The Utility of Exploiting Idle Memory for DataIntensive Computations

1
The Utility of Exploiting Idle Memory for
Data-Intensive Computations

• Anurag Acharya and Sanjeev Setia
• Presented by John Oleszkiewicz
• CSE 807 4/22/03

2
Problem Statement

• We would like to use the idle memory of a network
  of workstations to allow data-intensive
  applications to run more quickly

3
Questions

• What fraction of memory is idle at any given
  time on a NOW?
• What fraction of individual host memory is idle
  at any given time?
• How long will memory typically stay idle?
• What is the benefit of harvesting idle memory to
  the owner of the program?
• Will harvesting memory have an adverse impact on
  users of the workstations?

4
Workstation Environments

• Alpha
• 10 stations, 4-processor DEC, batch processing
• Umdres
• 14 stations, faculty and grad students
• Gmu
• 32 stations, faculty and grad students
• Comsearch
• 37 workstations, engineers and software developers

5
Traces

• Two traces collected for each workstation
• User activity (e.g. Mouse movement)
• Memory load information
• Each trace has a resolution of 90 seconds

6
Workloads

• OLAP 5 queries to department store database
• 1 million customers
• 100,000 products
• 10,000 employees
• 200 departments
• 100 million transactions
• Can't fit in available memory of NOW

7
Workloads

• DMINE application that extracts association
  rules from retail databases
• 50 million transactions
• Each transaction typically contains 10 items
• Smaller than OLAP, but still can't fit in idle
  memory of workstation.

8
Workloads

• LU
• Decomposition of a 8192x8192 matrix
• Can fit in idle memory of NOW

9
Metrics

• Memory Availability
• Benefits
• User Perception

10
Memory Availability

• Aggregate availability
• total memory in pool that is available
• Fractional availability
• fraction of workstations that have at least x
  memory available
• Idle
• No activity and memory load less than 30 for 5
  min
• Idle Interval
• How long a chunk of memory remains idle

11
Benefits

• Execution time
• reduction in execution time by using idle
  memory in NOW as cache
• Equivalent Memory
• Amount of memory that has to be added to the
  local workstation to get the equivalent speedup

12
User Perception

• Memory context size
• How much memory must be reclaimed when the
  workstation becomes active
• Directly impacts user-perceived slowdown

13
Parameters

• Recruitment policy
• Network Bandwidth
• Memory vacate time
• How fast should harvested memory be relinquished?

14
Results
15
Metrics

• Aggregate Availability
• 70-85 of total memory is available on average
• 60 is (almost) always available
• Fractional Availability
• 90 have 50 memory available
• Idle Interval
• Median 18-30 minutes

16
What does it mean?

• Parameter recruitment policy should be set to
  50
• Very large memory chunks are likely to survive at
  least 15 minutes
• Apps that access datasets several times in 15-20
  minutes will benefit

17
Workload Results

• LU
• Performance benefits on some clusters
• Equivalent Memory 379 MB
• EQ limited by network overhead
• DMINE
• Performance benefit reflect amount of idle memory
  available
• EQ highest at 1 GB
• EQ limited by idle memory

18
Workload Results (con't)

• OLAP
• Performance increase similar to DMINE, but
  benefit was much lower
• EQ benefits lower
• Bottleneck average I/O requirements much
  larger, bigger communication overhead.

19
Metrics User Costs

• In most cases, no delays
• Memory context size typically between 0-1.5 MB
• Harvesting at most 50 of memory has very little
  impact on users.
• What about increasing the recruitment policy to
  over 50?
• Beneficial to apps with footprint greater than
  idle memory available
• Cost user-visible delay.

20
Other Results

• Increasing network bandwidth increased equivalent
  memory
• Memory vacate times
• Numbers other than 0 never resulted in
  improvement of more than 10
• Small benefit probably not worth slowing down
  workstation owner.

21
Conclusions

• 30 of memory on a typical NOW can be had with no
  adverse impact on users
• If this memory is used as an intermediate cache,
  execution time can be reduced by 26
• Restricting maximum recruited memory to 50
  ensures no delays during reclamation.

22
(No Transcript)