Generally, this machine is extremely stressed, mainly on the processor and RAM aspects. Sometimes this is combined with intensive I/O needs. We thus equip such a machine with high-end processors (nowadays Pentium III Xeon) with the greatest level 2 cache possible (nowadays 2 MB). In RAM, we would tend to reach the maximum possible, thus 4 GB. For this type of server, we should adapt the swap such as it covers all the needs in memory required by the computing codes to execute. There is often no specific constraint on network performances, except in case of a computing cluster, in which case we may consider to equip the machines with 100 Mb/s cards with a switch between them, even with Myrinet, Dolphin or Gigabit products. Typically, a LP1000 2*PIII 1.4GHz 1 GB/2*9GB is the target of this type of server. The choice is extremely variable depending on the nature of the computation to do. We should underline that there are processors, such as PA-Risc, much more performant in floating point computation than today Intel processors (up to 6 times). The new IA-64 Itanium family seems also to bring a better compromise for price/performances for that environment.
This is by the way not the only approach as shown by the i-cluster project. Cf: http://www.hpl.hp.com/techreports/2001/HPL-2001-206.html