Cristian Tatu

Distributed computing aims to offer tools and mechanisms that enable the sharing, selection, and aggregation of a wide variety of geographically distributed computational resources in a transparent way. The research done in this team is based on the past expertise of the group, and on extending it towards the aspects of distributed computing that can benefit from this expertise. The team at BSC has a strong focus on programming models and resource management and scheduling in distributed computing ...

Project that aims to create the NeuroPlat portal for neurodrug design

Name: Matmul GPU Case 1 Cache-ON Contact Person: cristian.tatu@bsc.es 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: cristian.tatu@bsc.es 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: cristian.tatu@bsc.es 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: cristian.tatu@bsc.es 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: cristian.tatu@bsc.es 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: cristian.tatu@bsc.es 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

Name: TruncatedSVD (Randomized SVD) Contact Person: support-compss@bsc.es Access Level: public License Agreement: Apache2 Platform: COMPSs Machine: MareNostrum4

TruncatedSVD (Randomized SVD) for computing just 456 singular values out of a (3.6M x 1200) size matrix. The input matrix represents a CFD transient simulation of aire moving past a cylinder. This application used dislib-0.9.0