Overview
In response to ever increasing requirements for lower-latency and higher throughput file services to support modern AI/ML workloads, we have added an all-NVMe based WekaFS tier to our storage profile. This highest performance parallel filesystem is directly attached to our new HPC cluster serving I/O over HDR200/100 InfiniBand. The existing Isilon NAS has been moved exclusively to Campus-wide SMB service for the labs and desktops.
...
Storage Tiers | C2B2 Cluster | New Cluster |
|---|---|---|
home | home/PI_gp/userUNI | users/userUNI |
data | data/PI_gp/userUNI | groups/PI_gp/userUNI |
archive | archive/PI_gp/userUNI | archive/PI_gp/userUNI |
Users (Previously home)
Exclusive free storage space of 20GB 50GB assigned to each C2B2 Cluster user.
Used for small input files, source code files and executables, software builds etc.
POSIX/NFS shared
Backed up daily (multiple times).
Groups (Previously data)
Shared space of 1TB storage allocated by default to each PI_group.
General space for group software, common data, etc.
POSIX/NFS shared
Backed up nightly.
Additional storage can be requested in increments of 1TB.
Archive
Shared cold storage space of 1TB allocated by default for each PI_group.
PI can purchase spinning disk storage to store their data for long term.
SMB mounted only, can be accessed via Globus
Backed up monthly to another S3 block storage device (in the same building but different floor)
...
Additional storage can be requested in increments of 1TB.
Localscratch
Since all tiers on Weka system are equally fast, faster than traditional scratch space, there is no dedicated "Scratch" tier provided from Weka system.
1.5 TB of local scratch space is provided on an NVMe on each node, mounted at /localscratch.
As the name indicates, this storage space is local to the node, shared by all users, no quotas enforced, cleaned up upon termination of running jobs, and is provided free of cost.
Users who want to utilize the above scratch space during job execution, should create /localscratch/user_uni/ upon job start, pull data from /groups global space, process the data, copy processed data back to /groups permanent space, and delete /localscratch/user_uni/ before job termination.
...
No matter how fast filesystem you have, there is VERY HIGH overhead in opening a file for read or write ops. This will certainly add up if you have large number of small files, and you will be surprised how bad it could get. We cannot emphasize enough that one should avoid creating large number of small files at all cost!!! In addition, use only plain-English characters, without spaces, without special characters, to name the files & folders.
Groups and Archive tier storage are available in capacity increments of 1TB.
...