On Mon, May 24, 2021 at 2:34 PM Jeremy Allison <jra at samba.org> wrote:
> On Mon, May 24, 2021 at 02:27:51PM -0300, Daniel Lopes de Carvalho wrote:
> >Hi Jeremy, how are you doing?
> >
> >We are a petroleum engineering research lab. We work specifically with
> >reservoir simulation and we have this HPC to do this job for us.
> >
> >There are some studie case that a user needs to simulate 100, 200 or
500
> >cases at once. And these simulations take place in parallel, having an
> >intense reading and writing operation.
> >
> >Samba is installed in the headnode of this HPC, but the storage is in
> >another dedicated server and they are connected by NFS mount in an
> >infiniband network (100Gbps). The clients (users desktops) are
connected
> to
> >headnode with a gigabit network.
>
> Using NFS here isn't a great idea if you can avoid it.
>
Unfortunately I need to use NFS in this way...
> NFSv3 or NFSv4 mount ? I'm assuming this is the Linux
>
NFSv4 mount. All the HPC nodes (headnode, storagenode and compute nodes are
Linux).
> kernel NFS client ? What Linux distro, what kernel version ?
>
Kernel NFS client vers=4. Centos 8.1. Kernel 4.18.0-147.el8.x86_64
> >There are not many symptoms and after a certain time of running these
> >simulations, Windows returns this error (An unexpected network error
> >occurred) without giving details. In the Windows event logs there is
also
> >not much to investigate too. The error appears to be a momentary
network
> >outage.
>
> Network outage between clients and headnode, or between headnode
> and infiniband NFS server ?
>
The outage occurs between headnode server and clients desktops (Windows 7,
10, Server 2012 and 2019) only.
>
> >Some searches on Google have returned that it may be a Windows problem
> with
> >UAC. But I haven't investigated it in depth yet.
> >
> >I thought that Samba could have a specific configuration for these
intense
> >uses of reading and writing, so I decided to write to you and ask for
some
> >advice.
> >
> >Thanks and best regards
>
> I'd certainly advise using io_uring on Samba (presuming
> the headnode is a supported version of the Linux kernel).
>
> That should help performance-wise, but I'll need more
> info to be able to help more.
>
--
Daniel Lopes de Carvalho
daniel at cepetro.unicamp.br
unisim.cepetro.unicamp.br <https://www.unisim.cepetro.unicamp.br/>
+55 19 3521-1221