Software Architecture

As shown in the above diagram, the SCore Cluster System Software consists of the following components:

PMv2
PMv2 is a low level communication library for cluster computing. The PMv2 API (Application Program Interface) is carefully designed so that many types of networks and shared memory are accessed in a uniform way for cluster computing. PM II drivers for Myrinet, Ethernet, UDP, Shmem have been implemented.
SCore-D
A user-level global operating system to utilize cluster resources, such as processors, networks. Various PM network devices are managed and utilized by SCore-D. SMP clusters, heterogeneous clusters as well as homogeneous clusters are supported. SCore-D provides a flexible parallel job scheduling scheme. User parallel jobs are time-shared and space-shared simultaneously and effectively. Preemptive consistent checkpointing is provided by SCore-D, and any user programs can be checkpointed by only specifying a checkpoint interval time.
SCASH
A software DSM (Distributed Shared Memory) system using PM II. It employs the Lazy Release Consistency model with both update and invalidate protocols.
MPICH-SCore
MPICH-SCore, the former MPICH-PM/CLUMP, is based on the MPICH MPI library using PM II for interprocess communication. MPICH-SCore is a high-performance MPI on workstation, PC, and SMP clusters. For more information, see the MPICH-SCore Manual and Documents page.
PVM-SCore
PVM-SCore is a PVM implementation on SCore.
MPC++
A multi-threaded C++ using the C++ template feature. It provides synchronous/asynchronous remote function invocation facilities, synchronization structures, global pointer, and other such functions. For more information, see the MPC++ Level0: Multi-Threaded Template Library page.
MPC++ also supports a meta-level architecture to achieve an extensible and modifiable programming language system. The architecture makes it possible to incorporate new optimizers into the compiler. Library designers are able to provide an optimizer specific to their class/template library in the library header file. The library user may use such a high performance library by including the header file. The MPC++ meta-level architecture is not included in this release. SCore-D is written in MPC++.
TACO
TACO (Topologies and Collections) is a general purpose distributed object platform based on C++ templates. TACO provides a powerful dynamic group concept as the basic means for data-parallel programming. Entire distributed objects groups can be created, destroyed, selectively cloned and manipulated by various parallel collective operations. Thus high-level parallel programming becomes possible without any language extensions. Since TACO is implemented as a thin library layer on top of the MTTL it yields an excellent performance on cluster architectures.
SCOUT
The scout(1) program provides a SIMD-style remote UNIX shell environment. Any UNIX command can be executed on cluster hosts in parallel.

The following servers are run to maintain the system:

Cluster Database Server
All the cluster information required by the SCore Cluster System Software are maintained by the scoreboard(8) server.
Compute Host Lock Server
The Compute Host Lock Server (msgbserv(8)) provides a locking facility that prevents other users from using cluster hosts. The msgb command allows cluster users to browse the current status of a cluster. Users of parallel application programs are assumed to lock hosts via the msgbserv server. This locking management is done automatically by user programs when they are linked with the SCore runtime library.
System Information Server
The scbcast(8) server obtains system information generated by the SCore-D operating system and distributes it to clients. scbcast must be invoked before starting the SCore-D operating system and clients.

See the following sections for an introduction to the available environments:

PC Cluster Consotium

CREDIT
This document is a part of the SCore cluster system software developed at PC Cluster Consortium, Japan. Copyright (C) 2003-2004 PC Cluster Consortium.