The GPMIMD-2 Meiko CS-2
Back in 1992 a series of discussions on the introduction of parallel computing into HEP was held. The final outcome was the signing, by the then Director-General Carlo Rubbia, of a contract which, among other things, provided for the installation at CERN of a 32-node Meiko CS-2 (Computing Surface 2). This installation was to take place within the framework of a European Union ESPRIT project -- GPMIMD-2 -- and the computer was to be funded by the European Commission.
The purpose of having such a computer here is to determine whether it is possible to run typical HEP codes easily and effectively on this type of platform. To this end, the Commission is also funding a small group of Fellows who are endeavouring to adapt various long-running codes to the CS-2. Also funded are the manpower to set up a usable service on the machine and that necessary to run the overall project, CERN being the lead partner.
Parallel processing has been traditionally associated with numerically intensive computation and has the reputation of being difficult to implement. At CERN we are talking about parallel processing in the widest possible sense and for typical HEP applications. The principal concept is the allocation of several processors at the same time, ``in parallel'', in order to solve a problem or to obtain results in a shorter time. In its simplest form parallel processing just involves running several batch jobs at the same time, as has been done on workstation clusters for some years now, although certain other applications might use a finer grained parallelism taking advantage of the CS-2's high-bandwidth, low-latency communication system.
The CS-2 was installed last summer and quickly passed its acceptance tests. A lot of work has since been carried out to test and tune many aspects of the operating system and hardware. Each of the 32 nodes is a SuperSparc chip running at 40MHz (four at 50MHz) and each with 32MByte of memory (four with 64MByte) and 1GBbyte of disc (two with 2GByte). It runs standard Sun software -- Solaris 2.3 with the usual f77, C and C++ compilers -- as well as debugging tools such as BBN's Totalview. However, to reduce overall costs, the EPC f90 and C compilers have also been installed and these are available on all the 32 nodes without restriction. (These compilers are freely available for use elsewhere at CERN, see CNL 217.) The nodes are connected internally as a fat tree network with fast Elan interfaces acting as 8-way crossbar switches, giving a latency as low as 10s and transfer speeds of up to 50MBytes/sec point-to-point. Externally, the CS-2 has the usual Ethernet, FDDI and HiPPI connections.
The nodes are partitioned such that some are used for login and others allow large jobs to use up to 24 nodes at a time.
The CS-2 does not (yet) provide a general service, but is already being used to perform large amounts of simulation for the NA48 experiment. Its simulation program accesses multiple nodes by making use of CHIMP (an interprocessor communication package from Edinburgh University). There is also a large amount of simulation being carried out for DELPHI with a specially adapted prototype version of GEANT using MPI, a newly-emerging message passing standard. This version of GEANT is based on so-called event parallelism, in which one job controls the scheduling of event processing over many processors. Other applications include second-level triggering studies within the EAST project and the porting of PIAF with the intention, possibly, to extend the production service to this platform.
The CS-2 thus already provides significant capacity for the CERN physics programme and this will be quadrupled next year by an EU-funded replacement of each processor by two new ones running at 90MHz as dual shared-memory symmetric processors. At that time, the project would be happy to entertain further requests for significant amounts of parallel capacity from the physics community.