Placement Algorithms

1
Exploiting Sharing for Data Center Consolidation
Timothy Wood, Jim Cipar, Gabriel Tarasuk-Levin,
Peter Desnoyers, Emery Berger, Mark Corner,
Prashant Shenoy University of Massachusetts,
Amherst
Motivation and Challenges
Using the potential for memory sharing as a guide
for placing VMs can lead to substantial memory
savings. Memory requirements and the potential
for sharing fluctuate over time, thus the system
must monitor memory utilization to prevent
hotspots.
Memory is an expensive resource and can be the
limiting factor when consolidating virtual
machines with low CPU utilization. ESX Server
supports page sharing allowing virtual machines
to reduce memory consumption by sharing identical
pages.
Sharing Memory
If two VMs have an identical page in memory,
only store a single copy until one makes a
write. Matches are found by comparing hashes
generated for each page in a VMs
memory. Currently, ESX only monitors sharing
within a single host.
A VMware experiment reports memory savings of
over 30 for a group of Windows servers.
Table from Memory Resource Management in VMware
ESX Server, Carl A. Waldspurger, OSDI 2002
Finding Similar Virtual Machines
We must efficiently calculate the sharing
potential between VMs across large data
centers. Brute force comparison of page hashes
is costly in both computation and memory.
We propose an efficient Bloom filter based
fingerprinting technique. Bloom filters still
maintain a high degree of accuracy, and allow for
a tradeoff between storage requirements and
prediction accuracy.
Bloom Filter Accuracy
Sharing Estimation Time
Relative Error
Time (sec)
0 25 50 75 100 125
150Number of VMs
10 25 50 75
100 Sharing Rate
Placement Algorithms
VM
VM
VM
VM
VM
VM
Using sharing potential can help optimize
placement of virtual machines. Sharing reduces
memory requirements, increasing consolidation
possibilities. Additional memory tracing
techniques can help detect and prevent memory
hotspots.
Host
Host
Host
Taking advantage of memory sharing allows more
virtual machines to run on a smaller number of
hosts
VM
VM
VM
VM
VM
VM
Host
Host