Discussion about fast storage evaluation

Hi Yann,

we would be interested in testing these out in conjuction with a GPU node.
Would this be possible?

Our application is in reducing a dataloading bottleneck for the training of our networks with the latest GPUs.


Hi Johnny,

sure: I’ve created a directory for you:

[root@gpu002.cluster ~]# df -h /srv/flashblade/users/raine
Filesystem                 Size  Used Avail Use% Mounted on   55T     0   55T   0% /srv/flashblade/users

Thanks for the feedback.

1 Like

Dear Yann,

I would be also interested in trying out these fast storages.
Could you also create a repository for me?

Thanks and cheers,

Hi, it’s Done. /srv/flashblade/users/guthma.

1 Like

I was playing around a bit with reading and copying hdf5 files with python from /srv/flashblade/users/guthma and the writing is by a large factor slower than when writing e.g. on /srv/beegfs/scratch/groups/dpnc/atlas/

Manipulating 200 files takes some minutes on /srv/beegfs/scratch/groups/dpnc/atlas/ while on /srv/flashblade/users/guthma it took hours

I used the cpu node277 for testing this

Hi, you are right, it seems we have a network issue :thinking:

We’ll investigate next week and update this post.

edit: this should now be solved. I forgot to update the post end of the week.

In general, I also experience slow hdf5 reading on baobab GPU nodes.
On a laptop, reading the same file maybe 10x faster than on the GPU node.
(the hdf5 file is in my scratch/ folder)


Hi Julien,

Accessing the file from a CPU or GPU node should be the same. As you are talking about your scratch folder, it means we are talking about network shared disks. On your laptop, you are probably using a local SSD, which is clearly faster. The storage on the cluster may be slower or faster depending on the user load.

On the cluster, you may use a local storage as well: hpc:storage_on_hpc [eResearch Doc]

Hi, we’ll meet the storage vendor soon, do you have some feedback for us?