Copyright 1999- Jun Makino2009/02 2009/01 2008/12 2008/11 2008/10 2008/09 2008/08 2008/07 2008/06 2008/05 2008/04 2008/03 2008/02 2008/01
What kind of computers do you imagine when you hear the terms ``Supercomputing'' or ``High-Performance Computing''? Cray XT3/4/5? IBM Bluegene? Or a number of rackmount IBM/Intel/AMD servers with Infiniband or some other fast networks?修正後のは
These are certainly the machines on which many of large simulations are done. You can easily find many beautiful computer graphics from such simulations. However, to use these big machines is not the only way to do large scientific calculations.
One alternative is to use ``accelerator'' hardwares. The latest example is the IBM Roadrunner system which started operation in June 2008. It consists of approx 13,000 CELL B.E. processor originally developed for Sony PS3 game console, with double-precision enhancement (PowerXCell 8i). Two CELL processors are mounted on a blade, and two blades are connected to a dual-socket dual-core Opteron blade through a PCI-Express interface. Thus, conceptually, each of four cores of Opteron blade controls one CELL processor. This combination of one Opteron blade and two CELL blades is called ``Tri-blade''
The CELL processor itself consists of one IBM PowerPC processor core and eight simpler processors (SPEs, Synergistic Processor Element), each with two double-precision FMA units. With 3.2GHz clock rate, one CELL processor achieves the theoretical peak speed of 102.4 Gflops for double-precision operation.
What kind of computer do you imagine when you hear the terms "supercomputing" or "high-performance computing?" A Cray XT3/4/5? An IBM BlueGene? Or a number of rack-mounted IBM/Intel/AMD servers with Infiniband or some other fast network? Certainly, these machines do many large simulations, and from such simulations you can easily find numerous beautiful computer graphics. However, these big machines are not the only way to do large scientific calculations. The GRAPE and GRAPE-DR hardware (figures 1 and 2), developed at the Center for Computational Astrophysics, National Astronomical Observatory of Japan, are alternatives to typical supercomputing architecture.
Accelerator hardware is one alternative to the most widely used supercomputer architectures. The latest example is the IBM Roadrunner system ("Science-Based Prediction at LANL" SciDAC Review 4, Summer 2007, p33), which started operation in June 2008. It consists of approximately 13,000 Cell BE processors, originally developed for the Sony PS3 game console, with double-precision enhancement (PowerXCell 8i). Two Cell processors are mounted on a blade, and two blades are connected to a dual-socket, dual-core Opteron blade through a PCI-Express interface. Thus, conceptually, each of the four cores of an Opteron blade controls one Cell processor. This combination of one Opteron blade and two Cell blades is called TriBlade. The Cell processor itself consists of one IBM PowerPC processor core and eight simpler processors-synergistic processor elements, each with two double-precision fused multiply and add units (FMAs). With a 3.2 Hz clock rate, one Cell processor achieves a theoretical peak speed of 102.4 gigaflops (Gflops) for double-precision operation. The entire Roadrunner system consists of 3,060 TriBlades, connected with an x4 DDR Infiniband network.
# ping -c 2 dbs.c.u-tokyo.ac.jp PING dbs.c.u-tokyo.ac.jp (184.108.40.206) 56(84) bytes of data. 64 bytes from www1103.sakura.ne.jp (220.127.116.11): icmp_seq=1 ttl=53 time=14.7 ms 64 bytes from www1103.sakura.ne.jp (18.104.22.168): icmp_seq=2 ttl=53 time=14.1 ms --- dbs.c.u-tokyo.ac.jp ping statistics --- 2 packets transmitted, 2 received, 0% packet loss, time 1001ms rtt min/avg/max/mdev = 14.109/14.425/14.741/0.316 mssakura.ne.jp なんだ、、、
# df -B 1G Filesystem 1G-blocks Used Available Use% Mounted on /dev/mapper/sshgw-root 7 1 6 14% / tmpfs 2 0 2 0% /lib/init/rw udev 1 1 1 1% /dev tmpfs 2 0 2 0% /dev/shm /dev/sda1 1 1 1 7% /boot /dev/mapper/sshgw-home 214 67 137 33% /home cfcafs04.cfca.nao.ac.jp:/mnt/raid02 7451 1 7078 1% /misc/work042 cfcafs03.cfca.nao.ac.jp:/mnt/raid01 8382 868 7095 11% /misc/work031 cfcafs03.cfca.nao.ac.jp:/mnt/raid02 8382 858 7106 11% /misc/work032 cfcafs03.cfca.nao.ac.jp:/mnt/raid00 8252 181 7652 3% /misc/work030 cfcafs04.cfca.nao.ac.jp:/mnt/raid01 7451 1 7078 1% /misc/work041 cfcafs04.cfca.nao.ac.jp:/mnt/raid00 7451 4 7075 1% /misc/work040それなりに増強、みたいな。 1T 8 台でも GiB 単位だと 7451 GiB しかない。