Closed vgamayunov closed 9 months ago
PR #294 should resolve this - it will at least write the rule as documented here: https://learn.microsoft.com/en-us/azure/storage/files/nfs-performance#increase-read-ahead-size-to-improve-read-throughput
fix is available in the marketplace images microsoft-dsvm:ubuntu-hpc:2004:20.04.2023111801 microsoft-dsvm:ubuntu-hpc:2204:22.04.2023111801
Troubleshooting an issue with very slow read performance of a Lustre filesystem (~100x) we found that it is caused by the NFS read-ahead setting done by the HPC images.
This is how it is done in hpc-tuning.sh:
The value of
ATTR{read_ahead_kb}=15380
seem to be applied to all devices, not only NFS, including Lustre mounts.This is how teh issue can be seen - with 10MB block size the performance drops big time compared to 1MB:
Checking the read-ahead value for lustre - it is set to 15380:
Setting read-ahead to 0 fixes the slowdown:
Can the udev rule be written in a way that it only affects nfs mounts and not applied to all devices?