1. Research in Data Broadcasting
Michael Franklin
University of Maryland
November 1998
Joint Work with D. Aksoy, M. Altinel,
R. Bose, U. Cetintemel, J. Wang, and S. Zdonik
November 1998
© Michael J. Franklin

2. Data Dissemination Many emerging applications involve
large-scale data distribution.
stock and sports tickers
traffic information systems
software distribution
news and/or entertainment delivery
November 1998
© Michael J. Franklin

3. How well does it Scale? Election, Oscar, DeepBlue, etc. Result servers
use WWW server to disseminate results
Scalability?: Server crashes, intolerable delays,…
A Better Screen Saver (e.g., Pointcast)
“push-based” interface for customized news over the Internet.
Scalability?: large companies restrict usage due to heavy LAN traffic.
November 1998
© Michael J. Franklin

4. “Push” is a Potential Answer
Broadcast Media
Newspapers, Radio, TV, Junk Mail,…
Unicast
Mail, Telephone, ,…
Data Push
Teletext, BCIS, Datacycle, etc.
Publish/Subscribe
Webcasting, Internet Push
November 1998
© Michael J. Franklin

5. Push or Pull? - Webcasting (e.g., PointCast)
Web push technology is exploding --- even though there’s no such thing. Byte, 8/97
Client
Converter
User Interface
Server
PULL
PUSH
November 1998
© Michael J. Franklin

6. Push with Periodic Broadcast (e.g., Broadcast Disks)
Let’s look at another kind of push based system.
Teletext is a good example of this.
Broadcast disks is our own work.
Repetition creates a revolving disk.
Good for intermittent connection, limited memory, high turn-over, or huge client population.
November 1998
© Michael J. Franklin

7. Broadcast Disks Broadcast hottest data more frequently
Superimpose multiple “disks” on broadcast
A flexible, tunable, memory hierarchy
Client storage resources can be exploited to mitigate scheduling mismatch.
November 1998
© Michael J. Franklin

8. Push or Pull? - Broadcast Disks
CACHE
User Interface
Server
Client(s)
PUSH
PULL
November 1998
© Michael J. Franklin

9. What’s going on here?
1) Push vs. Pull is just one dimension along which to compare data delivery mechanisms.
- We focus on three.
2) Different mechanisms for data delivery can (and should) be used across different links.
- Enabled by Network Transparency.
November 1998
© Michael J. Franklin

10. Delivery Options Push Pull Aperiodic Periodic Unicast 1 to n request/
response
w/snoop
polling
polling
w\snoop
lists
publish/
subscribe
list
digests
Broad-
cast
disks
request/
response
All of these combinations make sense.
There are likely finer grained divisions in this hierarchy.
publish/
subscribe
November 1998
© Michael J. Franklin

11. Network Transparency Clients Brokers Sources
The type of a link matters only to nodes on each end
November 1998
© Michael J. Franklin

12. Using Network Transparency
An example: DB
Server
Proxy cache
A slightly more complex example.
Can vary dynamically
November 1998
© Michael J. Franklin

13. Large-Scale On Demand Broadcast
Server
downlink
Clients
uplink
November 1998
© Michael J. Franklin

14. Push vs. Pull Scalability
Pull-broadcast
Pull-unicast
Push-
broadcast
[few clients]
[many clients]
November 1998
© Michael J. Franklin

15. Previous Algorithms First Come First Served (FCFS)
Longest Wait First (LWF) [Dykeman et al.]
Most Requests First (MRF) “
MRF-Lowest (MRFL) “
Priority Index Policy (PIP) [Su and Tassiulas]
[Vaidya and Hameed]
November 1998
© Michael J. Franklin

16. Previous Algorithms - Performance
* Zipf distribution
* overhead ignored
LWF: wins but it is impractical.
November 1998
© Michael J. Franklin

17. Looking Deeper Hot Pages Cold Pages November 1998
© Michael J. Franklin

18. RxW - A Scalable, Tunable Scheduling Algorithm
Wait queue is FIFO
Request queue is sorted by # of reqs
Choose item with maximal RxW value
Pruning is effective (72% savings), but...
Approximate version can be as cheap as O(1).
November 1998
© Michael J. Franklin

19. Average Waiting Time (without scheduling overhead)
November 1998
© Michael J. Franklin

20. Scheduling Overhead increasing arrival rate increasing dbSize
November 1998
© Michael J. Franklin

21. Prototype Implementation
Server
IP-multicast
Downlink
(100 Mbps)
UDP
Uplink
(10 Mbps)
Clients
November 1998
© Michael J. Franklin

22. Responsiveness (Prototype)
Average Wait
November 1998
© Michael J. Franklin

23. Data Staging
Problem: Scheduling algorithms assume that all data are readily available for broadcast.
In reality, data may reside on disk, tertiary storage, or even at remote (e.g., WWW) sites.
We have integrated RxW with:
Server Cache Management
Disk Prefeteching
Opportunistic Scheduling (“postfetching”)
November 1998
© Michael J. Franklin

24. I. Caching - Love/Hate Hints
Goal: Reduce need to fetch data (i.e., misses).
Skewed access  many cold pages.
Least Recently Used handles cold pages poorly.
Approach: Love hot pages, hate cold ones.
A scheduled page is considered hot if both:
It is encountered on the R-list first.
It is within the top “hot range” R-list pages.
November 1998
© Michael J. Franklin

25. Caching Performance (RxW.90)
November 1998
© Michael J. Franklin

26. II. Prefetching
Prefetching is a common data staging technique (e.g., video-on-demand, WWW).
Since hot pages are cached --- prefetch the cold ones.
top “prefetch_window” pages in the W-list are prefetched.
November 1998
© Michael J. Franklin

27. III. Opportunistic Scheduling
More importantly -- keep the broadcast busy!
“Postfetch” missed page and send out another.
For data broadcast, this is best.
Square root rule.
Small latency penalty.
No guesswork.
November 1998
© Michael J. Franklin

28. Prefetch vs. Postfetch
November 1998
© Michael J. Franklin

29. On-Demand Broadcast Summary
For scalability, real solutions must take into account scheduling overhead and data staging.
RxW is 4 times faster than previous algorithms for same scheduling quality, and can be as cheap as O(1).
RxW provides hints for data staging.
Love/Hate caching plus “postfetching” obviate the need for speculative prefetching.
November 1998
© Michael J. Franklin

30. The DBIS Toolkit for Network Data Management
Clients
Brokers
Sources
November 1998
© Michael J. Franklin

31. Other Information Brokers
Information Broker Architecture
Other Information Brokers
Data Sources
Forwarded
Profiles
Data Items
Data Items
Data Source
Manager
Broker Manager
Data Sources
Decomposed Profiles /
Profile Updates
Pull Requests
Data Source Registration
Filtered Data HD
Cache
Catalog/Profile
Manager
Mapper
Profiles /
Pull Requests
Catalog Updates
IB Master
Broadcast
Manager
Scheduler
Client
Manager
Network Manager
Profiles /
Pull Requests
Acknowledgement
(Tune information)
Clients
Broadcast Medium
November 1998
© Michael J. Franklin

32. Map Dissemination Application
IBMaster, First IB, a Data Source
and a Client
Second IB and Client
November 1998
© Michael J. Franklin

33. DBIS Status Level 0 prototype constructed.
Supports Publish/Subscribe only.
Broadcast disks and On-demand scheduling to be added.
A key research issue is profiles.
How to express (is XML helpful?)?
How to manage and search 100’s of thousands?
How to automatically learn and maintain.
Clustering and categorization.
November 1998
© Michael J. Franklin

34. Summary Dissemination-based applications require new solutions.
Multiple types of data delivery can be combined easily due to network transparency.
We have developed scheduling and data staging techniques and are creating a toolkit .
Communication is important but a data management perspective is also essential.
Databases  Network Data Management.
November 1998
© Michael J. Franklin

35. An Analytical Treatment for RxW
Average waiting time:  N ui 
b-1  b i ui
W0 +
Wi + Wi
Wb =
i = b+1
i = 1
b-1  b i
1 - 
[ 1 - ] ui
i = 1
Average waiting time is bounded
inherent scalability in # of users
Square root ratio (optimal for push-based)
MRF - straight ratio, too much to hot pages
November 1998
© Michael J. Franklin

36. Worst Case Waiting Time
November 1998
© Michael J. Franklin

37. November 1998
© Michael J. Franklin

38. Extra graphs
November 1998
© Michael J. Franklin

39. November 1998
© Michael J. Franklin

40. November 1998
© Michael J. Franklin

41. November 1998
© Michael J. Franklin

42. Caching Policies Goal: keep hot pages in cache. November 1998
© Michael J. Franklin

43. November 1998
© Michael J. Franklin

44. November 1998
© Michael J. Franklin

45. Profiles [Cetintemel, Franklin, Giles 98]
Push requires models of user interests.
The accuracy of these models determines the perceived usefulness of the system.
The management of these profiles determines performance and scalability.
We have developed a novel, multi-modal technique for profile representation and learning.
November 1998
© Michael J. Franklin

46. Properties of On-Demand Broadcast
Broadcast is inherently scalable
Can do better than FIFO
Should approach optimal Push (assuming infinite server, no request overhead)
Need to balance average and worst case delay
Need algorithms that scale with the database size, bandwidth, and workload intensity.
November 1998
© Michael J. Franklin

47. DBIS Application Architecture
DS Code
DS Library
IB Executable
Client
Code
Library
Toolkit Parts
Application Developer Implementation
Data Source
Information
Broker
IB Master
Executable
Data
Catalog Info.
Other IB
Executables
November 1998
© Michael J. Franklin