HotOS 2009
Friday, May 22nd, 2009Opening Remarks and Keynote Address
Keynote Address: The City Is Here for You to Use
Adam Greenfield, Head of Design Direction, Nokia
————————————
My background:
- only undergraduate degree
- army
- go for .com
- -> information architect in Tokyo
- works for different companies -> gain experiences
- the ordinary audience after presentation of a products get very upset and feel very shameful because they think they are stupid
Ubiquitous computing
Network urbanism
By the end of 2008 , more than half the world’s human population lived in cities: 2007 un revision of world urbanization prospects
by the end of 2012, network sensors will account for 20% of non-video internet traffic: garner top ten prediction of 2008
Transitions:
1- component/resource
2- constant/variable
adapt to dynamic situation
3; latent/explicit
4- browse/search
searching for any kind of objecting instead of browsing them to find the favourite one. e.g. a Mexican restaurant close to my area
5- held/shared
6- expiring/persistent
It is about the information which enters the systems and can get expired after a while
7- deferred/real-time
8- passive/interactive
9- wayfinding/wayshowing
how do you find your way around by maps? GPS shows your location in the map
10- object/service
e.g. cell phone object would not be useful unless it is connected to the network and deliver you the intended services
11- vehicle/mobility
instead of owing a particular vehicle we would have lots of heterogeneous mobility services which connect us to every given point
12- ownership/use
instead of owing the music and listening to it, we have the music on a server and whenever we want to listen to it, we are forced to listen to a commercial as well.
13- community/social network
14- consumer/constituent
derived by emergence of open source community
15- ?/?
Q: technical challenges to accelerate this process:
A: open standards, open services, open architects which people can learn them, understand them and connect them. e.g. open API
Q: may be things go towards opposite direction! (digital divide) because of the cost you might not use cell phone and loose connectivity
A: poor people are forced to expose themselves to the networks
but in future the price could be so low that everybody can use that
Q: it does not mean that the free public objects are useful
A: I think ultimately it is a design question, by careful design we can handle that
Q: the cities you are talking about are rich, poor cities do not have these facilities?
A: it is wrong, the question might be who is the producer and who would be the consumer. However, having the connection in future will happen.
Q: What is special about urban part?
A:
It’s Dead, Jim
Session Chair: George Candea, EPFL
Hierarchical File Systems Are Dead
Margo Seltzer and Nicholas Murphy, Harvard School of Engineering and Applied Sciences
———————————————-
categorizing objects can be done by size, color, or size.
We choose the one according to how we are going to use them.
Why force the users to pick only one attribute for organization.
Problem: hierarchical namespace
large volume -> databases to rescue
lighter than DBMS, easier than raw device?
Web: tried to impose a hierarchical name space
Why we do not use the full hierarchy in web name? auto-complete, bookmarks, search
data quantity + web = search
What do we need from FS really?
- backward compatibility
- separate naming from access
-unstructured data access
- direct access to data
q: things search does not do but DBMS does, like existence of the object
a: you can ask the existence question as a search query
Proposal: hierarchy must be the base, and hierarchy on top
q: database people are doing that fro 40 years
a; their underlying is tables
Why not just build indexed on top if Posix?
Posix is limiting in some ways
q: people do not use local FS usually? use facebook instead
a: yes, but it is not an issue, because it does not matter where the data lives
q: but about removing objects? We would have broken links?
a: I think that is OK
q: it makes people to get a copy of data?
a: it is a commercial question
q: hierarchy is useful for security on whole directory?
a: I think it will be better by our technique
q: we need hierarchical tags in Gmail
a: OK but do not impose it on the application, let is decide
q: but we need support on the FS
Heads in the Clouds
Session Chair: Garth Gibson, Carnegie Mellon University and Panasas, Inc.
On Availability of Intermediate Data in Cloud Computations
Steven Y. Ko, Imranul Hoque, Brian Cho, and Indranil Gupta, University of Illinois at Urbana-Champaign
——————————-
intermediate data as a first-class citizen for dataflow programming frameworks in clouds
mapReduce: massive-scale data processing, distributed and parallel execution on cluster
- 2-stage computation with all-total communication
Pig: build atop map reduce
If the intermediate data is not available, the reduce can not proceed.
Sometimes it needs cascaded re-execution which is expensive
Solution: replicate the intermediate data
but not enough: challenge is minimal interference between this and other transmissions
-> modify HDFS to reduce overhead with synchronous replication
Research direction: replication using spare bw
deadline-based replication
plication based on a cost model
Q: how serious is the problem?
Q: there is a research towards reducing the number of failures? Why not that?
Small Is Beautiful
Session Chair: Kimberly Keeton, Hewlett-Packard Labs
Mobility Changes Everything in Low-Power Wireless Sensornets
Prabal Dutta and David Culler, University of California, Berkeley
—————————–
Mobility vs. Power: we target low power, high speed
Mobility invalidates assumptions in many static designs:
network discovery
routing
link layer packet retransmission
We can turn the mobility bug into a feature:
we can easily detect motion and do things differently
mov metrics: chock, vibration, acceleration, free-fall …
is mov an event?
There are enough energies that we can use: cell phone transmission signal, TV transfer antenna
Q: option, added logic, harder to debug?
A: yes, trade-off
Augmented Smartphone Applications Through Clone Cloud Execution
Byung-Gon Chun and Petros Maniatis, Intel Research Berkeley
——————————
people expect desktop functionality from smart phones
Idea: let the smart phone host its expensive exotic applications by augmenting its execution seamlessly
We propose cloneCloud: smart phone would be the interface of application in cloud
q: power consumption for data transfer?
a: must be done when it is connected to electricity
q: what is the difference between your case and the things that we already have?
a: now we have to do the partitioning manually, choose that do we need to use the backbone or the smart phone is good enough for that
Migration without Virtualization
Michael A. Kozuch, Michael Kaminsky, and Michael P. Ryan, Intel Research Pittsburgh
——————————-
OpenCirrus: planet lab for data centers
tashi: cluster management software for big data
Migration:
process migration
vm migration
we propose OS migration: pros. access to hardware, performance - cons: not supported by current OS
Design space: same machine, different machines same hw, diff hw
reboot, hibernate, sleep, live migration
diff hw + live migration:
Problem; device driver state
solution: use abstract state + import + export
q: does it really work?
a:
Operating Systems Should Provide Transactions
Donald E. Porter and Emmett Witchel, The University of Texas at Austin
——————————
use transactions to express consistency requirements to OS
Why needed? Posix API is broken.
How do it? Extends Linux by +9000, 14000 modification
how to keep old and new data? Operate on private copy and update the data after commit
Simple api, hard to implement
Your computer is already a distributed system. Why isn’t your OS?
Andrew Baumann, Simon Peter, Adrian Schüpbach, Akhilesh Singhania, and Timothy Roscoe, ETH Zurich; Paul Barham and Rebecca Isaacs, Microsof
——————————
Access latency varies for different caches
We expect node heterogeneity
Dynamic changes
Extreme position: clean state design
no regard for adaptability
Message passing vs. Shared memory:
cache access is like blocking RPC
Trade-off between message passing vs. shared memory: shard memory is not scalable
We need to replicate state
Change programming model
Optimization: sharing as an optimization in multi-kernels
q: why not have vm on each core?
a: we have some similarities
q: why stick to the same machine?
a: optimization
Hardware
Session Chair: Petros Maniatis, Intel Research Berkeley
Operating System Support for NVM+DRAM Hybrid Main Memory
Jeffrey C. Mogul, HP Labs, Palo Alto; Eduardo Argollo, HP Labs, Barcelona, Spain; Mehul Shah, HP Labs, Palo Alto; Paolo Faraboschi, HP Labs, Barcelona, Spain
————————————-
dram is expensive, takes too much power
why not flash? slow read, very slow write, low lifetime
flam: flash + vram
use os-level knowledge to avoid wear-out