[SYSTEMDS-3481] FrameFromMatrix Improvements #1755
Conversation
|
And now with parallelization: 22/12/21 22:03:53 ERROR frame.FrameFromMatrixBlockTest: 197.328218 |
|
Any ideas on how to add (or where to add) threads to all methods/internals/instructions. |
|
Reading time of boolean matrix blockSize (k), before , after
0.5, 0.557+-0.025 , 0.134+-0.005
1.0, 0.296+-0.003 , 0.111+-0.001
2.0, 0.210+-0.010 , 0.107+-0.006
4.0, 0.159+-0.009 , 0.106+-0.007
8.0, 0.137+-0.003 , 0.110+-0.006
16.0, 0.139+-0.009 , 0.108+-0.005
32.0, 0.152+-0.023 , 0.129+-0.006
64.0, 0.159+-0.012 , 0.130+-0.002
128.0, 0.112+-0.004 , 0.117+-0.004 |
Baunsgaard
force-pushed
the
MatrixBlockToFrame
branch
from
7fc78c4 to
2a258f7
Compare
This Commit introduce various updates and refinements to the FrameBlock infrastructure. In specific the modification and changing of MatrixBlock to FrameBlock is optimized. In this process the parallelization of the instructions is critical, and therefore contained in this commit is a larger change to Unary instructions to now all contain a thread count in the instruction string. This change is also effecting instructions that does not nessesarily need a thread count such as broadcast, but it did give the opportunity to refine the applySchema, toMatrix, toFrame, and other instructions to be parallel. Example: changing a Matrix 64kx2k to boolean frame: before 2.2 sec single thread, after single thread 0.9 sec, and parallel 0.13 sec. Also improved is the reading time of frames, where before the time varied drastically depending on block size saved, it is now improved from 0.56 sec to 0.13 sec on 500 block size. A final update to also imrove overall execution is compiletime, I observed that the compiletime if we include IO operations increase to 0.6 sec. While if we do not have IO operations it is 0.3 sec. This is due to the hadoop IO we are using taking up to 70% of the compile time in cases where we have simple scripts with only read and a single operation. This is a constant overhead on the fist IO operation that does not effect subsequent IO operations, to improve this i have moved this to a parallel operation when we construct the JobConfiguration. This improve the compile time of systemds in general from ~0.6 sec when using IO to ~0.2 sec.
Baunsgaard
force-pushed
the
MatrixBlockToFrame
branch
from
2a258f7 to
21c0c07
Compare
1 participant
This commit update the change of matrix to frame to more efficiently change the MatrixBlock to frames.
The previous implementation has nice cache blocks and allocation for direct double to double change, this PR simply adds support for this change in the case of changing into other types, like boolean.
Changing a Matrix 64kx2k to boolean frame:
After:
22/12/21 19:56:11 ERROR frame.FrameFromMatrixBlockTest: 1055.994364
22/12/21 19:56:12 ERROR frame.FrameFromMatrixBlockTest: 1039.756463
22/12/21 19:56:13 ERROR frame.FrameFromMatrixBlockTest: 946.029085
22/12/21 19:56:14 ERROR frame.FrameFromMatrixBlockTest: 928.161053
22/12/21 19:56:15 ERROR frame.FrameFromMatrixBlockTest: 943.132151
22/12/21 19:56:16 ERROR frame.FrameFromMatrixBlockTest: 950.212744
22/12/21 19:56:17 ERROR frame.FrameFromMatrixBlockTest: 964.515222
22/12/21 19:56:17 ERROR frame.FrameFromMatrixBlockTest: 966.944032
22/12/21 19:56:18 ERROR frame.FrameFromMatrixBlockTest: 965.85695
22/12/21 19:56:19 ERROR frame.FrameFromMatrixBlockTest: 956.783357
Before:
22/12/21 19:59:56 ERROR frame.FrameFromMatrixBlockTest: 2199.846241
22/12/21 19:59:58 ERROR frame.FrameFromMatrixBlockTest: 2373.381971
22/12/21 20:00:01 ERROR frame.FrameFromMatrixBlockTest: 2270.362306
22/12/21 20:00:03 ERROR frame.FrameFromMatrixBlockTest: 2324.07255
22/12/21 20:00:05 ERROR frame.FrameFromMatrixBlockTest: 2294.39046
22/12/21 20:00:08 ERROR frame.FrameFromMatrixBlockTest: 2284.978142
22/12/21 20:00:10 ERROR frame.FrameFromMatrixBlockTest: 2295.71655
22/12/21 20:00:12 ERROR frame.FrameFromMatrixBlockTest: 2297.712022
22/12/21 20:00:14 ERROR frame.FrameFromMatrixBlockTest: 2311.518135
22/12/21 20:00:17 ERROR frame.FrameFromMatrixBlockTest: 2467.055097