ProcessingInMemory Technology for Knowledge Discovery Algorithms

1
Processing-In-Memory Technology for Knowledge
Discovery Algorithms

• Jafar Adibi et.al. USC/ISI
• Presented by Chen Qian

2
Outline

• The goal of this paper
• Link Discovery Algorithm
• Processing in Memory (PIM)
• Graph clustering simulations
• LD on PIM Performance Analysis
• Conclusion

3
Goal

• To gain insight into whether processing-in-memory
  (PIM) technology can be used to accelerate the
  performance of link discovery algorithms

4
Link Discovery

• LD is a new challenge in data mining whose
  primary concerns are to identify strong links and
  discover hidden relationships among entities and
  organizations.
• Data-intensive and highly parallel.

5
PIM

• PIM chips that integrate processor logic into
  memory devices offer a new opportunity for
  bridging the growing gap between processor and
  memory speeds.
• especially for applications with high
  memory-bandwidth requirements.

6
PIM and LD

• Parallel computing power has the potential of
  providing dramatic computing speedups.
• This paper evaluate the LD algo to a PIM
  workstation-class architecture.

7
Link Discovery

• A major problem in LD is the discovery of hidden
  organizational structure such as groups and their
  members.
• Visible relationship (groups) class
• Invisible terrorist groups, drug cartel

8
Two functions

• Group detection can be further broken down into
• (1) discovering hidden members of known groups
  (or group extension)
• and (2) identifying completely unknown groups.

9
Algo in this paper

• The KOJAK Group Finder.

10
KOJAK Group Finder

• 1. A seed generator outputs a set of seed groups
  using deductive and abductive reasoning.

11
KOJAK Group Finder

• 2. A mutual information model finds likely new
  candidates for each group, producing extended
  groups.

12
KOJAK Group Finder

• 3. The mutual information model is used to rank
  these likely members by how strongly connected
  they are to the seed members.

13
KOJAK Group Finder

• 4. the ranked extended group is pruned using a
  threshold to produce the final output.

14
Transfer space to graph

• Each node represents an entity and each link
  represents the set of actions (email, phone call
  etc).

15
Transfer space to graph

• Graph Partitioning Problem

16
Link Discovery

• LD algo are data-intensive and highly parallel.

17
PIM
18
PIM

• However the limitation of memory bandwidth will
  also cause slow processing.
• Processor speed increases at 60/year
• Memory speed increases at 7/year.
• Increasing the CPU speed is not enough.

19
PIM

• PIM chips that integrate processor logic into
  memory devices offer a new opportunity for
  bridging the growing gap between processor and
  memory speeds.
• especially for applications with high
  memory-bandwidth requirements.

20
DIVA Key ideas

• First smart memory PIM that is
• Capable of executing independent threads of
  control
• Designed to support in-memory virtual addressing.
  
• Target applications Image processing and
  multimedia (streaming)

21
DIVA system architecture

• PIMs are separated by host memory and host
  processor.

22
chip architecture

• Two components. Two interconnects.

23
chip architecture

• Fast processing can be performed inside the
  memory (PIM units) without loading into host
  processor.
• Drawback
• the ability of process logic in PIM is limited.
• Special design for the process logic

24
DIVA PIM

• Can provide dramatic computing speedups for
  paralleled algorithms.

25
Parallel Graph Clustering Algo

• To take the advantage of PIM machines.

26
Sum.

• The data and computation are partitioned among
  PIM nodes as follows
• Each PIM keeps a fraction of the PairTable, that
  is, a subset of the pairs of nodes in the graph,
  where a pair is a set of two nodes with at least
  one link between them.
• A PairTable is partitioned so that, for a given
  node, all pairs containing that node are kept on
  the same PIM.

27
Sum.

• Hence the whole graph is represented in a link
  table of twice the number of links times the
  number of all link types plus 2 (for two nodes).
• This duplication of information avoids
  unnecessary cross communication among PIMs.

28
Sum.

• At each iteration each PIM computes the node,
  among the nodes in its subset, with highest
  InOutRatio.
• After that, all PIMs communicate to find the node
  with highest InOutRatio across all PIMs.
• Finally, at the end of each iteration, each PIM
  adds the new global best node to its local copy
  of the current group.
• Algo iterates until a desired no. of new nodes is
  added or there are no more to be added.

29
Benchmarks

• The measurements on two common bandwidth
  benchmarks tests out the hypothesis that PIMs
  really do offer a bandwidth advantage.
• StreamAdd measures the performance of a stream
  of floating point additions and updates.
• RandomAccess is designed to stress the memory
  system by performing updates to randomly selected
  entries of a large table.

30
Benchmarks

• PIMs were able to deliver comparable bandwidth to
  the host-only machine with just a single PIM, and
  8X more bandwidth with 8 PIMs

31
Simulation

• Simulation for Graph clustering algo.
• Incresing the number of PIMs reduces the cost of
  overall computation.
• However, the cost of communication also increased

32
Simulation

• Load imbalance problem exists.
• The amount of work at each PIM varies
  significantly. But the imbalance reduces as the
  of seed members increases.

33
Experiment

• Compared with Itanium-2

34
Experiment

• For Mutual Information
• Note that the Itanium-2 clock is operating at 900
  MHz while the PIM clock rate is 140MHz. The of
  cycles is used for comparison. The PIM code is
  executing 18 fewer cycles than the Itanium-2.

35
Experiment

• For Graph Clustering
• Significant improvement. Far more promising than
  that of mutual information

36
Experiment

• For Multi-PIM
• For both kernels, our PIMs were able to deliver
  comparable bandwidth to the Itanium-2 with just a
  single PIM, and 8X more bandwidth with 8 PIMs

37
Summary

• What is the most important
• LD ?
• PIM ?
• How to perform graph clustering in parallel and
  manage the communication overhead.
• There still exists more potential.

38

• Questions?
• Thank you !