Transcription

1

2

3

4

5

6

7 The Algorithm (DSC): Mark all nodes as unscheduled WHILE there are unscheduled nodes DO In descending order of priority, sort list of free nodes In descending order of priority, sort list of partial free nodes Let n x be the free task with highest priority. Let n y be the partial free task with highest priority. IF(priority(n x )!priority(n y )) THEN IF ( CT) THEN accept zeroing for edge (c i, n x ) where c i gives min(st(c i, n x )) ELSE schedule n x on new cluster ENDIF ELSE IF ( CT & CT) THEN accept zeroing for edge (c i, n y ) where c i gives min(st(c i, n y )) ELSE schedule n x on new cluster ENDIF ENDIF Mark n x as scheduled Recalculate priorities & find new free & partial free nodes ENDWHILE DEFINE CT: /*Guarantees that parallel time is not increased if node nxis assigned to a Predecessor */ For all clusters c i that are parents of n x (min_start_time = MAXINT) { IF ( min_start_time > ST( c i, n x )) THEN min_start_time = ST(c i, n x ) } IF ( startbound(n x ) < min_start_time ) THEN return TRUE ELSE return FALSE ENDIF ENDDEF DEFINE CT: /*Guarantees that the starttime of partial free nodes is never increased */ For all scheduled clusters c i that are parents of n y { IF (startbound(n y ) < ST(c i, n y )) THEN return FALSE ENDIF } return TRUE ENDDEF

8 A B C C C unscheduled = { } scheduled = " PT = 0 C free= {} partial free = " unscheduled = { } scheduled = {} PT = 0 C free = { }} partial free = " unscheduled = { } scheduled = { } PT = C free = { } partial free = " D E F C C unscheduled = { } scheduled = { } PT = free = {} partial free = {} unscheduled = {} scheduled = { } PT = free = {} partial free = " unscheduled = " scheduled = { } PT = 0 free= " partial freel = "

9 The Algorithm (MCP): List_of_PEs = ". Perform ALAP binding and assign the resulting ALAP time T L to each node in the graph.. For each node n i create a list l(n i ) which consists of T L s of n i and all it s descendants, sort l(n i ) in decreasing order of T L.. Create L by concatenating the T L values for each l(n i ) and sort in decreasing order.. Schedule head(l) to a Processing Element (PE ) and remove head(l) List_of_PEs # PE (head(l)). While L is not empty Compute min(start_time) when head(l) is placed on PE i in List_of_PEs. If min(start_time exist_cluster ) > start time of head(l) when placed on a new PE then add new PE to List_of_PEs and place head(l) on new PE else place head(l) on PE that gives min(starttime)

10 A TL(n) = 0 TL(n) = 76 TL(n) = TL(n) = TL(n) = 00 l(n) = (n, 00), (n, 76), (n, ), (n, ), (n, 0) l(n) = (n, 00), (n, 76) l(n) = (n, 00), (n, ), (n, ) l(n) = (n, 00), (n, ) l(n) = (n, 00) L = {n, n, n, n, n} B PE head(l) = n $ PE # n C PE 0.0 D PE 0.0 E PE PE PE head(l) = n $min( start_time(n)) $PE # n 0.0 head(l) = n $min(start_time(n)) $ PE # n 0.0 head(l) = n $min(start_time(n)) $ PE # n F PE 0.0 PE head(l) = n $min(start_time(n)) $ PE # n PT = 0

11 The Algorithm. Parse the PDG into a hierarchy of clans. The root is the entire graph and the leaves are the graph nodes. The original node execution costs are applied to the leaves.. Traverse the tree from the bottom up making local decisions at linear clans. The decision made at a linear clan is the best sequence of clustering and concurrency for its independent children.

12 A 0.0 B C L C I C L L - Linear Clan I - Independent Clan 0.0 Parse Tree C 0.0 PE PE Parallelization of independent clan C causes node to execute separately from nodes and. Schedule completes in parallel time

13

14

15 Granularity ANCHOR ANCHOR ANCHOR ANCHOR Node Weight Range # of Graphs Node Weight Range # of Graphs Node Weight Range # of Graphs Node Weight Range # of Graphs G< <G<0. 0.<G<0,8 0.8<G<.0.0<G

16 Granularity Range CLANS DSC MCP G < < G < < G < < G < < G 0 0 0

17 0.00% 00.00% 0.00% 0.00% MCP G < 0.08 DSC 0.08 < G < < G < 0.8 CLANS.0 < G 0.8 < G <.0 CLANS DSC MCP Vertical axis: Average Relative Parallel Time Base axis: Granularity Range Granularity Range CLANS DSC MCP G < % 69.% 0.00% 0.08 < G < 0..% 0.%.% 0. < G < 0.8.8% 6.%.% 0.8 < G <.0 6.8%.% 0.%.0 < G.7% 0.6% 0.%

18 < G 0.8 < G <.0 0. < G < < G < 0. G < 0.08 CLANS DSC MCP CLANS DSC MCP Vertical axis: Average Speedup Base axis: Granularity Range Granularity Range CLANS DSC MCP G < < G < < G < < G < < G...

19 G < < G < < G < 0.8 MCP 0.8 < G <.0 DSC CLANS.0 < G CLANS DSC MCP Vertical axis: Average Efficiency Base axis: Granularity Range

20 Granularity Range CLANS DSC MCP G < < G < < G < < G < < G Node Weight Range CLANS DSC MCP Node Weight Range CLANS DSC MCP % %.088% % % % 0-00.% 8.90% 8.66%

21 % % % % % 0-00 CLANS DSC 0-00 MCP 0-00 CLANS DSC MCP Vertical axis: Average relative parallel time Base axis: Node weight range Node Weight Range CLANS DSC MCP

22 MCP DSC CLANS 0-00 CLANS DSC MCP Vertical axis: Average speedup Base axis: Node weight range DSC MCP 0-00 CLANS CLANS DSC MCP Vertical axis: Average efficiency Base axis: Node weight range

23 Node Weight Range CLANS DSC MCP Anchor OD CLANS DSC MCP Anchor OD CLANS DSC MCP

24

25

26