Poor efficiency on lustre filesystems

[privefl]

I have been using mio to allow some memory-mapping access to disk files in my packages (R/C++); it works quite well, thanks for that.

However, a few users who use a HPC cluster with lustre filesystem, report very bad performance when accessing the files via memory-mapping, especially when trying to do so in parallel.

Anyone has experienced this issue? And could provide some workaround?