A Novel Home Migration Protocol in Home-based DSM

1
A Novel Home Migration Protocol in Home-based DSM

• W.J. Fang C.L. Wang W.Z. Zhu F.C.M. Lau
• System Research Group
• Department of Computer Science
• The University of Hong Kong

2
Outline

• Background
• Problem Statement
• Our Approach
• Performance Evaluation
• Conclusion

3
Distributed Java Virtual Machine (DJVM)
import java.util. class worker extends
Thread private long n public worker(long
N) nN public void run() long sum0
for(long i0 iltn i) sumi
System.out.println(Nn Sum"sum) public
class test static final int N100 public
static void main(String args) worker w
new workerN Random r new Random() for
(int i0 iltN i) wi new
worker(r.nextLong()) for (int i0 iltN i)
wi.start() try for (int i0 iltN i)
wi.join() catch (Exception e)

• A distributed Java Virtual Machine (DJVM)
  consists of a group of extended JVMs running on a
  distributed environment to support true parallel
  execution of a multithreaded Java application.
• A DJVM provides all the JVM services, that are
  compliant with the Java language specification.
• Presents a Single System Image (SSI) of clusters
  to Java programs an illusion that the program
  is running on a single machine, yet more
  powerful.

(Single System Image)
Global Object Space
DJVM
Exec
Class
Thread
JVM
JVM
JVM
JVM
Java thread
HKU JESSICA2 project http//www.srg.cs.hku.hk/
4
Global Object Space

• Virtualizes a single object heap spanning on the
  whole cluster.
• A DSM service in the distributed JVM.
• Java Memory Model Resembles lazy release
  consistency.

5
LRC - Homeless Multiple-Writer Protocol

• False Sharing
• Two processes simultaneously write on the
  different parts of the same shared object.
• Ping-pong effect if using single-writer protocol.
  
• Twin and Diff
• Twin the copy before writes.
• Diff the difference between the current copy and
  the twin.
• Detect and batch the writes ever performed.
• Homeless
• The diff is saved and managed at each process.

6
Drawbacks of Homeless Protocol

• Communication overhead
• Fetch diff from each previous writer.
• Redundant operations
• Fetch and apply diffs at each faulting node.
• Memory consumption of twin and diff
• Requires a global garbage collection.
• Not scalable!

7
LRC - Home-based Protocol

• Each shared object has a home to which all writes
  (diffs) are propagated and from which all copies
  are derived.
• Demonstrated to be more scalable than the
  homeless protocol.
• Less communication, lower diff overhead, less
  memory consumption.
• Access asymmetry
• The access overhead at the home node is more
  lightweight than that at non-home nodes.

8
Outline

• Background
• Problem Statement
• Our Approach
• Performance Evaluation
• Conclusion

9
Home Assignment Problem
A home-based multiple-writer protocol
imple-menting LRC
Create twin
Create diff
Access symmetry ? Home assignment problem ?
Find the optimal home
Unlock
Apply diff
10
Related Work

• JiaJia
• Barrier manager detects single-writer pattern
  between two barriers.
• JUMP
• The process requiring the page may become the new
  home.
• Blind to access pattern.
• Jackals Lazy Flushing
• Home is fixed.
• A process may become the exclusive owner.
• Blind to access pattern.
• Some homeless DSMs (ADSM, Adaptive TreadMarks)
  can switch between single-writer protocol and
  multiple-writer protocol.

11
Home Migration

• A must in home-based protocols.
• Adapts to memory access pattern.
• Targets at the single-writer pattern.
• Home migration makes little difference in the
  multiple-writer situation as long as the home
  node is one of the writers.
• Negative effect
• Home migration notices
• We need a timely and precise home migration
  protocol.

12
Outline

• Background
• Problem Statement
• Our Approach
• Performance Evaluation
• Conclusion

13
Sensitivity of Home Migration Protocol
remote read (fault-in)
remote write (diff)
! Migrate home
! Migrate home
! Migrate home
14
Robustness of Home Migration Protocol
remote read
remote write
HM notice
P0
P1
P2
P3
A home migration protocol should be robust
enough to prevent unnecessary home migrations.
15
Sensitivity vs. Robustness

• At a first glance
• Sensitivity ?, robustness ?, and vice versa.
• Can we achieve both sensitivity and robustness?
• Single-writer pattern
• Lasting single-writer pattern
• Transient single-writer pattern
• A good home migration protocol should demonstrate
• sensitivity to the lasting single-writer pattern
• robustness against the transient single-writer
  pattern
• home migration is prohibited.

16
Our Previous Work

• A distributed JVM with GOS support
• A home-based, object-based, multiple-writer
  protocol
• Runtime detection of distributed-shared object
• Adaptive optimizations
• Object home migration (fixed threshold)
• Synchronized method migration
• Connectivity-based object pushing

Weijian Fang, Cho-Li Wang and Francis C.M. Lau,
On the Design of Global Object Space for
Efficient Multi-threading Java Computing on
Clusters, Parallel Computing Journal, Vol.29,
pp. 1563-1587, November-December 2003.
17
An Adaptive Threshold HM Protocol

• An adaptive per-object home migration threshold.
• decreases with increased likelihood that a
  particular object exhibits a lasting
  single-writer pattern.
• A lower threshold ? more likely a lasting
  single-writer patter ? quicker home migration.
• A higher threshold ? less likely a lasting
  single-writer patter (probably a transient
  single-writer pattern) ? more conservative home
  migration
• Continuously adjusted at runtime according to the
  feedback of previous home migration decisions for
  each object.

18
Runtime Feedback

• Consecutive remote writes
• Issued from the same remote node and not
  interleaved with the writes from either the home
  node or other remote nodes.
• To detect single-writer pattern.
• Exclusive home writes
• There is no remote write between an exclusive
  home write and an earlier home write.
• Positive feedback of previous home migrations.
• Redirected object requests
• Home migration notices.
• Negative feedback of previous home migrations.
• Home access coefficient
• The overhead ratio of one eliminated pair of
  object fault-in and diff propagation to one home
  migration notice.

19
Formalization

• Ci the number of consecutive remote writes
  since the (i - 1)th home migration.
• Ti the value of the adaptive home migration
  threshold since the (i - 1)th home migration.
• Tinit the initial threshold, which is equal to
  1.
• Ri the number of redirected object requests
  since the (i - 1)th home migration.
• Ei the number of exclusive home writes since
  the (i - 1)th home migration.
• ? the home access coefficient.
• ? the feedback coefficient. It is set to 1.

20
Formalization (contd.)

• Home migration is taken if
• Ci Ti
• where
• Ti max(Ti-1 ?(Ri - ?Ei)), Tinit
• where
• T0 Tinit 1

21
Outline

• Background
• Problem Statement
• Our Approach
• Performance Evaluation
• Conclusion

22
Effects of Home Migration
23
Adaptive Threshold vs. Fixed Threshold
24
Sensitivity and Robustness Analysis
25
Contributions

• This is the first home migration protocol taking
  home migration notice overhead into account.
• Our protocol is both sensitive to the lasting
  single-writer pattern and robust against the
  transient single-writer pattern.
• We formulate the protocol.
• The protocol uses a per-object adaptive migration
  threshold, which decreases with increased
  likelihood that the object presents the lasting
  single-writer pattern.
• The threshold is continuously adjusted according
  to the feedback of previous home migration
  decisions at runtime.
• We implement and evaluate this protocol in a
  distributed JVM.

26

• Questions?