CSE598C Project:Dynamic virtual server placement

Yoojin Hong

Mobile VM servers (From the vMatrix)

Static mirroring vs. Mobile virtual machine servers

Mobile VM servers

Basic ideas VM servers can be hosted in any real machines across

networks VM servers can be instantiated on demand VM servers can move closer to end users

Advantages of mobile virtual machine servers Higher availability Better response time Absorbing flash crowds Network bandwidth savings Lower cost of ownership of server machines

Mobile VM servers

Two-tier architectures

Mobile VM servers

Disadvantages of mobile virtual machine servers Difficulty to virtualize large size of database server

Closer front-end VM server is beneficial only when # of packets via connection A is larger than that via connection B

Overhead of migrating OS and applications for VM servers

FRONT END BACK END

Connection A Connection B

Scenarios

The service is provided mainly by compute-intensive application server

Locations of majority of end users are changing Suppose users of CNN.com has an access pattern to vi

sit the site around 4-5p.m. Time difference between east coast and west coast

Problem description

Dynamic provisioning + Placement of web server replicas

Problems Determine when a new VM server needs to be added

(When) Change of locations of end users

Select a real machine, which is located in an optimal location from end users, to host the new VM server (How)

Location of real machines available Location of end users Location of back-end server relative to real machines Size of VM files relative to size of server requests Cost of bandwidth during different times of day Location of real machines to host VM servers currently

Algorithms

When When locations of end users are changing

# of users located where current server cannot guarantee a certain response time is increasing

Algorithms

How – Modified k-mean algorithm 1. Find k number of centroids at random 2. Assign each end user location to its closest centroid 3. Update the centroids as follows:

: # of end user locations assigned to cluster j : location of end user j : application-specific constant weight ( 0) B : location of back-end server

4. Repeat 2, 3 steps until the centroids are converged 5. Select RMs located closest to the centroids

1

i

ji N

BdZ

iZ

iN

jd

Experiments

Simulation Comparison …