NEC and Unisys Tag Team on Monster Xeon Server
Two heads, 96 Xeon cores are better than Itanium
After nearly three years of engineering work, Japanese server maker NEC and its partner, American server maker Unisys, will today take the wraps off a co-developed big iron box nicknamed "Monster Xeon" server and sold under their respective brands globally.
The Monster Xeon server is a cell-based symmetric multiprocessor that makes use of the "Dunnington" six-core Xeon 7400 series processor that Intel is announcing today as well.
The design is very similar to the non-uniform memory architecture (NUMA) clustering developed by Sequent more than a decade ago, where you have a group of motherboards (or cells) clustered together to present a single system image to an operating system. (Because of the high-bandwidth of the interconnect, the NUMA-oid architecture looks and smells like an SMP server to an operating system and its applications, even though all processors are not, strictly speaking, connected to all memories and to each other locally on one system board.)
Just about every commercial server I can think of sold today uses a similar design, and IBM is a big fan of this kind of big iron, having happily used this approach in its high-end X64 and Power systems since it acquired Sequent in the summer of 1999.
In the case of the Monster Xeon server from NEC and Unisys, each cell board has four processor sockets, each of which can be equipped with a six-core Dunnington chip. This board fits inside a 4U rack-mounted chassis and anywhere from one to four of these boxes can be linked together as a single system through high-speed optical links that lash the systems together using the 80 GB/sec of bandwidth supplied by the MX chipset at the heart of the Monster Server.
NEC and Unisys are supporting the six-core X7460 Dunnington chip as well as the four-core E7440 chip in the machine, which means it can span up to either 64 or 96 cores, depending on the chip used.
Big Iron jobs
Mike Mitsch, general manager of the Enterprise Server Division at NEC's American unit, says that this is the first X64-based server platform that NEC is putting into the field through that division. ESD creates and sells vector supercomputers, the Itanium-based "Azuza" and "Asama" servers (which bear the Express5800 brand), as well as a cloud-friendly dense servers called Eco Center.
NEC's Client and Server Division makes X64-based Series 3800 rack and tower servers (one- or two-socket boxes) as well as the SigmaBlade blade servers and the 300 Series fault tolerant servers (which are co-developed with Stratus Technologies). The significance of his unit making a big Xeon box is that NEC is saying that Xeon-based machine can take on that kind of big iron jobs.
The NEC variant of the Monster Xeon server has the kind of catchy name that NEC is famous for, and is called the Express5800/A1160. The server supports 667 MHz fully buffered DIMM main memory, and right now 4 GB DIMMs are supported for a maximum of 128 GB per node; when 8GB DIMMs are available in the first quarter of 2009, the Express5800/A1160 will support a maximum of 1 TB of main memory on a fully configured 64-core or 96-core system with 128 total memory slots.
Each node in the Express5800/A1160 has six PCI-Express x8 peripheral slots. The 4U chassis has six drive bays, which hold 2.5-inch SAS drives in 73 GB, 147 GB, and 300 GB capacities, and an integrated RAID disk controller. The machine has two Gigabit Ethernet ports per server node, and each node also has a flash disk option for hypervisors.
Interestingly, the dynamic partitioning feature of Windows Server 2008 was developed by Microsoft and NEC, and is first implemented on the Monster Xeon box. It will be available in the first quarter of next year, says Colin Lacey, vice president of systems and storage for the Systems and Technology group at Unisys. With this feature, a Windows partition can be dynamically reconfigured to add or subtract whole cell boards--meaning it can be changed without shutting the system down or rebooting Windows instances.
Sponsored: Benefits from the lessons learned in HPC