26–30 Oct 2009
Lawrence Berkeley National Laboratory
America/Los_Angeles timezone

Optimizing tape data access

27 Oct 2009, 14:30
25m
Bldg. 66 Auditorium (Lawrence Berkeley National Laboratory)

Bldg. 66 Auditorium

Lawrence Berkeley National Laboratory

1, Cyclotron Road, Berkeley, CA, 94720 USA
Storage Storage II

Speaker

Jonathan Schaeffer (CCIN2P3)

Description

TReqS is our Tape Request Scheduler. Based on BNLBatch, its goal is to get between dCache and HPSS and to reorder the files requests. Since May of this year, a first implementation hit our production system. We will present here : - the problematic of tape access for LHC experiments - the solution we implemented - TReqS in its production environement, our 5 month experiment with the beast

Summary

All LHC experiments use dCache as a front end to file access at most T1 sites.
At CCIN2P3, dCache is interfaced whith HPSS. While writing to the mass strorage system is effective, reading from it has shown to be most unsatisfying.
The major cause of this issue is that dCache has no knowledge about the files location on tape and the reading requests are rather chaotic from the HPSS point of view. It generates a lot of useless tape mount and dismount and the average dCache activity impacts the HPSS system badly.

We present here TReqS, a method for scheduling the file requests and submit them to HPSS to ensure a staging order and minimize the access latency due to the tape movements in the library. TReqS is based on BNLBatch, itself based on OakRidge batch Scheduler.

We will present the global concept and show the results and the conclusions of 6 month of tape scheduling.

Primary author

Presentation materials