Fermilab Offline Computing Farms

1
Fermilab Offline Computing Farms

• Stephen Wolbers
• Fermilab
• May 22, 2001

2
Farms at Fermilab

• Farms have a long history at Fermilab, providing
  large CPU at low cost.
• Ill discuss the reconstruction farms, used for
  event processing and for Monte Carlo simulation
  and for large-scale calculations.
• The farms are all dual-Pentiums running Linux,
  logically divided into CDF, D0, CMS and Other.
• Batch system is Fermilab-written FBSNG, a product
  which is the result of many years of evolution
  (ACP,cps,FBS,FBSNG).

3
Characteristics

• The farms are designed for small numbers of large
  jobs and a small number of expert users.
• The goal is to get physics results as quickly as
  possible from the experiments and to provide very
  large cost-effective CPU to projects.
• There are 314 PCs in the farms and 124 more are
  on order
• CDF 88 nodes 4998 SpecInt95 (64)
• D0 100 nodes 5702 SpecInt95
• CMS 40 nodes 2856 SpecInt95 (20)
• Other 86 nodes 3019 SpecInt95 (40)

4
Run II CDF PC Farm
5
OLD
6

NEW
7
Data Volume per experiment per year (in units of
109 bytes) CPU needs increase at least as
quickly.
Data Volume doubles every 2.4 years
8
Future

• The offline farms do their job well. They are
  inexpensive, run well, dont require large
  amounts of system or software support.
• I expect the current model to continue to hold,
  with the machines continuing to increase in speed
  and also having better connectivity.
• Other models of PC cluster computing will take
  hold, but with different use patterns and
  management issues.
• Disk Farms, Analysis Farms, Server Farms
• Could we in 5 years see only PCs in our big data
  centers?