Existing URCs and GRCs are summarized in the Table 1, which lists their primary characteristics, including number of nodes, processor type and speed, aggregate memory, disk, and tape storage, and internal and external networking bandwidth.
A typical Tier 2 center will contain many of the following elements:
A hundred or more compute nodes. These will typically be Intel Pentium, AMD Athlon, or Compaq Alpha-based workstations or PCs, running the linux operating system.
A private 100 Mbps or Gbps network with a fully meshed 100 Gbps intersection bandwidth dedicated switch.
A tape robot with a capacity of tens to hundreds of Tbytes.
Specialized RAID servers that provide reliable access to very large data sets.
Front-end nodes that control the computational cluster and serve as world-visible portals or entry points.
External network connections that scale up from OC3 and OC12 (2001) over time.
URCs and GRCs vary in their physical characteristics according to the needs of their primary applications. For example, analysis of high-energy experimental data has fairly low bandwidth requirements for the local private network, and compute throughput is largely limited by the speed of integer operations. Thusthe centers for these purposes will probably use 100 Mb/s Ethernet and dual-processor nodes. On the other hand, the analysis of gravitational wave data using frequency-domain techniques uses larger data sets, and is largely limited by the speed of floating-point arithmetic and access to memory. For these centers, single processor systems and Gb/s networking may be more suitable.
Because commodity computer hardware is evolving so rapidly, it is neither possible nor desirable to give a detailed design for centers that will be built more than a year in the future. Optimal choices of CPU type and speed, networking technology, and disk and tape architectures vary with a time scale of a year. Because the needs of the physics experiments drive the deployment of Tier 2 hardware resources, this means that iVDGL will be composed of diverse and heterogeneous systems.
iVDGL Deployment Schedule
Initial iVDGL deployment will target six existing URCs and GRCs, selected based on capabilities available local expertise available. (Caltech, FNAL, U.Chicago, UT Brownsville, UW Madison, and UW Milwaukee are current targets.) These centers will be provide resources for Grid computation and testing.
In the second year, we will include the outreach centers and roughly five of the European centers that have pledged support as detailed in the supplementary documents part of this proposal. Each European center is expected to have equipment similar to what is proposed here, and candidate centers will be integrated on a timeline based again on capabilities and local expertise. If DTF is funded, then these resources will also be made available for periods of peak or burst operation.
In the third year, the number of US and international sites within the iVDGL will increase sharply, to of order fifteen US sites and fifteen European sites, including NSF PACI sites.
In the fourth year, Japan and Australia will be added, as will additional sites in the US.
By the fifth year, we anticipate that there will be of order 100 Tier 2 centers in iVDGL, in addition to the ones funded here.
The proposed initial hardware purchases for the iVDGL compute facilities are summarized in Table 2. The type and scope of later purchases will be determined at a later time.