View allAll Photos Tagged Datacenters
The fibers are being fed from the spools on the floor to the blower. (Le fibre passano dalle bobine al soffiatore).
Nobreaks.
Tubos de aço trefilados com e sem costura, peças tubulares acabadas e semi-acabadas e corte a laser
Site: www.engetref.com.br/
Email: engetref@engetref.com.br
Telefone: (11) 4066-1611
The fibers are being positioned to be spliced. On the right, the heat shrink protection tubing on an already spliced fiber is being heated. (Macchina per la giunzione della fibra ottica. Questa sofisticata apparecchiatura controlla elettronicamente la bontà della giunzione).
Telecomunicações.
Tubos de aço trefilados com e sem costura, peças tubulares acabadas e semi-acabadas e corte a laser
Site: www.engetref.com.br/
Email: engetref@engetref.com.br
Telefone: (11) 4066-1611
Spotted at a new datacenter: an interesting, custom SPARC-based computing system for very large scientific workloads — this rack is part of a computer cluster that can solve a system of linear equations with more than ten million variables.
Each SPARC CPU is a 8-core chip clocked at 2GHz, and each core has 256 (!) double precision floating-point registers and four multiply-add units. That number of FP registers is sufficient to compute a 8x8 matrix multiplication without requiring any access to RAM beyond the initial loading and final storing of the FP data. Accesses to the "slow" L1, L2 caches and RAM are thus minimized, allowing the CPU to crunch numbers at high speed.
Operations on large matrixes can be efficiently divided e.g. into 8x8 block decompositions that fit in the register file.
Each multiply-add unit can output on each clock cycle the result of an operation of the form D := A * B + C where A, B and C are double precision FP numbers.
The SPARC CPU's maximum FP throughput is thus 2GHz * 8 cores * 4 fused mutiply-adds = 128 GFLOPs/CPU. Each SPARC CPU has a memory bandwidth of 64GBytes/s.
A SPARC CPU, together with 16GB of RAM and an Interconnect Controller (ICC), form a unified "compute node".
The ICC combines, on a single VLSI, four 5GBytes/s DMA interfaces and a crossbar switch / router with ten 5GBytes/s bidirectional links. These ten links connect to other compute nodes, forming a virtual 6D fused torus / mesh network structure.
Compute nodes can access the memory of other nodes using virtual addressing, as a remote DMA operation. The ICC of the destination node performs the required virtual to physical address translation and the actual DMA. The ICC can also perform simple arithmetic operations on integers and FP data, enabling the parallel computation by the communication fabric itself of barrier operations, without having to involve the SPARC CPU.
Four compute nodes are integrated on each system board, and each rack holds 24 hot swappable system boards.
The picture shows the upper twelve system boards in a rack. Also visible are the nine air-cooled, redundant power supply units, the six I/O controller units, as well as two blade-like, redundant rack supervisor controllers and a Fujitsu storage array containing the operating system boot disks.
The six I/O controller units are water-cooled, and each contains one unified compute node. These I/O controllers connect the rack to other racks and to a high-speed clustered local storage system with a capacity of about 11 petabytes, and a global file system of about 30 PBytes. The operating system of the unified compute nodes is a custom fault-resilient multi-core Linux kernel; the mass storage system is based on Lustre.
The peak FP performance of each rack is 128 GFLOPs/compute node * (4 compute nodes / system board * 24 system boards + 6 I/O controller compute nodes ) / rack = 128GFLOPs * (4*24+6) = 13056 GFLOPs, or 13.056 TeraFLOPs; the total memory size per rack is 1632 Gigabytes.
Each rack requires about 10KW of electrical power, and the high-speed 6D torus inter-node connection fabric has been designed to efficiently extend to hundreds of such racks. Beware that electricity bill...
In this data center, a cluster of 864 of these racks form a massive parallel supercomputer, with 1400 Terabytes of RAM, and a theoretical peak FP performance of 13.056 TFLOPs * 864 = 11.280 PetaFLOPs — i.e. more than eleven million gigaFLOPs.
The effective LINPACK performance is about 93% of that theoretical peak.
The main intended application area seems to be the life sciences, with an emphasis on molecular modelling ab initio — simulating complete molecules starting from the quantum behavior of elementary nucleons and electrons — to assist the design of new drugs, simulate biochemical processes like chemotherapy agent resistance of cancer cells at the molecular level, model neural processes etc.
Climate modelling, atomic level simulation of novel nanomaterials and computational fluid dynamics applications are also in the input queue.
Exaustão.
Tubos de aço trefilados com e sem costura, peças tubulares acabadas e semi-acabadas e corte a laser
Site: www.engetref.com.br/
Email: engetref@engetref.com.br
Telefone: (11) 4066-1611