Vladimir Litvin, Harvey Newman,

1
Long-Term Massive Production Runs on Alliance
resources Experience

• Vladimir Litvin, Harvey Newman,
• Sergei Shevchenko
• Caltech CMS

2
Introduction

• Caltech Higgs diphoton decay channel study is in
  the second stage when 10M of background events
  with full detector simulation is simulating
  reconstructing and analysing.
• 4.5M of events has been done since April 2003
• Physics results have been presented on Les
  Houghes conference and reported on CMS weeks and
  in CMS notes.
• 2004-2005 suppose to be 100M run on TeraGrid
  Alliance facility.

3
Summary
4
Analysis Chain

• Production analysis chain
• - cmsim125 and ORCAv6 (with ObjyDB) have been
  used
• - FORTRAN part can be running everywhere
• - ORCA C part was running on Caltech pTier2
  only due to ObjyDB RH6x restrictions

5

Hierarchy of problems

• There is a hierarchy of problems which might
  appear on any level
• We will concentrate mainly on the lowest,
  infrastructure, level including clusters,
  network, mass storage system, social issues
• Even at this level there are a lot of open issues

• 3-tier architecture

6
NCSA(I)

• NCSA Platinum Cluster Technical Summary
• IBM eServer x330 thin server (dual-processor)
• ECC SDRAM 1.5GB (compute nodes)
• Access nodes 4 (8 CPU)
• Compute nodes 484 (968 CPU)
• Storage nodes 32 (64 CPU)
• Intel PIII 1GHz, 256kB full-speed L2 cache (peak
  performance 1 Gflop)
• Network Interconnect
• Access node Gigabit Ethernet
• Compute node Myrinet 2000
• Disks local 10GB per node and 4 NFS mounted FSs
  650GB each

7
NCSA Statistics

• NCSA per day SU usage

• NCSA total usage

8
NCSA Statistics

• NCSA fraction of idle (Q) and running (R) time
  per day

9
NCSA Statistics

• NCSA fraction of idle (Q) and running (R) time vs
  number of jobs completed per day

10
NCSA (II)

• Should be mulitple runs submitted in one PBS
  job.Due to NCSA computing policy, PBS cannot
  allocate one CPU it always allocates one NODE.
  The same will be on future TeraGrid
• Two different tasks of two different users can
  ask large RAM in both tasks
• During one CPU allocation, second CPU is sitting
  idle anyway.
• Smaller number of allocated nodes - lower the
  priority of this PBS job
• Large number of nodes allocated per one PBS
  job is not good from Caltech HPSS usage point of
  view
• Fair maui policy is even more unfair

11
NCSA (III)

• Job submission
• !/bin/csh
• PBS -q standard
• PBS -N cmsim
• PBS -l nodes4ppn2prod
• PBS -l walltime120000
• set PBSHOST hostname
• foreach node (cat PBS_NODEFILE)
• if ( node PBSHOST ) then
• GEN_EXEout.file
• else
• ( ssh -a -x -q node "GEN_EXEout.file
  )
• endif
• end
• wait

12
NCSA (IV)

• Walltime limit exceeding problem
• Huge LAN traffic from another computing nodes in
  the same segment
• Job was started incorrectly
• Random order of running jobs
• If 100 same jobs were started, the first running
  job might be any job hard to predict the chunk
  of already ready events
• PBS MAXJOB limit was set to 50 and jobs will be
  killed by PBS when number of jobs will exceed
  this limit

13
NCSA (V)

• Blocked Jobs
• maui can start a job and if it fails, job will be
  blocked and sitting in a queue without any
  notification. Unblock must be done by hand asking
  support team to do so. There isn't any way to be
  notified by maui/PBS.
• Accounting Problem
• PBS accounting cannot calculate correctly CPU
  used time when ssh was used to start runs on
  another allocated nodes
• NCSA own Sybase based accounting system is wrong
  as well (overestimate the CPU usage)

14
NCSA (VI)

• NCSA custom made utility usage is overestimates
  used CPU hours
• corrected PBS accounting is too crude for the
  estimation as well

15
Caltech pTier2 (I)

• Server node
• Dell PowerEdge 4400, 2GB SDRAM
• 3 RAID arrays 3TB in a total
• Computing nodes
• 20 rack mounted dual CPU PIII 800MHz Intel
• SDRAM 512 MB and 133 MHz
• 10GB local disk

16
Caltech pTier2 (II)

• Caltech pTier2 layout

17
Caltech pTier2 Statistics

• Caltech per day SU usage

• Caltech total SU usage

18
Caltech pTier2 Ganglia monitor

• Screenshot

19
Caltech pTier2 Statistics

• Caltech pTier2 fraction of idle (Q) and running
  (R) time per day

20
Caltech pTier2 Statistics

• Caltech pTier2 fraction of idle (Q) and running
  (R) time vs total number of jobs

21
Caltech pTier2 (III)

• NFS related problem high CPU load average
  without any real work to do.
• Typical diagnostics in /var/log/messages
• Jul 24 062738 t007 kernel nfs server
  tier2 not responding, still trying
• Jul 24 062738 t007 last message repeated 4
  times
• Jul 24 062930 t007 kernel nfs task 3281
  can't get a request slot
• Jul 24 064252 t007 automount5189
  expired /data/raid1
• Jul 24 064406 t007 kernel nfs task 3283
  can't get a request slot
• Looks like NFS server cannot send a data on
  request due to overloading or other reasons and
  when the length of pending requests from one
  client starts to exceed the certain threshold

22
Caltech pTier2 (IV)

• Good node (Ganglia)

• Bad node (Ganglia)

23
Caltech pTier2 (V)

• Good node (Ganglia)

• Bad node (Ganglia)

24
Condor (UW-Madison)

• Condor flock of chaotically distributed nodes
• Intel/Linux 614 nodes
• Intel/WinNT50 113
• SUN4u/Solaris28 105
• SUN4x/Solaris28 6
• We are using it for FORTRAN part only

25
Condor (UW-Madison) Statistics

• Condor per day SU usage

• Condor total usage

26
Condor (UW-Madison)

• NFS problem evicted job cannot open existed
  file after start on a new node
• Failure rate depends on type of job and varied
  from
• Reason is unknown

27
Caltech HPSS

• Hardware
• 5SP2 four processor Silver nodes
• 1 SP2 eight processor High node
• IBM 3494 Robotic tape library with
• 6 IBM Magstar 3590 drives (10GB per tape, 9MB/s
• 2300 tape slots (23 TB capacity)
• StorageTek 4410 Robotic tape silo
• 4 STK Redwood drives (50GB per tape, 11MB/s)
• 6000 tape slots (300TB capacity)

28
Caltech HPSS

• Cannot keep more than 200-300 connections at the
  same time. 400 connections are killed whole
  system
• Future TeraGrid Caltech Unit will have 100TB of
  disks under PVFS

29
Conclusion (I)

• In addition to Application and Grid Middleware
  problems we have a set of infrastructure
  technical problems, policies and social open
  issues which can prevent fully utilize future
  grid capacity
• Technical issues are closely connected with
  Computing Center policies.
• PBS has a set of limitations MAXJOB limit,
  finest granularity is node, not CPU, wrong
  sequence of running jobs

30
Conclusion (II)

• Lack of good monitoring and accounting system.
  This is the basis of any future activity of
  resource brokers and it is practically impossible
  to have robust broker without detailed
  statistical information
• Reliable Mass Storage System is the critical
  issue for LHC data handling. Current limitation
  on opened connections at Caltech HPPS is raised
  questions about MSS reliability. Further
  investigations with HPSS and PVFS are essential