thin-long 행렬과 long-thin 행렬의 곱셈 성능 측정

[kmu-leeky]

스파크 MLLib 활용 매트릭스 사이즈는 1M * 1K 정도로 block 나누는 법은 행별, 열별, 블락별을 각각 따로 시도해봐야 할듯

[namnam21]

test1 (초기에 동일한 블럭개수에서 numMidDimSplits 값만 바꾼 경우) val leftRowParallelism = 2, leftColParallelism = 8 val rightRowParallelism = 8, rightColParallelism = 2 val rowBlockSize = 1024, colBlockSize = 1024

val leftRows = 2048, leftCols = 8192 val rightRows = 8192, rightCols = 2048

val leftRowBlocks = 2, leftColBlocks = 8 val rightRowBlocks = 8, rightColBlocks = 2

numMidDimSplits=1 numMidDimSplits=4

Screen shot - Job4 numMidDimSplits=1

numMidDimSplits=4

[namnam21]

test2 val rowBlockSize = 512, colBlockSize = 512 val leftRows = 2048, leftCols = 8192 val rightRows = 8192, rightCols = 2048 val leftRowBlocks = 4, leftCol Blocks = 16 val rightRowBlocks = 16, rightColBlocks = 4 numMidDimSplits=1

val rowBlockSize = 1024, colBlockSize = 1024 val leftRows = 2048, leftCols = 8192 val rightRows = 8192, rightCols = 2048 val leftRowBlocks = 2, leftColBlocks = 8 val rightRowBlocks = 8, rightColBlocks = 2 numMidDimSplits=4

Screen shot - Job4 numMidDimSplits=1

numMidDimSplits=4