CDF

1
CDFs Utilisation of Switched Circuit Optical
Networks
2
Introduction

• CDF is a particle physics experiment attempting
  to elucidate the fundamental nature of matter.
• It is currently taking data from the Tevatron at
  Fermilab just outside Chicago

• The experiment produces approximately 1PB of raw
  data per year.
• Analysis of this data is a global task.

3
Data Transfer Needs

• Secondary datasets are presently 1-50TB
• Typical transfer rates using the production
  network are 25Mb/s, thus a 50TB dataset would
  take over 6 months to transfer.
• Transfer rates of 700Mb/s would allow the same
  dataset to be retrieved in a week.
• The total size of these datasets is expected to
  quadruple in the next 2-3 years !

4
Network topology

• At UCL the CDF software is installed on a machine
  directly connected to the UKLight link
• At Fermilab the CDF data pools are on a fast
  network

5
Anticipated Usage

• Transfer data sets from Fermilab to UCL required
  for analysis by UK physicists
• Generate simulated data using Grid at UCL and
  send back to Fermilab
• Open up facility to other UK HEP groups with data
  at Fermilab

6
Usage so far

• Established UKLight StarLight Connectivity.
• Transferring files to test integrity and rate of
  transfer.

7
Results Experiences

• Router to Router tests have achieved a sustained
  1000 Mb/s. However it should be noted that this
  is not a FNAL PC to UCL PC transfer.
• Memory to Memory tests using UDP indicate a
  bandwidth of approximately 800Mb/s before packet
  loss is experienced.
• Packet reordering is observed above 400Mb/s

8
Results Experiences

• In the simplest disk-to-disk test, where there
  are no issues from other concurrent transfers and
  disk access from other CDF users a maximum rate
  of 250 Mb/s has been achieved.
• This is not a disk I/O limitation tests using
  the dd command have shown dzeromon capable of
  sustaining 550 Mbit/sec and cdf00-uklight of 750
  Mbit/sec for disk I/O.
• The CDF DCache to UCL disk to disk transfer rate
  was also in the 200-250 Mbit/sec rate despite the
  additional issues of other concurrent CDF user
  network and disk access.

9
Usage of storage resource managers

• Circumvent possible bottleneck of gridftp doors.

• dCache pools have about 10 parallel requests by
  different users, therefore part of the advantage
  given away.

• Further machines needed at UCL in order to start
  more parallel transfer requests in the dCache
  system