GPU[15-19]

[Predrag Punosevac]

FYI scratch directories are created. Due to the fact that NVMe SSD
used for OS is very small 256GB these scratch directories are
miniscule. However servers do have a bunch of 3.5" drive bays and I
can add some old drives or we could even buy new once as they are
super cheap (12 TB is about $250).

Also people should familiarize themselves with the differences

https://www.simplylinuxfaq.com/p/major-differences-between-rhel-8-and-7.html

If everything goes as planned no new instances of branch 7 will be
deployed. Existing branch 7 instances will be gradually replaced with
8.XXX but that process might take a while (read more than a year as we
have over physical machines running branch 7).

Best,
Predrag

On Tue, Jan 28, 2020 at 12:02 AM Predrag Punosevac
<predragp at andrew.cmu.edu> wrote:
>
> Dear Autonians,
>
> I just provisioned five new GPU computing nodes bought by Dr. Jeff
> Schneider before Christmas. Each server had 40 CPU threads, 192 GB of
> RAM and four RTX 2080 Ti cards. OS is installed on 256 GB SSD.
>
> For the first time our computing nodes are running Red Hat branch 8 (8.1
> version to be precise soon to be upgraded to 8.2). That is the reason it
> took me so long to provision them as I wanted to make sure drivers and
> packages are available.
>
> Nvidia drivers seems to be in working condition according to my tests. I
> did install CUDA 10.2. I didn't install cuDNN libraries as I see only
> cuDNN RPMs avaiable for Red Hat branch 7. However, I do see
> cudnn-10.2-linux-x64-v7.6.5.32.tgz which I assume is a source code so
> you are welcome to try to compile this on your own. Note that you have
> to have NVidia devel account to access cuDNN libraries.
>
> You will find miniconda3 in its usual place. /opt. Note that RedHat 8.1
> is build using gcc-8.3.1. System python is now 3.6 but I suggest you use
> the 3.7.4 from miniconda. Python 2.7 is no longer available. I am
> pushing right now MATLAB. Your scratch directories will be created
> tomorrow after I get some sleep. R is there as well as Git 2.18.2. Yes I
> did test network login both legacy as well autofs accounts. Note that
> /opt/rh for now is non-existing as SCL are still not available. Namely,
> default software is rather new so there is no need yet for newer
> versions.
>
> Please stay away from these machines if you are not willing to help
> debugging remaining  issues.
>
> Best,
> Predrag
>