Locality Optimizations in OceanStore

1
Locality Optimizations in OceanStore
An introduction to introspective techniques for
exploiting locality in wide area storage
utilities.

• Patrick R. Eaton
• Dennis Geels

2
Agenda

• OceanStore Review
• Problem Overview
• Previous Work
• Proposed Solution
• Prefetching Algorithm
• Preliminary Results
• Future Work

3
OceanStore Review

• Properties of OceanStore relevant to
  introspective locality optimizations
• implemented in the extremely wide area
• has many places to put any single piece of data
• cannot rely on users to make relationships among
  data explicit
• depends on effective locality optimizations for
  improved performance
• No possible way to solve exactly

4
Problem Overview

• Passively observe data accesses
• data shared among multiple users
• single users accessing the network from different
  physical locations
• data is replicated across the network
• Optimize the location of data to provide quicker
  access to users
• cluster semantically related data
• replicate data to move it closer to consumers
• migrate primary replicas toward the source of
  updates

5
Measurable Attributes

• File Temperature
• A measure that indicates the frequency of access
  to the file
• A hot file is frequently accessed
• Semantic Distance (Kuenning)
• Any measure that can quantify relationships
  between files on the range 0,?)
• Local distance relates one instance of a file
  access to another
• Reference distance is an aggregate measure that
  summarizes all local distances for a pair of
  files
• Typical measures use access order or timing
  information

6
Prefetching Techniques

• Automatic Prefetching (Griffoen and Appleton)
• construct a probability graph that records
  accesses which follow within a lookahead period
• predict a prefetch when the chance of an access
  is above a tunable parameter
• Context Modeling (Kroeger and Long)
• record in a trie all access sequences which have
  been observed
• maintain pointers to all nodes which represent
  current contexts
• predict a prefetch when the chance of an access
  to a child of a current context is above a
  probability threshold

7
Our Approach

• Exploit the ideas of semantic distance to compute
  relationships among data
• Cluster data based on the observed relationships
• Store a summary of these relationships with the
  data
• Migrate (prefetch) files based on familiar
  patterns in the access stream
• recognize higher order correlations as in context
  modeling
• tolerate noise in the access stream

8
Motivation for Prefetching Algorithm
A
Y
Many patterns can be predicted only by
observation of higher-order correlation--combining
several pieces of past history.
K
B
Z
A
Other patterns can only be detected through
identification and filtering of noise.
B
C
9
General Prefetching Algorithm

• Update
• Record the most recent file accesses in the file
  history buffer (FHB)
• Each time a new file S is accessed, extract all
  triples of the form (FHB(i), FHB(j)) ?S from the
  FHB and update in the second-order distance table
• Predict
• Each time a new file S is accessed, examine the
  distance table entries of (FHB(i), S)
• Prefetch files that are predicted with confidence
  above a certain threshold
• Problems
• O(k2) work to update distance table
• Noise infects distance table

10
Optimizations to the Prefetching Algorithm

• First-order distance table
• Records files that are close, as measured by
  semantic distance
• Allows reverse lookup
• Use first-order distance tables to filter out
  irrelevant file relationships
• Update only relevant entries in the second-order
  distance table
• Search for predictions based on only relevant
  access pairs

Indicative FHBs
11
Prefetching Algorithm Example

• Update
• Extract relevant triples by intersecting the FHB
  with the results from the reverse lookup in
  first-order tables

• Predict
• Extract relevant doubles by intersecting the FHB
  with the results from the reverse lookup in the
  first-order tables
• Prefetch if the second-order table predicts a
  future access with sufficient confidence

12
Preliminary Results (Local System)
13
Future Work

• Retarget the simulations to model OceanStore
• Continue to refine the prefetching algorithm
• Examine the potential of higher order prefetching
• Combine prefetching and clustering
• Look for opportunities to test the ideas on
  different workloads