HEP Computing Status Sheffield University Matt Robinson Paul Hodgson Andrew Beresford
Interactive Cluster • • • 30 self built linux boxes AMD Athlon XP cpu’s, 256/512 meg ram OS Scientific Linux 303 100 megabit network Use NIS for authentication, NFS mount /home etc System install using kickstart + post install scripts Separate backup machine 15 Laptops mostly dual boot Some MAC’s and one Windows Box 3 Disk servers mounted as /data 1 /data 2 etc (few TB)
Batch Cluster • • 100 cpu farm Athlon XP 2400/2800 OS Scientific Linux 303 NFS mounted /home and /data Open. PBS batch system for job submission Gigabit Backbone with 100 MBit to worker nodes Disk server provides 1. 3 TB as /data Raid 5 Entire cluster assembled in house from OEM components for less than 50 k • Hard part was finding air-conditioned room with sufficient power
Cluster Usage
Software • • • PAW, CERNLIB etc Geant 4 ROOT Atlas 10. 0. 1 FLUKA ANSYS, LS-DYNA
Comments - Issues • Have tightened up security in last year • Strict firewall policy, limited machine exemption • Blocking scripts prevent ssh access after 3 authentication failures within 1 hour • Cheap disks allow construction of large disk arrays • Very happy with SL 3 for desktop machines • Use FC 3 for Laptops – 2. 6 kernel
The Sheffield LCG Cluster
Division of Hardware • 162 x AMD Opteron 250 (2. 4 GHz) • 4 GB RAM/box (2 GB/CPU) • 72 GB U 320 10 K RPM local SCSI disk • Currently running 32 bit SL 303 for maximum compatibility with grid. • ~2. 5 TB storage for experiments. • Middleware: 2. 4. 0 • Probably the most purple cluster in the grid.
Looking Sinister
Status
Usage so far • We can take quite a bit more.
Monitoring • Ganglia with modified webfrontend to present queue information
Installation • Service nodes connected to VPN and Internet • PXE Installation via VPN allows complete control of dhcpd and named • Red. Hat kickstart + post install script • ssh servers not exposed • RGMA always the hardest part • Stumbled across routing rules. • WN install takes about 30 minutes, can do up to 40 simultaneously.
Matt Robinson: Future plans • Keep up with middleware updates • Increase available storage as required in ~3 -4 TB steps • Fix things that break • Try not to mess anything up by screwing around • Look toward operating with 64 bit OS.
Скачать презентацию HEP Computing Status Sheffield University Matt Robinson Paul
19cda765c02c4e99aecda7c7cdb03995.ppt