GPUDirect Storage

I used the provided tool, gdsio, in the GDS lib to test the GPU Storage performance. The underlying SSD is Samsung 970 pro and GPU is Quadro RTX 8000. I performed seq read test and transferred 128MB from the file whose size is 4GB to the GPU GDS buffer. Then I found the measured throughput could be as high as 10GB/s. Is it possible the throughput can be higher than the native SSD throughput (seq read 3500MB/s)?

Ideally no if testing with -x 0 option and populated files. However if the data is all sparse data, then the result can be high as you indicate.

please check if the data is sparse data or the data is coming from the page cache.

1 Like

Sparse data means there are holes in the file? The file size does not match with the real data size, am I correct?

Yes thats correct.

Can GDS work on GTX 1650 GPU?

I should be supported in compatibility mode

Hi kmodukuri,

Could you provide further details about this specific type of sparse data? I’m particularly interested in understanding if, in the absence of compatible mode, achieving a bandwidth higher than SSD specifications is still feasible from GDS.