Technical Specifications for the E12k
System Boards, Processors, Cache and Memory:
Technical Specifications for the E10k
System Boards, Processors, Cache and Memory:
There are actually two (2) interconnected 64 processor Sun Microsystems HPC10000 systems in the facility, located side by side in CoRE 635. Also dubbed as "Starfire", "E10k", and "E10000", they represent Sun's high-end multi-processor servers. The Rutgers configuration is the HPC-10000, which consists of an Enterprise 10000 platform with the Sun HPC software installed. Total list price for the two systems is approximately $7 million dollars, although Sun made considerable contributions to the project including the donation of one system at no cost.
Each system is configured with sixty-four 400 MHz SPARC processors (maximum processor configuration for all e-class systems), 32GB of ram, and approximately a terabyte of disk. Each system supports 16 processor boards with 4 processors per board. Each processor utilizes 4Mbytes of L2 cache. Processor boards can be dynamically divided into domains as described below. The E10000 belongs to the category of shared memory parallel machines. All processors in a given domain share each other's memory, and can access data with the same transfer speed. The shared memory model is one of the two most popular computational models today. The other is a distributed memory parallel machines. A group of Personal Computers (or workstations) connected in network (i.e. cluster) is one example of a distributed memory system.
Sun has sold over 1500 E10k class systems within the first two years of the product line, and has become one of the most popular shared memory systems in the industry. It is most popular with Internet providers because the high transactions per second spec required for Web clients. It can achieve over 120 million transactions per day across the Internet. Over thirty E10000s have been sold in New Jersey,
Each Starfire server's backbone is comprised of the Gigaplane-XB™ interconnect, which enables fast, uniform, conflict-free memory access throughout the system. Far superior to bus architectures, the Gigaplane-XB interconnect provides up to 12.8 Gbytes/sec data bandwidth at a low constant latency - less than 500 ns. Given the system's bus speed and L2 cache size, the E10k can achieve up to four to five times the performance of a similar distributed PC cluster with the same number and speed of processors.
Each Enterprise 10000 was delivered with the hardware capability for division into four physical domains. Every Dynamic System Domain remains logically isolated from any other domain in the system, isolating it 100% from any software error or CPU, memory, or I/O error generated by any other domain. Our domain configuration will be as follows:
The peak rating (i.e. advertised) of the combined facility is 100 GFLOPS, with maximum LINPACK performance rated around 90 GFLOPS. In actual LINPACK tests performed on the Rutgers systems (all 128 processors), a rating of 79.36 was obtained. This put the Rutgers facility at ranking 83rd in the June top 500!! In addition, according to SUN publications, the 64 processor E10000 is the top performer on the TPC-C (Transactions Processing Council) benchmarks, which is considered the best benchmark for database performance. It has won the Datamation 1999 Product of year award for IT managers. (For those that insist on MIPs ratings, its rated at 6000 mainframe MIPs. And equally significant for the mechanical engineers, its calculated at 49.25 horsepower)
Each domain has its own 100Mbit Ethernet connection to the rest of CAIP’s Cisco 5500 switched network. In the near future, it will be on RU Net 2000. The type and speed of connection is still under consideration. In addition, there exists a private gigabit fiber link between red and blue used for MPI jobs.
Summary of system resources:
|
|
|
|
|
|
|
|
|
|
128 Gbyte OS |
|
|
|
|
|
Gigabit link to magenta |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|