mlcommons
Is it possible to select a specific disk on which to run the benchmark?
Hi, I have some questions.
-
I want to run the benchmark on an nvme disk that is not mounted on the file system. However, it seems that this evaluation is conducted in a logical test(file system mount) environment, not a physical test) If I want to run the benchmark and measure performance on the nvme disk I want to test, do I need to mount the nvme disk to the file system and install the tool in that location?
-
Is there any guides about parameters(ex. --num-accelerators, --num-parallel, --client-host-memory-in-gb) to run mlperf-storage when performing "datagen" and "run" step? I'm confused about what parameters are appropriate and what parameter should be set to suit my computer and storage environment. +a) Do I have to match the number of parameters of "--num-parallel" and "--num-accelerators"?
-
Is it not possible to run the BERT model? According to the mlperf-storage benchmark results, there seems to be a BERT model workload, but I can't find it. You are also explaining the BERT-Model Workload on your website (https://mlcommons.org/2023/06/introducing-the-mlperf-storage-benchmark-suite).
I'll be waiting for your answer of above questions. Thank you.
I followed this guide, and it was helpful for me. https://github.com/mlcommons/storage/blob/main/README.md
