PP16 Lec26 CC3&Embedding1

download PP16 Lec26 CC3&Embedding1

of 12

Transcript of PP16 Lec26 CC3&Embedding1

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    1/12

    Parallel

    ProcessingLec26

    Dr M Shamim Baig

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    2/12

    Multiprocessors CacheCoherence protocols

    Snoopy based protocols: Broadcast oriented e.g

    shared bus Directory based protocols

    Selective communication

    oriented e.g MIN

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    3/12

    Directory Based multiprocessor ccoherence protocol

    snoopy CC-protocol isinherently

    limited to busbased

    as each coherence operation is broadcasted to all pr

    When IN is not a bus e.gMIN!" broadcast is no

    # requiresadditional hard$are # time delay to

    as here point to point comm is %avored.

    Why not send coherence re&uests to onlythose p

    that need to be noti%ied'

    It is done usinga directory"which maintains presenc

    %or each data item cache line! along $ith its glo

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    4/12

    (rchitecture o% typical directory based systems:a a centrali)ed director b a distributed director

    Directory Based CC-System

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    5/12

    Performance ofDirectory Based CC-Schemes

    *he need %or a broadcast media is replace

    the directory. *he underlying net$or+ must be able to ca

    all the coherence re&uests.

    *he additional bits to store the directory m

    add signi%icant overhead. ,uanti%y memory re&uired %or the t$o Dire

    based CC-schemes Normal vs %%icient!.

    *he directory is a point o% contention" there

    distributed directory schemes may be used

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    6/12

    Normal Directory based CC-schem Cache-bloc! "ntries structure

    States bits

    States bits

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    7/12

    ma# simultaneous cachingbounded

    (r log2n + 2)

    "$cient Directory based CC-schem Cache-bloc! "ntries structure

    i% r n

    States bits

    / M $here M 001

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    8/12

    "#ample %&ef class notes'Memory ()erhead for both

    Directory based CC-schemes

    NormalvsEfcient

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    9/12

    *opology "mbedding

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    10/12

    +hy *opologies Mapping, 2%ten" $e need to embed a +no$n

    communication pattern into a giveninterconnection topology.

    We may have an algorithm designed

    %or one net$or+" $hich $e are portinto another topology.

    For these reasons, it is useful to understand mapp

    between graphs

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    11/12

    raphs. topology Mapping Me

    When mapping a graph G(V,E)intoG(V,E)the

    metrics are important:-

    Dilation: Ma3imum number o% consecutive! lin

    that any! one edge in Eis mapped onto is

    dilation o% mapping. (!istance delay effect) Congestion:Ma3imum number o% edges (

    consecuti"e)in that are mapped onto any on

    in E is called congestion o% mapping. (#omm

  • 7/26/2019 PP16 Lec26 CC3&Embedding1

    12/12

    *opologies "mbedding *ypes Embedding Sparser networks in Denser n

    4inear (rray into a Mesh

    4inear (rray into a 5ypercube

    Mesh into a 5ypercube

    Simultaneous Multiple topology embeddin

    Embedding Denser networks in Sparser n Mesh into a 4inear (rray

    5ypercube in 4inear (rray 5ypercube into a Mesh