OpenAFS logo


June 1-5, 2009

Stanford University

Thursday 4 June 12:00pm

Location: Old Union - Clubhouse

Speaker: Hartmut Reuter

Title: Direct Client Access to AFS Vice Partitions

Abstract:

Direct client access to AFS vice partitions have been tested by Andrei Maslennikov at gridka.de (the german tier 2 centre for the new experiment at CERN) on behalf of HEPIX. Andrei has compared different access methods to data stored on a disk server for many parallel running jobs on a number of clients.

It turns out that today Lustre and GPFS offer the best performance compared with xrootd, dcache and AFS.  However, with my direct vice partition access extensions AFS comes back into play because clients with mounted GPFS /vicepa partition are nearly as fast as native GPFS, but the big advantage is that outside the cluster the data are world-wide visible and accessible as well, protected by normal AFS means. For Lustre no tests have been performed, but I hope we will have by the time the workshop takes place and I also hope the results will be similar to those with GPFS. In these cases the fileserver acts mainly as a metadata server, checking access rights and returning the coordinate of the file to the cache manager.  Of course, this technique should be used only in a trusted cluster...

Slides:

Open Office PDF