Plans for the CERNSP and CERNVM Services
CERN recently signed the contract for the replacement of the IBM ES9000 running CERNVM by a half-sized machine and the purchase of a 64 node IBM Scalable POWERparallel 2 (SP2) system. The SP2 was delivered to CERN on 30 October. Note that the original plan to install a second set of 64-nodes in 6 months' time has now been cancelled on the advice of the HEP Computing Coordinating Committee. We will initially dedicate 16 nodes to interactive use, 16 for serial batch, 16 for server functions (disk, tape and network servers and new FATMEN + HEPDB platforms) leaving 16 nodes for running parallel interactive and batch. The main impact of the reduction is on the parallel applications. That part of the machine will now be used to evaluate and demonstrate the foreseen parallel applications but not offer a production service. The machine remains inherently scalable should these parallel applications demonstrate a convincing case for production services.
The ES9000 will be replaced by a 3090-600J over the Christmas shutdown. This machine has 6 CPU units as does our present machine but runs at half the clock speed. We do not believe this will degrade interactive response but batch job turnaround times will increase, with priority being given to interactive work during normal working hours. There will be no Vector Facility on the replacement machine but this was only ever used by a very small community. The 3090-600J will run the full CERNVM system and all user mini and maxi disks will be kept initially. We will reduce the pool of staging disks by about two hundred gigabytes from its current size of three hundred gigabytes but still expect good lifetimes since the main user has been the LEP batch work that will be stopping on CERNVM at the end of this year.
The new 3090 should deliver about 115000 CERN Unit hours per quarter. Each LEP experiment is allowed 15000 hours per quarter, combining batch and interactive, leaving 55000 hours for non-LEP usage, which matches their current usage of CERNVM. In this model there is little point in going through the labour intensive exercise of asking groups to make a formal request for CERNVM time next year. The Cocotime committee has agreed to the proposal that groups be allocated CERNVM budgets next year equal to what they have actually used this year and these budgets have now been published (see a recent Computing news).
The SP2 is currently running an experimental service for selected users. We anticipate a general SP2 service from early next year and Cocotime has agreed that its use need not be budgetted immediately but that this should be applied without too much delay. We would like groups who are interested in using this service early next year to contact us (Tony.Cass@cern.ch or Harry.Renshall@cern.ch) for setting up a group environment in AFS. The SP2 public serial node partition will have about seven times the capacity of CERNVM so those groups migrating to it from CERNVM or newly starting small to medium sized experiments should plan their future computing work with this in mind.