Name: SparseLU Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs

Description

The Sparse LU application computes an LU matrix factorization on a sparse blocked matrix. The matrix size (number of blocks) and the block size are parameters of the application.

As the algorithm progresses, the area of the matrix that is accessed is smaller; concretely, at each iteration, the 0th row and column of the current matrix are discarded. ...

COMPSs Matrix Multiplication, out-of-core using files. Hypermatrix size used 2x2 blocks (MSIZE=2), block size used 2x2 elements (BSIZE=2)

Name: Matrix multiplication with Files, reproducibility example, without data persistence Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs

Description

Matrix multiplication is a binary operation that takes a pair of matrices and produces another matrix.

If A is an n×m matrix and B is an m×p matrix, the result AB of their multiplication is an n×p matrix defined only if the number of columns m in A is equal to the number ...

Name: Matrix multiplication with Files, reproducibility example Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs

Description

Matrix multiplication is a binary operation that takes a pair of matrices and produces another matrix.

If A is an n×m matrix and B is an m×p matrix, the result AB of their multiplication is an n×p matrix defined only if the number of columns m in A is equal to the number of rows m in B. When multiplying ...

Name: Matmul GPU Case 1 Cache-ON Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs Machine: Minotauro-MN4

Matmul running on the GPU leveraging COMPSs GPU Cache for deserialization speedup. Launched using 32 GPUs (16 nodes). Performs C = A @ B Where A: shape (320, 56_900_000) block_size (10, 11_380_000)             B: shape (56_900_000, 10)   block_size (11_380_000, 10)             C: shape (320, 10)                block_size ...

Name: Matmul GPU Case 1 Cache-OFF Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs 3.3 Machine: Minotauro-MN4

Matmul running on the GPU without Cache. Launched using 32 GPUs (16 nodes). Performs C = A @ B Where A: shape (320, 56_900_000) block_size (10, 11_380_000)             B: shape (56_900_000, 10)   block_size (11_380_000, 10)             C: shape (320, 10)                block_size (10, 10) Total dataset size 291 ...

Name: K-Means GPU Cache OFF Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs Machine: Minotauro-MN4

K-Means running on GPUs. Launched using 32 GPUs (16 nodes). Parameters used: K=40 and 32 blocks of size (1_000_000, 1200). It creates a block for each GPU. Total dataset shape is (32_000_000, 1200). Version dislib-0.9

Average task execution time: 194 seconds

Name: K-Means GPU Cache ON Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs Machine: Minotauro-MN4

K-Means running on the GPU leveraging COMPSs GPU Cache for deserialization speedup. Launched using 32 GPUs (16 nodes). Parameters used: K=40 and 32 blocks of size (1_000_000, 1200). It creates a block for each GPU. Total dataset shape is (32_000_000, 1200). Version dislib-0.9

Average task execution time: 16 seconds

Name: Dislib Distributed Training - Cache ON Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs Machine: Minotauro-MN4

PyTorch distributed training of CNN on GPU and leveraging COMPSs GPU Cache for deserialization speedup. Launched using 32 GPUs (16 nodes). Dataset: Imagenet Version dislib-0.9 Version PyTorch 1.7.1+cu101

Average task execution time: 36 seconds

Name: Dislib Distributed Training - Cache OFF Contact Person: [email protected] Access Level: public License Agreement: Apache2 Platform: COMPSs Machine: Minotauro-MN4

PyTorch distributed training of CNN on GPU. Launched using 32 GPUs (16 nodes). Dataset: Imagenet Version dislib-0.9 Version PyTorch 1.7.1+cu101

Average task execution time: 84 seconds