM: 8 GB ECC Front-Side Bus: 21.6 GB/sec Disk: 2 x 300 GB SAS Disk Bandwidth: 600 MB/sec Network Interface 2x Gigabit Interfaces to other blades IBM Bladecenter H with14 Blades

15

RHEL 5.1 Xen 3.0.3

Hadoop 0.20.1 RHEL 5.2

VM

Blade

slide-16
SLIDE 16

Benchmark Tests and Applications

  • CPU: SysTester
  • Memory: STREAM
  • Disk: Bonnie++
  • Network: Netperf

Systems Benchmarks

  • Hadoop Sort
  • Hadoop Wordcount
  • Hadoop TestDFSIO

Hadoop Workloads

Physical Host Physical Host Physical Host Physical Host

V M V M V M V M

Physical Host Physical Host

V M V M

Physical Host

V M V M V M V M V M V M

4 VMs across 4 blades 4 VMs across 1 blade 4 VMs across 2 blades

Executed on Synthetically Configured Infrastructure

16

slide-17
SLIDE 17

Results of Systems Benchmarks

17

Network

No Variation

RAM

25% drop in bandwidth for Layouts 2 and 3

Disk

60 – 80 % drop in bandwidth for Layouts 2 and 3

Physical Host Physical Host Physical Host Physical Host

V M V M V M V M

Physical Host Physical Host

V M V M

Physical Host

V M V M V M V M V M V M

Layout 1 Layout 3 Layout 2

CPU ~ 4x speedup for Layout 3

slide-18
SLIDE 18

256 512 1024 2048 4096 Layout 3 38.5 99.6 287.1 1339.7 6644 Layout 2 30.3 44.3 113.7 527.6 2400.7 Layout 1 29.8 38.3 66.4 362.9 1311.3

1 10 100 1000 10000

Time in Seconds (Log Scale) Size (MB)

Hadoop Sort

18

5x performance variation

Layout 3 Layout 2 Layout 1

Physical Host Physical Host Physical Host Physical Host

V M V M V M V M

Physical Host Physical Host

V M V M

Physical Host

V M V M V M V M V M V M

Layout 1 Layout 3 Layout 2

slide-19
SLIDE 19

DFSIO Benchmark

~ 5x performance variation

19

0.000 5.000 10.000 15.000 20.000 25.000 30.000 35.000

1x4 2x2 4x1 Throughput (mb/sec) Layout (VMxHosts) Read Write Layout 3 Layout 2 Layout 1

Physical Host Physical Host Physical Host Physical Host

V M V M V M V M

Physical Host Physical Host

V M V M

Physical Host

V M V M V M V M V M V M

Layout 1 Layout 3 Layout 2

slide-20
SLIDE 20

Analysis

  • Upto 5x performance drop in both
  • Disk contention is the reason

Correlation between Sort and DFSIO Benchmark

  • When all VMs are on one blade, they

compete for disk I/O bandwidth Hadoop designed to leverage parallel I/O

20

slide-21
SLIDE 21

Hadoop Wordcount

1 2 4 8 4Vx1B 182 339.9 660.2 1633.8 4Vx2B 167.2 321.9 620.4 1414.4 4Vx4B 167.1 317.2 631.5 1356.8 200 400 600 800 1000 1200 1400 1600 1800 Runtime (Seconds) Input Size (GB)

~20%

21

Layout 3 Layout 2 Layout 1

Physical Host Physical Host Physical Host Physical Host

V M V M V M V M

Physical Host Physical Host

V M V M

Physical Host

V M V M V M V M V M V M

Layout 1 Layout 3 Layout 2

slide-22
SLIDE 22

Conclusions

22

  • VM placement on same resource
  • Higher bandwidth for inter-VM communication
  • Constraints Memory and Disk

Tradeoffs

  • It’s impact on performance varies across

application domains

  • Up to 5x performance variation for I/O-bound

Provisioning Variation

slide-23
SLIDE 23

Future Studies and Directions

  • Different classes of scientific applications (CPU, Memory, I/O

Bound)

More Studies on other Applications

  • To inform provisioning to meet QoS

Application Profiling on the Cloud

  • Dynamic application adaptation to variations in cloud resources

Resource-aware Applications

We have only scratched the surface!

23

slide-24
SLIDE 24

Join Us!

Postdoctoral Positions Carnegie Mellon Qatar http://qatar.cmu.edu/~msakr/postdoc

24

Contact Prof. Majd Sakr or Email Me: msakr@qatar.cmu.edu suhailr@qatar.cmu.edu