Farms/Clusters of the future

1
Farms/Clusters of the future
Large Clusters O(1000), any existing examples
? Yes Supercomputing, PC clusters LLNL, Los
Alamos, Google No long term experience (except
google) ASCI Installation Powerstation (xMW),
Well organized recover management (see Eniac)
2
Farms/Clusters of the future
Reliability By Martin H. Weik, 1961 Ordnance
Ballistic Research Laboratories, Aberdeen Proving
Ground, MD The ENIAC's first few years at the
Aberdeen Proving Ground were difficult ones for
the operating and maintenance crews. The
result was a huge preventive-maintenance and
testing program, which, in the end, led to some
major modifications of the system. Tubes were
life-tested, and statistical data on the failures
were compiled. This information led to many
improvements in vacuum tubes themselves.
Procurement of large quantities of improved,
reliable tubes, however, became a difficult
problem. Power-line fluctuations and power
failures made continuous operation directly off
transformer mains an impossibility. The
substantial quantity of heat which had to be
dissipated into the warm, humid Aberdeen
atmosphere created a heat-removal problem of
major proportions. Down times were long
error-free running periods were short.
3
Farms/Clusters of the future
Avoid local disks ?
4
Future CPUs - Low Power (mobile ?) ?
5
Architecture Results Results
Architecture Amanda Simulation Amanda Reconstruction
Host oceanide1, new desktop PC, 533 MHz FSB Pentium4, 2.40GHz, 512 KB Cache, 256 MB Memory 1625.460u, 10.530s 138.680u, 0.140s
Host minerva, theory PC, 400 MHz FSB XEON-P4, 2.00GHz, 512 KB Cache, 2 GB Memory 1968.150u, 4.620s 167.110u, 0.120s
Host euterpe, network services Pentium III (Tualatin), 1266MHz, 512 KB Cache, 1 GB Memory 3158.580u, 9.370s 154.830u, 0.740s
Host ice53, farm node Pentium III (Coppermine), 800 MHz, 256 KB cache 512 MB 4889.670u, 17.010s 250.290u, 0.310s
6
LQCD - Dirac Operator Benchmark (SSE) 16x163,
single P4/XEON CPU
Intensive cache pre-fetch
No benefit from cache
Dirac operator
Linear Algebra
MFLOPS
7
Future Cluster/Farm Architectures - Blade
Servers ?
NEXCOM Low voltage blade server 200 low voltage
Intel XEON CPUs (1.6 GHz 30W) in a 42U
Rack Integrated Gbit Ethernet network
Mellanox Infiniband blade server
Single XEON Blades connected via a 10 Gbit (4X)
Infiniband network
MEGWARE, NCSA, Ohio State University
8
Intel Centrino Mobile Technology Intel 855GM
chipset
Intel Pentium M Processor
VID
IMVP IV
PSI
TPVDVO
DDR memory
DVO (2 ports)
LVDS
Intel 855GM
Hub Interface
Intel Pro/Wireless Network Connection
Integrated LAN
ICH4-M
PCI 33 MHz
Cardbus
Modem
AC97 2.3
9
Benchmark list hyper threading vs non hyper
threading
Architecture Results Results Results Results
Architecture Amanda Simulation Amanda Reconstruction Theorie form3
Host cube3, farm node, 400 MHz FSB XEON-P4, 2.4 GHz, 512 KB cache, 2 GB Memeory Non hyper-threading kernel 1626.200u 1.480s 131.250u 0.320s Time 61.33 sec Generated terms 35999900
Host cube8, farm node, 400 MHz FSB XEON-P4, 2.4 GHz, 512 KB cache, 2 GB Memeory Hyper-threading kernel 1617.420u 1.530s 154.210u 0.430s Time 62.52 sec Generated terms 35999900