1. NL Service Challenge Plans Kors Bos, Sander Klous, Davide Salomoni (NIKHEF) Pieter de Boer, Mark van de Sanden, Huub Stoffers, Ron Trompert, Jules Wolfrat (SARA), Hans Blom, Freek Dijkstra, Paola Grosso (UvA), Peter Michielse (NCF) Serv.Chall.Mtng. – CERN – February 24, 2005

2. 2 Planning as presented at RAL  January setting up  February 7-11 disk-disk tests with test cluster  March 14-25 sustained 100 MB/s disk-disk  April 11-15 disk-disk tests with new hardware  May 9-14 first disk-tape tests  May 30 – June 10 10 Gbit/s tests  July 4-29 sustained 50 MB/s disk-tape  August vacation I’ll report Changes !

3. 3 January network ENTERASYS N7 10 GE Switches CERNH7 Router FORCE10 1G 10 Itanium dual CPU 1.5 GHz, 2-4 GB RAM CERN SARA 10 Opteron dual CPU 2 GHz, 4 GB RAM 3 x 1G 10G transparent lambda 10GE LAN PHY We used 3 times 1 GE between CERN and SARA Cisco 6509 ONS 15454 HDXc Force10 3 x 1GE OC192 POS SURFnet ONS 15454 SURFnet 3 x 1GE

4. 4 Rembrand Cluster thanks to group of the UvA  Used a cluster of 4 nodes:  The whole cluster is bigger (9 nodes)  64-bit dual Opteron at 2 GHz and 3 GB RAM  2 (SATA) data and 1 system disk 250 GB per node  1 Gb/s fiber NIC  Now upgrading to:  12 disks (250 GB) per node  can do striping  10 Gb/s fiber NIC on 6 nodes

5. 5 Results Tests 7-12/2 Double stream:4x2 disks both sides4 Nodes time throughput 327 sec 196 MB/s Double stream:3x2 disks both sides3 Nodes timethroughput Best 275 sec 174 MB/s Average 348 sec 143 MB/s Worse 460 sec 104 MB/s Double stream:2x2 disks both sides 2 Nodes time throughput Best 280 sec 114 MB/s Average 323 sec 100 MB/s Worse 386 sec 97 MB/s Double stream:2 disks both sides 1 Node time throughput Best 270 sec 59 MB/s Average 285 sec 56 MB/s Worse 296 sec 54 MB/s Single stream:1 Node time throughput Best 143 sec 56 MB/s Average 198 sec 41 MB/s Worse 246 sec 33 MB/s From cern to Amsterdam 8 Gbyte files Disks RAID0 Using GridFTP

6. 6 What we have learned  With this setup:  We can do SC2: 100 Mbyte/s disk-to-disk  We reached ~200 Mbyte/s  Limitation is the disk writing speed: 50 Mbyte/s  Load balancing is difficult  MTU is important (stay away from jumboframes?)  What else we learned:  Still takes few days to set up and understand  Sander at CERN was usefull  Need more monitoring  And a proper detailed test plan !!

7. 7 What we will improve  On the Rembrand cluster:  Add more disks: 12 x 250 GB per node  Use striping for better speed:  250 Mbyte/s per node?  Use RAID for more safety  performance reduction  With 4 nodes  1 Gbyte/sec?  Could use more nodes  What else :  Switch to 10 GE  No more problems with load balancing on network

8. 8 March Sevice Challenge 2  Disk-to-disk performance tests: 14-25/3  Sustained file transfer tests  Goal transfer rate 100 Mbyte/s  6 T1’s : Lyon, Fermi, A’dam, FZK, RAL, CNAF  With current setup:  We could do more than 100 Mbyte/s  Detailed test plan needed !!  More monitoring needed

9. 9 April new hardware to connect to SARA SAN  3 Dell Power Edge 1850  32-bit dual Xeon at 3.0 GHz and 4 GB RAM  2 SCSI system disks  I/O Riser card  1 Dell PowerConnect 5324  SGI TP9300 Storage array  14 x 250 Gbyte SATA disks  2 x 9940B tape drives

10. 10 Hardware to be purchased

11. 11 April set up Internal Gigabit Network Force10 SAN HSM Shared FS Fiber Channel 1 Gb/s UvA cluster 10 Gb/s New hardware LCG tape drives

12. 12 April 10 Gbit/s tests  10 G/sec set up: 11-15/4  Problem: WAN-PHY to LAN-PHY conversion !!  Disk-to-disk performance tests  Install new hardware  Software development  SRM on DMF (prefered)  SRM plus Dcache on DMF  Decide on replacement of UvA cluster: 18/4

13. 13 May tests with tape  New set-up tests: 2-6/5  Data from CERN into HSM  Goal transfer rate 50 Mbyte/s  First tape tests: 9-14/5  Data from CERN onto tape at SARA  2 new 9940B tape drives  Expected rate 50 Mbyte/s  Software tests:  SRM

14. 14 June other 10 Gbit/s tests  Set up network for Vancouver: 30/5 – 3/6  Use full SURFnet lambda  no other users  10 Gb/s tests Vancouver: 6/6 – 17/6  After 17/6  Other 10 Gb/s tests? RAL? Univ.of Alberta?  Back to 1 Gb/s for RAL, Prague, Taipei, Vancouver, Alberta, Moscow  Install and test UvA cluster replacement

15. 15 July Service Challenge  Preparations: 27/6 – 1/7  Network: 10 Gb/s operation  Hardware: servers/disks/tapes  Manpower: operators/manpower/vacation  Operation mode: monitoring/communication  Service Challenge: 4-29/7  50 Mbyte/s from CERN to SARA tape (100 tapes)  Using SRM  Monitoring and publishing

16. 16 August vacation