The PS resources in Maxwell serve the Photon Science staff members. It consists of the PS partition, interactive compute nodes and GPFS storage infrastructure.
For details about the setup of the partitions and compute nodes please have a look at Compute Infrastructure page, and the limits applying.
Interactive Login Nodes
Preferably use max-display.desy.de or https://max-display.desy.de:3443/ to connect to the Maxwell cluster, and connect from there to interactive compute nodes - or submit batch jobs.
Please have a look at Access and the Interactive Login. max-display is directly accessible from outside (no tunneling required).
Login nodes are always shared resources sometimes used by a large number of concurrent users. Don't run compute or memory intense jobs on the login nodes, use a batch job instead!
For other information have a look at Getting started.
Interactive Compute Nodes
Summary
- For login to CPU workgroup server connect to max-fsc.desy.de. Don't use individual node-names.
- Be aware that you’re not alone on the systems.
- max-fsc is not directly accessible from outside tunnel. Tunnel through bastion.desy.de, or use max-display.desy.de
- For login to GPU workgroup server connect to max-fsg.desy.de. Don't use individual node-names.
- Be aware that you’re not alone on the systems
- max-fsc is not directly accessible from outside tunnel. Tunnel through bastion.desy.de, or use max-display.desy.de
- Be aware that you’re not alone on the systems
- If you cannot access the system (not a member of netgroup @hasy-users), please contact fs-ec@desy.de
- If you’re missing software or for support, contact maxwell.service@desy.de (and in cc: andre.rothkirch@desy.de)
- dCache instances petra3 and flash1 are mounted on max-fsc/g (/pnfs/desy.de/ ; no Infiniband)
- A gpfs-scratch folder is available by /gpfs/petra3/scratch/ (no snapshot, no backup, automatic clean-up [will come]).
it's scratch space in a classical sense, i.e. for temporary data. This space is used by default as temporary folder for Gaussian and you can create a folder for yourself (e.g named after your account) for other purposes. It's shared among all users and thus you're ask to check (in particular in view of gaussian) and to clean up / free scratch disk space if no longer needed.
Currently you can use the resource without scheduling. We will monitor/see how it goes without such a scheduling measure. In case of ‘crossfire’ / users disturbing each other we have the option to put a scheduling similar to HPC or Maxwell core (i.e. SLURM) in front of the resource, the how / in which way needs to be discussed then within photon science (e.g. scheduler only for max-fsg or for max-fsg and 50% of max-fsc or ….).
Hardware environment
The photon science (FS) maxwell resource consists currently of 11 nodes:
- 8 workgroup server equipped with 32 cores and 512GB ram.
- CPUs are Intel(R) Xeon(R) CPU E5-2698 v3 @ 2.30GHz
- check the benchmarks
- 3 workgroup server / compute nodes with additional Cuda card for GPU computing
- 32 cores and 512GB ram
- 1 x Tesla K40m
- cuda 7
The workgroup server are connected by Infiniband (FDR, 56Gb) to the GPFS storage system for PETRA III and thus should e.g. allow for faster access/analysis of larger amounts of data.
Personal environment
The home-directory is not the usual AFS but on a network-storage (currently hosted on GPFS) located in /home/<user>:
[user@max-p3a001 ~]$ echo $HOME /home/user
This working directory resides on GPFS and thus you’ll have the same “home” on all workgroup servers belonging to the new resource. Please note that this “working directory” should NOT be considered as replacement for your AFS or Win Home-directory. In particular the home-directory on max-fs is NOT in backup!
The usual afs-environment will not be available on max-fs! However, upon login you will obtain an afs-token and kerberos-ticket:
[user@max-p3a001 ~]$ tokens Tokens held by the Cache Manager: User's (AFS ID 3904) tokens for afs@desy.de [Expires Jun 20 09:40] --End of list-- [user@max-p3a001 ~]$ klist Ticket cache: FILE:/tmp/krb5cc_9999_4GphrK1aA2 Default principal: user@DESY.DE Valid starting Expires Service principal 06/19/2015 11:01:51 06/20/2015 09:40:08 krbtgt/DESY.DE@DESY.DE renew until 06/21/2015 09:40:08 06/19/2015 11:01:51 06/20/2015 09:40:08 afs/desy.de@DESY.DE renew until 06/21/2015 09:40:08
Tokens and tickets will expire after about 24h. In contrast to AFS-homes, accessing the home-directory does not require tokens or tickets, which means that long-running jobs will continue after token expiry. Just ensure not to introduce dependencies into your AFS-home. If needed, tokens and tickets can be renewed as usual with e..g:
[user@max-p3a001 ~]$ k5log -tmp
Your home-directory is subject to a non-extendable, hard quota of 30GB. To check the quota:
[user@max-p3a001 ~]$ mmlsquota max-home
Block Limits | File Limits
Filesystem type KB quota limit in_doubt grace | files quota limit in_doubt grace Remarks
max-home USR 76576 0 10485760 0 none | 520 0 0 0 none core.desy.de
Storage environment
The GPFS core filesystem is mounted on the system as
/asap3
Below /asap3, you will find several facilities, e.g. /asap3/petra3/gpfs or /asap3/flash/gpfs
Underneath you’ll find data taken recently at the facilities, sorted in subfolders by beamline with additional substructure of year/type/tag, e.g. data in 2015 at beamline p00 with beamtime AppID 12345678 would reside in /asap3/petra/gpfs/p00/2015/data/12345678/, followed by the directory tree as given/created during beamtime. Folder “logic” is same as during beamtime, i.e. assuming you’ve been named participant for a beamtime and are granted access to the data (controlled by ACls [access control list]), you can read data from subfolder “raw” and store analysis results in scratch_cc (temporary/testing results) or processed (final results).
For further details concerning folders and meaning please see subsections in the ASAP3 Confluence Space.
The folder exported read-only to the Beamline PCs foreseen for documentation, macros etc. can be found in folder /asap3/<facilitiy>/gpfs/common/<beamline>.
AFS is fully available on all max-nodes.
dCache is currently not available on max-fs nodes. If access to dCache hosted data is required, use the P3WGS6 nodes, which mount both dCache and GPFS.
Remarks
Please note that these servers are currently accessible without a scheduling system as used for HPC. Thus the systems (or a single WGS) are shared resources among all of you and multiple users can be logged in consuming given RAM and cores in parallel. Please keep this in mind and pay attention on e.g. core consumption and RAM demands when running own multithreaded jobs (i.e. restrict your jobs not to use all cores/RAM available on a single server) and do not try to ‘optimise’ program execution such as you might do if a WGS is exclusively yours alone. You are also recommended not to distribute jobs on multiple nodes.
With respect to the GPU (Cuda) nodes, please keep in mind that currently GPU cards do not allow for running jobs in parallel by e.g. different users. Thus do not execute/start GPU processing while the card is already in use, it will lead to crashes. You may check the Cuda status by calling “nvidia-smi” from the shell, besides others it will report on running processes.
The environment incl. storage system mounts will further evolve, e.g. next to come are that data from external beamtimes carried out by DESY staff will be put to GPFS, and a “scratch” space will be added to allow for processing of previously taken data (to be copied from dCache then) or processing data not originating from a PETRA Beamtime (e.g. simulations/theory).
Though the nodes are all connected via infiniband, please do not run any processes on multiple hosts (for example with MPI). Due to current restraints, the workgroup server are split over two infiniband-fabrics, which limits the possibilities of multi-host processes.