The Future of Virtualization: The Virtualization of the
Future
Larry Rudolph, VMware
© 2010 VMware Inc. All rights reserved
Caveats
 The opinions expressed in this talk are
• Entirely my own
• Do not come with any guarantees
• Subject to interpretation
• Based on years of experience having little to do with virtualization
• (I joined VMWare a year ago but I am still an academic )
2
2
Outline -- Disembodied Computation
 Beyond the OS
• More migration
 Beyond PC’s and Servers
• bigger, smaller
 Beyond Computers
• Don’t let the cat out of the bag
• No cat, no bag. No, really
3
3
OS Scheduler
 In a time-sharing
• want shortest jobs first
 Now think out of the box
Process Priority
• cpu intensive jobs run in background
I/O
Bound
Compute
Bound
Time Quantum Length
4
4
Location Scheduler
 In a rich environment
• Jobs should move near their
most critical resources
• I/O devices
• Storage
• Compute Devices
• Resource Allocation
• You can’t always get what you
want, but if you try sometime, you
get what you need.
5
5
Load Balancing
 Socialism, Capitalism, or
Environmentalism?
 Marginal Utility Curve per
job
• How does job’s performance
vary with its location
 Location Marginal Utility
Curve per Load
• increase number of jobs at a
location may slow all
 Solve by initial guess,
then iterate
6
6
Virtualization makes this possible
 Start job on local PC:
• If it needs more compute power,
• migrate it to server
• migrate it to supercomputer
• If it is always accessing a remote file system,
• migrate it to better SAN access
• migrate it to data center / cloud
• If it needs user interaction,
Assumptions:
* Virtualization nearly
everywhere
* Authorization to use
* etc.
• migrate to display
• If multiple high needs, split or thrash
7
7
Review: Benefits of Virtalization
 Isolation
• no connection between virtual machines
 Disembodiment
• The app & OS disconnected from HW
• CPU, MMU, Devices from each other
 Resource Sharing
• Time sharing done right
8
8
High Performance Computing
 HPC want every cycle of performance for their apps
• suspicious of virtual memory; VMs take too much overhead
 HPC applications optimally balance computation with communication
• virtualization that hides or modifies the ratio --> LPC (Low PC)
 HPC applications often require low-latency communication
• barrier synchronizations, collective ops such as reduce
• data-centers care about bandwidth and throughput, not latency
• e.g. infiniband not supported under current virtualization offerings
 HPC programmers use special purpose (expensive) supercomputers
9
9
9
HPC and the Cloud
 Despite negatives, there is good reasons to support HPC
 High-end HPC mostly government funded; just enough machines to meet
current needs
 Desire for elastic resource
• more computational scientists in academic & industry
 Massive data centers will have 100,000’s of cores and are efficiently
managed
 Co-locate small (windows-based) computations that feed into HPC task
• e.g. analyze biological slides on PC, then global analysis
 Fast decision making based on models -- e.g. finacial
10
10
1
Big Computation
Today’s Supercomputers have
• 100,000’s cores, hi bandwidth, lo latency
• Thrilling to use, but infrequent
Supercomputing for the masses
 languages, packages, tools all in place
Need seemless scaling from S -> M-> L
 Need to debug/refine only small part
 Create VMs, one per core, under-full simulation
• record neighbor communication (small fraction of time)
• debugging: full sim. island nodes, synthetic execute rest
• when effects changes neigh, change them to real full simulation.
11
1
Petascale
Ocean Model Numbers
1
Divide simulation
domain into subregions
(white lines on picture).
Each subregion covers
n (256 in this exercise)
processors.
Save messages at
subregion boundaries
To replay one year of simulation for region marked
Ņ1 Óabove. Would require
~300GB of stored data (assuming no user directed optimization) and
~256 processors.
This would enable an exact rerun of region
Ņ1 Ófor visualization or analysis of
~71TB of data.
For the whole simulation archive requirement is ~35TB, to allow any region to
be replayed.
12
Little computations
13
1
Larry’s Mission – Open Please
 Phone Stakeholders:
 Handset Manufacturer
 Carrier
 ISV
 Where does the owner fit it?
 When one buys a phone, one should be able to
install any OS on it, e.g. root password
 Clouds -- Vertically integrated, implementation
details hidden.
 Academic research relegated to toy systems?
14
1
Virtual-physical boundary
 Remember “Do-over’s?”
• I wish life was like that
• E.g. sent email but wish you didn’t?
 When do actions commit?
• when they are observable in the physical world
• someone reads their email, output to printer, launch missle
• checkpoint & rollback but across a system
• VM’s on distinct machines may cooperate to roll-back
15
1
Virtual Machines meet Virtual Reality
16
1
Mixed Reality
 2nd life on handheld or handheld in 2nd life
17
17
1
Blurring the boarders
 Seamless move between virtual and real world
 Human “world switch” is expensive
 Virtual world can keep track of all activities for later reference
 eliminates any need to sync between all different devices
 We can use whatever (computer-based) device is convenient
The virtualization of
the future
18
Confidential
1
19
19
1
Descargar

The Future of Virtualization: The Virtualization of the …