The NumGRID is a joint Russian-French project of the M2P2 laboratory (UMR 6181 CNRS-Universits d'Aix-Marseille et Ecole Centrale, Marseille) and ICM&MG
SB RAS, Novosibirsk. The NumGRID provides middleware for joining geographically distributed computational clusters in order to run large-scale scientic applications (such a joint computational resource is also referred to as a NumGRID further on). Other important goals are:
- to prolong lifetime of the old clusters by consolidating them with newer systems or with other obsolescent clusters;
- to allow for multi-part applications where each part requires specialized hardware or system software. The NumGRID will join specialized clusters and place parts of the application according to their specic requirements.
The development of the NumGRID started with the following basic requirements:
- clusters should be joined on a basis of common communication layer for the processes located on worker nodes,
- communication layer should be based on MPI  standards specications,
- NumGRID should enable running distributed jobs without major changes to the local cluster administrative policies,
- clusters can be heterogenious (dierent CPU performance, dierent memory capacity, etc),
- each cluster is composed of the head node and a number of worker nodes. Worker nodes are used for running jobs and are connected to each other via high performance network. Also, they are connected to each other and to a head node via less capable TCP/IP private network which is used for shared le system and job control. A head node has another TCP/IP network interface to accept user connections from outer world (internet). A head node is used to compile, queue and monitor jobs.