Compaq, Alpha and High Performance Technical Computing (HPTC) - New German Partnerships

HPTC using Alphas

Dr. Jochen Krebs, well known as an expert in the HPTC field presented the new Alphas. Compaq want to push the sales and invests 500 Mio US$ to expand the market, together with Samsung and API. IBM supports the Alpha with its copper and silicon on isolator (SOI) technology. Actually The Alpha EV6/EV67 delivers 4 integer and 2 floating point results per clock, allows out-of-order execution up to 80 instructions and speculative branch. The external cache bandwidth sums up to 5 GB/s, the memory bandwidth up to 3 GB/s - 2 GB/s sustained. The EV67 processor has a clock of 667/730/750 MHz. Th successor EV68AF has 1 GHz, EV68CB (copper based) 1.25 GHz and the EV69 SOI will have 1.45 GHz and is scheduled for 2002. The same year the EV7 (0.18 micron) will appear with more than 1 GHz. In 2004 we can expect the EV8 in 0.125 micron and more than 1.5 GHz, all the logic is integrated in the processor, and in 2006 the EV9 with 0.1 micron.

In the workstation arena - the building blocks of the clusters - there is the DS series with up to 2, the ES with up to 4 and the GS with up to 32 processors. The GS320 will be extended with the EV7 up to more than 64 processors per SMP. The small systems use the DS10, the DS10L as a rack system (466 MHz) or the AlphaStation XP1000 (500/667 MHz) all with one processor.

Both, the Myrinet and the Quadrics (QSW) SAN allow the clustering of 1000s of processors and nodes. In 2001 QSW plans to cluster 256 nodes with each being a 64 SMPs, resulting in approximately 30 TFlop/s. The improved processor speed will deliver clusters of the same size but 40 TFlop/s peak performance in 2002 and 100 TFlop/s in 2004.

The Partners Offering

As the Myrinet interconnect is less costly than the Quadrics, and the software is open source, the partners chose it for their cluster solution. A packet prize is about 125 000 EURO, consisting of 8 nodes XP1000 rack mounted (667 MHz), delivering nearly 11 GFlop/s, 5 slots (2x64 bit PCI, 2x32 bit PCI 1x shared ISA/PCI), Myrinet highspeed interconnect (1.28 Gbit/s), Fast Ethernet, Compaq Tru64 Unix or Linux and ParaStation with standrad MPI pre-installed.

In the near future the partners will switch to the recently announced Myrinet-2000 standard which offers a data rate of 2+2 Gb/s and a lower latency, boosting the already high performance of their cluster solution. Myrinet-2000 will - in addition to serial copper - also support native Fiber interconnect allowing distances of up to 200 meters between nodes, giving the customer more flexibility where to locate the cluster nodes.

ParTec AG with its ParaStation technology and cluster environment is a spin-off of the University Karlsruhe. They support Alpha workstations, Intel and AMD PCs and the Myrinet interconnect. The operating systems are Tru64 Unix on Alpha and RedHat and SuSE unix on Alpha and Intel/AMD. More than 1000 nodes are supported and interconnected with a fat tree topology. ParTec measured a latency of 25 micro seconds and a throughput of 150 MByte/s. The cluster environment allows process and node management, fault tolerance, load balancing, partitioning and co-scheduling. New features like accounting, batch processing, high-speed IP and checkpoint restart will be added soon.

One example is the ALiCE cluster at the University Wuppertal with 128 DS10 with the Alpha 21264, 600 MHz, 1.2 GFlop/s peak. The total system sums up to 154 GFlop/s peak.