System Software

SYSTEM SOFTWARE

Cluster Management System

	SIIA System: Cluster Administration System
	CMagic: Cluster Replication and Configuration System
	CMon System: Cluster Monitoring System
	Pcmd: Parallel Unix Commands
	CNETmon: Cluster Network Monitoring System
	CSec: Cluster Security Management System

Standard Communication Libraries

	Optimization of MPI for Efficient Collective Communication Operations
	The global focus of this research is to propose efficient solutions for collective communication operations in software and hardware. For this purpose, we study the behavior of some collective communication solutions measuring their execution times and their influence on the overall performance of several parallel programs.
	In order to attain efficient solutions in software is fundamental to develop solutions that take advantages of interconnection networks properties. We are developing software solutions based on the multicast property of some networks. We also are developing solutions that hide messages latencies based on the messages scheduling method into network.
	Our preliminary performance results appoint to develop a hardware solution like as TTL_PAPERS with direct access without O.S interference and larger bandwidth (32 or 64 bits).
	Our platform consists of clusters of Dual (Quad) Pentium Pro processors interconnected via 100Mb/s Fast Ethernet and (0.64+0.64 Gb/s) Myrinet networks. In order to carry out this research, we are using the MPICH implementation version 1.1 of MPI library. We implement a TTL_PAPERS prototype for 8 nodes and we have included the barrier synchronization and broadcast solutions of TTL_PAPERS as collective operations in MPICH version 1.1.

User Level Communication Libraries

FULL2 - User level communication libraries, such as the VIA standard proposal, are currently the best way to deal with the high overheads imposed by the operating systems communication layers. We are working with a implementation of VIA based on Berkeley's VIA and NERSC's M-VIA, both for LINUX and WinNT OS.

SW - Distributed Shared Memory (DSM)

DSMplus - SW-DSM is a very promising approach to provide shared memory environment to cluster or network of workstations, using basically message passing operations to provide memory consistency. We are evaluating and enhancing the performance of existing SW-DSM, such as Rice Treadmarks and Brazos, both for LINUX and NT OS.

PULSAR - First SW-DSM implemented at LSI-EPUSP.

Hardware Development | System Security and High Speed Networks
Benchmarking | Applications | Other

DELTA | CMagic | CMon | Cluster Administration | PARIS