Docstoc

CSS Parallel Distributed Computing

Document Sample
CSS Parallel Distributed Computing Powered By Docstoc
					Ivy: Memory Coherence in Shared
     Virtual Memory Systems
      Presenter: Munehiro Fukuda




                 CSS434 DSM        1
Outline
   Architectural Overview
   Review of Write Invalidation
   Ownership Implementation
       Fixed Ownership
       (Improved) Centralized Manger
       Fixed Distributed Manager
       Broadcast Distributed Manager
       Dynamic Distributed Manager
   Contribution and Drawback
                       CSS434 DSM       2
Ivy Architecture
  Appolo      Appolo                 Appolo
   Node 1     Node 2                  Node n
             Memory                  Memory
  Memory



   Mapping   Mapping                  Mapping
   Manager   Manager                  Manager




                Shared Memory

                        Token Ring
                       CSS434 DSM               3
      Write Invalidation
                            Client wants to write:

                                    new copy P3            owner = P3
Problem:                                                  Copyset = {P3}
                                                        Copyset = {P0, P1, P2}
1. How to locate owner
2. Where to store copyset
                                     2. Replicate block

                                                           3. Invalidate block
                   3. Invalidate block
                                               1. Request block


       a copy of                                                          a copy of
                                         block
         block                                                              block
                                                    owner = P1
         P0                               P1      Copyset = {P0, P2}
                                                                                 P2

                                         CSS434 DSM                                   4
Page Ownership Implementation

They focused on write invalidation.




                         CSS434 DSM   5
   Fixed Ownership Strategy
   (They didn’t event consider about it.)

                         Owner == I can write it.


  P0              P1               P2               P3     P4




   W      R        W      W         W                W      W
         Okay             No!


Page 0          Page 1          Page 2          Page 3   Page 4


                                 CSS434 DSM                       6
Centralized Manager

              4. Page Transfer
   P1                                     P2                         P3


                                     Page 0
        1. Query for Owner                          2. Invalidate
                        3. Write Request

 5. Confirmation
                               P0         Page#   Owner   Copy set
                                          0       P2
                                                   P1     P3
                                               Locked from 1 to 5
                             CSS434 DSM                                   7
           Improved Centralized Manager
                                              Copy set
Copy set                                      P3
P3
                        3. Page Transfer
             P1                                     P2                 P3


                                               Page 0
                                                            4. Invalidate
                  1. Query for Owner
                                  2. Write Request


1’ New requests forwarded to P1          P0         Page#   Owner
                                                    0       P2
                                                             P1



                                       CSS434 DSM                           8
             Fixed Distributed Manager

    P0                     P1             P2              P3                  P4


                                   Hot spot
Page#   Owner      Page#   Owner     Page#    Owner   Page#   Owner   Page#    Owner
0       P2         2       P0        4        P4      6       P1      8        P4
1       P3         3       P3        5        P1      7       P2      9        P0



             It works only if memory accesses are uniformly distributed.
             But…



                                         CSS434 DSM                                9
     Broadcast Distributed Manager
        Page & copy set


P0          P1              P2                P3               P4



                          Page 0           Request




             Request

                                 When I can find if I got an ownership?
                                  Should I repeat sending a request?
                                    Doesn’t it saturate network?
                           CSS434 DSM                            10
         Dynamic Distributed Manager
         (Read fault: A Naïve Algorithm)

                       request        request            request
  P0              P1             P2                P3              P4


                                                     owner
                                       owner
owner         owner


Page 0                 Problem: a long link to a real owner




                                 CSS434 DSM                             11
     Dynamic Distributed Manager
     (Read fault: A Correct Algorithm)



P0            P1        P2            P3           P4


                                         owner
                              owner
          owner                                  owner
                                      owner

                                                 Page 0




                        CSS434 DSM                      12
     Dynamic Distributed Manager
     (Write Fault)



P0            P1     P2            P3         P4


                                    owner
                           owner
          owner
            owner                           owner


            Page 0                          Page 0




                     CSS434 DSM                    13
Further Improvement
     Periodically broadcasting the current owner’s
      location to all processes
          Simulation for 1024 processors



#page faults before broadcast        #messages taken to the owner
256                                  2.34
1024                                 3.64




                                CSS434 DSM                          14
Contributions
   Full examination of page ownership
    implementations
   Performance evaluation
       Partial differential equations: excellent
       Merge sort: good but bound to logN
       Dot product: too small granularity
       Matrix multiplication: excellent
   Memory usage per CPU decreased as
    increasing #CPUs participated in computation

                          CSS434 DSM                15
Drawbacks
   Little information about actual
    implementation: How much did they
    change OS?
   No performance evaluation using
    pathetic applications: Computation
    granularity versus communication
   How much did #forwarded messages
    occupy the entire overhead?
                 CSS434 DSM              16
Questions?




             CSS434 DSM   17

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:4
posted:10/9/2011
language:English
pages:17