observingClouds / slkspec

fsspec filesystem for stronglink tape archive
5 stars 2 forks source link

Can the StrongLink disk cache be used? #27

Open d70-t opened 1 year ago

d70-t commented 1 year ago

To my understanding, StrongLink will always put files into its own disk cache when retrieving from tape. Is there a way to access this disk cache and have a command similar to slk retrieve, but instead of copying the file to another location (e.g. /scratch), the tool could return the path to the StrongLink cache.

This would

observingClouds commented 1 year ago

This would be ideal but isn't working/planned as far as I know. cc @neumannd

neumannd commented 1 year ago

@d70-t

StrongLink can be mounted via NFS. If this is done, all files, which are stored in StrongLink, are "visible". Thus, you can go through directories via cd and list there content.

Currently, a auto-recall feature is activated which means that files are automatically copied from tape to cache ("recalled" as StrongLink calls it) when they are accessed via NFS. It might cause some trouble if someone does a grep -R "a" /arch. Therefore, it is somehow dangerous to allow NFS access. If the auto-recall is not activated then users need to do an slk recall to get the files from tape into the HSM cache. The part which causes issue in the moment is the tape access. Copying data from HSM-cache to lustre is quite fast with slk and works quite smooth (if a file is already in the HSM-cache). Therefore, the NFS mount would not considerably improve the current situation. Additionally, I am not sure if we can simply turn on and off the auto-recall feature.

We have installed StrongLink version 3.2 (this is the version of the server software). StrongLink Version 3.3 is said to support Lustre mount -- but version 3.2 does not. However currently, StrongLink 3.3. cannot be connected to our old tape library. Therefore, we cannot upgrade to StrongLink 3.3 in the moment. There are different solutions discussed. If were on StrongLink version 3.3, the Levante Lustre filesystem could be mounted to StrongLink and the system could copy data to the Lustre file system. However, we never used this feature and we do not know how well it works -- performance, error rate, ... .

florianziemen commented 1 year ago

The stronglink cache is much smaller than the lustre file system (and further away from the compute). We currently have ~ 2 PB of cache, while Lustre is 130 PB. In the current situation, where the cache is rather full, files will get deleted more-or-less daily in the HSM cache, while users can influence removal times in their (shared) cache.

So, I think we should stick to downloading things to lustre.