Broadcast Disks: Data Management for Asymmetric Communication Environments

1
Broadcast DisksData Management for Asymmetric
Communication Environments

• 1997. 10. 24
• Sohn Young-Chul

2
Introduction

• Asymmetric communication environment
• physical bandwidth limitation (mobile)
• information flow of application(server/client)
• Broadcast disk
• Assumptions
• static access pattern
• data is read only
• no prefetching
• no request to server

3
How to Broadcast?

• Reduce expected delay
• flat / skewed / regular(multi-disk) broadcast
• advantage of multi-disk over skewed(random)
  broadcast
• low expected delay
• enable prefetching(regular)
• enable sleeping

4
Broadcast Algorithm

• some broadcast slot may be unused
• must be fine-tuned(increase the delay for pages
  in slower disks)

1. Order pages from hottest to coldest 2. divide
into multiple disk(group) 3. set relative
frequency for each disk 4. split each disk into a
number of smaller units
num_chunks(i) max_chunks/rel_freq(i)
5. run a program (p. 339)
5
Client Cache Management

• Role of client caching - caching hottest data?
• broadcast pages are not all equidistant from the
  client
• reasons making problem harder
• inaccurate access distribution, multiple client
• Cost based page replacement
• Store when access prob. is greater than freq. of
  broadcast
• prob. / freq. of broadcast(PIX)
• not practical because
• perfect knowledge of access probabilities,
  comparison of PIX for all cache resident pages at
  replacement time.

6
Experiment-modeling

• Page access model
• Zipf distribution
• prob. of accessing any page numbered i is
  proportional to (1/i)?
• Server execution model
• rel_freq(I)/rel_freq(N) (N-I)? 1
• ? 0 is flat broadcast

7
Result - no caching, 0 noise

• Increasing skew may hurt performance ( in case of
  D1 )
• more disk level does not necessarily ensure
  better performance ( in case of D2, D5)

8
Result - noise and no caching

• Multi-disk can have worse performance than flat
  disk

9
Result - caching and noise

• Caching policy P cache the hottest page (fig.
  8)
• noise sensitive
• noise increase caching of fastest disk pages
• hit rate remains same but cache miss become more
  expensive

10
Cost Based Replacement Algorithm

• PIX performs better with noise
• Lower cache hit ratio does not mean higher delay
  (fig. 11)

11
Implementing Cost Based Policies

• LRU
• LIX
• probability estimate
• pi c/(CurrentTime -
• ti)(1-c)pi
• lix pi / frequency of page i

12
Result - LIX vs. LRU
13
Conclusion

• Better broadcasting method than flat one.
• Cache can smooth dynamic access pattern of client
• Future work
• dynamic data(not read-only)
• prefetching