State-Slice: New Paradigm of Multi-query Optimization of Window-based Stream Queries

1
State-Slice New Paradigm of Multi-query
Optimization ofWindow-based Stream Queries

• Song Wang
• Elke Rundensteiner
• Database Systems Research Group
• Worcester Polytechnic Institute
• Worcester, MA, USA.

Samrat Ganguly Sudeept Bhatnagar NEC
Laboratories America Inc. Princeton, NJ, USA.
2
Computation Sharing for Stream Processing
Register Continuous Queries
Streaming Data
s
Streaming Result
w1
?
Agg
w2
s
Agg
s
w3
SPJA Query Network

• New Challenges
• In-memory processing of stateful operators
• Stateful operators with various window
  constraints

3
Window Constraints for Stateful Operators

• Time-based sliding window constraints
• Each tuple has a timestamp
• Only tuples within W timeframe can form an output

• Observations
• States in the operator dominate memory usage
• State size is proportional to the input rate and
  window length
• Join CPU cost is proportional to the state size

4
A Motivation Example
Q1 SELECT A. FROM Temperature A, Humidity
B WHERE A.LocationId B.LocationId WINDOW w1 min
Q2 SELECT A. FROM Temperature A, Humidity
B WHERE A.LocationId B.LocationId AND
A.ValuegtThreshold WINDOW w2 min
Let w1ltw2

• Observations
• State AW1 overlaps with state AW2
• State BW1 overlaps with state BW2
• Joined results of Q1 and Q2 overlap

5
Sharing with Selection Pull-up CDF02, HFA03
Q2
Q1
Router
sA
Ta-Tb
ltW1
all
R

Aw2
Bw2
B
A

• Selection pull up
• Using larger window (w2)

• CDF02 J. Chen, D. J. DeWitt, and J. F.
  Naughton. Design and evaluation of alternative
  selection placement strategies in optimizing
  continuous queries. In ICDE02.
• HFA03 M. A. Hammad, M. J. Franklin, W. G.
  Aref, and A. K. Elmagarmid. Scheduling for shared
  window joins over data streams. In VLDB03.

6
Sharing with Selection Pull-up CDF02, HFA03

• Pros
• Single Join Operator
• Cons
• Wasted Computation without Early Filtering
• Wasted State Memory without Early Filtering
• Per Output-Tuple Routing Cost

7
Stream Partition with Selection Pushdown KFH04
Q2
Q1
Router
all
ltW1
Ta-Tb
Union
U
R
A1
B1
A2
B2

Aw1
Bw1
Aw2
Bw2
1
2
lt
Split
S
gt
Threshold
B
A

• Split stream A by A.Value
• Route shared join results

• KFH04 S. Krishnamurthy, M. J. Franklin, J. M.
  Hellerstein, and G. Jacobson. The case for
  precision sharing. In VLDB04.

8
Stream Partition with Selection Pushdown KFH04

• Pros
• Selection pushdown no wasted Join Computation
• Cons
• Multiple Join Operators
• Duplicated State Memory in Multiple Join
  Operators
• Per Output-Tuple Routing Cost

9
State-Slice New Sharing Paradigm

• Key Ideas
• State-Slice Concept for Sliding Window Join
• Pipelined Chain of Join Slices
• Prospective Benefit
• Fine-grained Selection Push-down
• Pipelined Join Operators
• Avoiding Per-tuple Routing Cost

10
One-way State Sliced Window Join

• Iower bound of sliding window w1,w2
• B tuple only probes A tuples that are older at
  least W1, but at most W2, than itself

11
The Chain of One-way State-Sliced Joins
Joined-Result

• Split state memory into chain of joins
• No overlap of state memory in chain of joins

12
From One-way to Two-way Binary Join
Joined-Result
U
Union
female
A Tuple
State of Stream A 0, w1
State of Stream A w1, w2
male
Queue(s)
male
B Tuple
State of Stream B 0, w1
State of Stream B w1, w2
female
J2
J1

• Intuitively a combination of two one-way join
• Two references for each A or B tuples
• Male tuples are used to probe states
• Female tuples are inserted and cross-purged to
  respective states

13
State-Sliced Join Chain The Example

A1

• States of sliced joins in a chain are disjoint
  with each other
• ? Minimize State Memory Usage
• Selection can be pushed down into middle of join
  chain
• ? Avoid Unnecessary Resource Waste
• No routing step is needed
• ? Avoid Per Output-Tuple Routing Cost Completely

14
Summary State-Sliced Join Chain

• Pros
• Minimized Memory Usage
• Reduced Routing Cost
• No Need of Operator Synchronization in the Chain
• Cons
• Stream traffic between pipelined joins
• Purge cost

15
Sharing via Chains Memory-Optimal Chain

• No Selection

• With Selection

16
Mem-Optimal Chain? CPU-Optimal Chain?

• Overheads
• Too many operators may increase system context
  switch cost
• Too many sliced states increase purging cost

17
Merging Sliced Joins

• Tradeoff
• Gain from Merging
• Reduce number of Join operators
• Reduce extra purging cost
• Loss from Merging
• Introduce routing cost
• Increase memory usage due to selection pullup
• Cost Model for CPU Usage

18
CPU-Opt. Chain Search Space Solution
Legend Vi window start/end time Vi toVj one
slice window
v0
v1
v2
v5
v3
v4

• w0

• w2

• w1

• w3

• w5

• w4

Shortest path problem
19
Summary Mem-Opt. vs. CPU-Opt. Join Chain

• Mem-Optimal
• Minimized Memory Usage
• Higher System Overhead
• Higher Purging Cost
• CPU-Optimal
• Minimized CPU Usage
• More Memory Usage if Selection is Pulled Up to
  Merge Slices.

20
Experimental WPI Stream Engine CAPE
Software Demonstration VLDB04
21
Experiment Study 1 Memory Consumption
22
Experiment Study 2 Total Service Rate
23
Experiment Study 3 Mem-Opt. vs. CPU-Opt.
Window Distributions Used for 12 Queries.
Small-Large 12 Queries
Small-Large 24 Queries
24
Conclusion

• Pipelined state sliced join chain
• Mem-Optimal chain construction
• CPU-Optimal chain construction
• Implemented in CAPE
• Performance evaluation

25
Thank You!
Visit CAPE Homepagehttp//davis.wpi.edu/dsrg/CAP
E/index.html
Supported by
CRI grant CNS 05-51584