NorduGrid technical meeting

17 June 2004, Copenhagen

Minutes

Present:

Daniel Kalici, Mike Gindonis, Jonas Bardino, Balázs Kónya, Aleksandr Konstantinov, Oxana Smirnova, Jukka Klem, Henrik Thostrup Jensen, Jesper Ryge Leth, Jakob Langgaard Nielsen, Martin Folkman, Mattias Ellert, Anders Wäänänen, Michael Grønager, Niels Elgaard Larsen

Next meeting:

In Reykjavík, September (? TBD).

Knoppix CD:

Presentation by Niels

Status:

CD working as:

cliet and server (frontend + worker node) standalone
client with NorduGrid certificate read from: Floppy/USB/mounted Windows harddrive

Not working as server with host certificate - ideas:

Nice to have an Unsafe CA - Automatic for Knoppix use
Do we want screensaver science

(feature req for Frontends running behind NAT... from Daniel to Aleksandr)

Job Manager:

Presentation by Henrik

Notes from the following discussion:

The present user interface will be implemented using the Job Manager (as a test) – to day only job submission can be handled.
Very pluggable: Just a new handler for each of the new ng* commands.
Failover: Two (or more) Job Managers can be run simultaneously – one is just taking over if the other one fails.
No state save to disk – considering to use a database instead (probably Zope Object database).
In one month (middle July) current functionality (ng*) will be implemented.
For more see: the proposal
Implemented using XMLRPC (a lightweight/predecessor version of SOAP)

ATLAS DC2 – CA-policy

Problem: Grid manager has to know the CAs of a submitted job – otherwise job might fail... Solution ideas:

Advertise the CAs you authorize against
Make a list of needed CAs based on Proxy, etc...
Associate CAs with runtime envs...
Make an auth-plugin that checks if the job can connect to all storages – Aleksandr will look into this: Basically the local cluster can use an auth-plugin to check if access to input and output files of the script can be authorized – before the job enters the ACCEPTED state.
The Auth-plugin kind of functionality could also make it possible to e.g. run a ldd on all executables in the script before entering the ACCEPTED state...

Managing keys:

Install an RPM called ca-policy-atlas (or eugridpma [ed: ouch]), which depends on the needed CA-packages.

Other issues:

RLS: If you upload a file – ask RLS for this, if not found (or server down) should ask another place for this but it just fails (after trying same place 5 times).

Release of ARC: Clusters need at least 0.4.1. – when is next release: 0.4.3 - TODAY!

Status:

LRMS Port (Balázs):

Fork: Problems: More information from PBS than from fork (history, queue etc). Grid manager front end for fork not working correctly. Finished in 2 days
Sun Grid Engine: Looking into this after fork...
Condor: up and running, problems with ngclean (Mathias) – implemented as a queue – not a cluster.
All above should be up and running in a month: July 17th 2004.

Frontend model (Daniel):

Running fine
Implemented on "Horseshoe" {620CPU Linux clusters}, "Theory" {Altix, Onyx, Origin} same frontend, different queues. Installation on IBM PowerPC cluster in progress and "Niflheim" with 800CPUs.
Big test is ATLAS DC2 running on Horseshoe...

Runtime Environenment (Mike for Juha) in progress...

(Balázs) Swegrid would like to administer their own namespace – to be sure a RE really means the same - from now on to be discussed in NDGF.

Accounting (Martin):

Global Bank service, Logger (LUTS) service and different account services per project/ per VO etc.
Cluster-vise: JARM(Job account res. Manager). Auth-plugin. (clients to LUTS and BANK)
Based on Java program from GT3 (needs 1.2 and up) used in auth-plugin as well as in the global bank.
First alpha release of Admin web interface ready. Tested on Hagrid, but no big testing.
Advanced quoting schemes were discussed – like get a quote/short time reservation etc. – some could already be implemented by Johan Tordsson from Umeå.
A scheme to specify when you want the job run, and how much you want to pay.
For more see e.g. the talk at the Swegrid inauguration
Aleksandr presented NGLogger – a graphical interface to the database. (See slides by Ugur)

Mailing list

Everyone happy: with current status (Anders, Balázs, Oxana)

GM turnaround time:

Everybody agrees signalling should be implemented; Anders fills the feature request.

Data Management:

RLS problems (Jakob):

Globus RLS
Freezes regularly – randomly stops accepting connections.

Reasons for freezing are being investigated.
RC was dying, too, but that was easy to catch (error message) and restart. RLS doesn't die, just freezes.

Does not support collections.
Command-line tool is really horrible and lacks some functionality, and is also illogical.
It also leaks memory on non-authenticated searches.
All in all, it is very easy to bring RLS down.
Positive things:

allows free-form attributes to get associated with entries
works nicely with authentication.

Storage element (Aleksandr):

Currently, most SEs are NG GridFTP servers with static access control. Indexing: 1 RC, 2 RLS servers are up.

Smart SE (SSE): GM for data tasks, kind of. Has internal states (non-exposed):

                  Collecting       Downloading
                        \             /
                         \           /
                          \         /
                           \       /
                           Complete
                               |
                               |
                               |
                               |
                           Verified

Reliable file transfer. Communication: https, transfer: http, control: Web-services. So far works fine. Can support various indexing services.

File transfer is slower than FTP, but multistream transfer is possible.

An active indexing service for files is needed (see Trond's proposal). At the moment there is no network of SEs, just a bunch of disks. No automatic replication. No bulk replication.

New tool: ngrequest – asynchronous ngcopy. SSE makes the transfer and possibly LFN resolution. To be documented (Aleksandr).

SRM: Oxana to send specs. Would be nice to have SRM-compliant SEs and a bulk/automatic replication tool.

Indexing service: Jakob to write an architecture proposal.

Aleksandr will add replication funcationality to ngcopy: copying LFN to itself should create a new replica.

Issue: access rules for SE files. GACL won't scale for a VO case: it will be big, and is always static. VOMS, CAS, PERMIS: Aleksandr will have a look.

Minutes taken by M.P.G. and O.S.