Information Technology

1
T. Bowcock
Liverpool Sept 00
2
University of Liverpool

• Successes
• Issues
• Improving the system
• Comments

3
Architecture
4
Successes

• Many events generated for
• LHCb
• H1
• DELPHI
• ATLAS
• (CDF)
• Software
• Fault tolerant on transfers
• Can distribute MCs, update system etc.
• Prototype a big system

5
Issues

• Ease of use
• Expert system
• Robustness
• OS (RH5.1 and 6.2)
• Storage
• Up-time since Jan00 only - 90
• Development Recovery
• Typically run with 95 CPUs in MAP

6
Issues

• COMPASS nodes
• Do not appear as a single volume
• 10 disks /scsi1 /scsi2..
• User unfriendly? Yup.
• Do we care?
• Who is our user?

7
Bigger Issues

• Hardware failures
• NIC (20) low-cost but annoying
• Replacement fast.
• Disks
• Power Supplies
• Problem with sleave-bearing fans. Leading to
  higher failure rate. Higher quality fans will be
  installed (Sept 00)
• Down-time expected of about 1 week.
• Hubs replaced by Switches

8
Improvements needed

• Multi-Redundant Masters
• Plan 6-fold redundancy for security and
  simplicity of operation (Oct 00)
• MAP-FCS
• Need to bullet-proof it and make it idiot proof
• Interface to User
• Grid or remote login needs development
• Storage Transfer
• How do disks appear to the outside world?
• Data-Analysis Capability

9
Software Improvements

• Sept 00
• Complete upgrade to RH6.2
• Complete upgrade of MAP-FCS
• Oct-Dec 00
• reduce system vulnerability
• Large improvement from above
• GLOBUS interface?

10
Expanding MAPs role

• MAP conceived of as a MC engine
• Throwaway MC(or ntuples)
• Keeping MC involves moving data to COMPASS
• Done
• How do we (re-analyse) larger chunks of data?
• Assuming we want to do this.

11
Challenge -

• Example(a)
• LHCb. Want to produce 106 events. Reprocess it
  once or twice and then analyse it for a while.
• Optimistically 106 is about 1TByte of space
• Solution
• Increase disk store on each MAP node, and store
  data there. Analysis/repro possible.
• But implies we now need resource management
• Disks can get full up. Who gets to play?!

12
Challenge -

• Example(b)
• CDF. Want to import data from FNAL. Want to
  analyse at Liverpool.
• Solution(none yet!)
• Importing data tricky. Rely on transfer from tape
  to stage (e.g. COMPASS).
• At 5MBytes/s (Fast Ethernet) 1Tbyte would take
  200000s!!!! To get onto nodes (2days). Using 6
  COMPASS nodes about 8 hrs
• Installation Gbit
• Expensive, reduces to about 1Hr.

13
Architecture Modification
Currently (MC-mode) 6-fold redundant master can
control 300-10000 PCs
Split into subfarms to increase the I/O bandwidth
14
Subfarm Solution

• Possible
• but substantial development of system
• MC is the biggest problem
• But we still need to analyse the data
• Where is the balance?

15
so

• Suggest following steps
• Complete installation of the 6 MAP masters
  (COMPASS nodes), 0.5Tbyes/each.
• O/P from jobs can be directed there
• Increase disk capacity on existing nodes
• Purchase 300 20Gbyte disks (about 50KChF)
• Hopefully by Oct 1. (Total disk capacity of
  6Tbytes on MAP, 3 on COMPASS nodes).
• Allow users to create persistant stores on MAP
• Business (bazaar style) as usual

16
Further Improvements

• Make MAP accessible!
• Globus
• Care required
• More users
• More store
• More management.
• Package the software for distribution
• But does anybody want it?????
• Hardware Upgrades
• More nodes

17
Comments

• Can any one system provide all the facilities and
  capabilities?
• cpu, storage, data-access, i/o?
• How do institutes/regional centres really fit in?
• Balance of politics and effectiveness
• Lessons for 2004