Date: Wed, 22 Mar 2000 10:55:23 +0100
Subject: Gros cluster a UNM de IBM's Netfinity's sous Linux. Succes a
  terme de cette strategie predite pour le HPC.

From Computergram

+ UNM to Build Big Linux Supercluster

By Timothy Prickett Morgan 

While vendors, analysts and customers debate the ultimate 
applicability of Linux to their data center operations, 
everybody in the high-performance supercomputer community seems 
to agree that Unix is living on borrowed time in the HPC arena. 
Conventional Unix-based supercomputers offer higher bandwidth, 
bigger memory blocks and more reliability than supercomputers 
based on PC server components, but the combination of 
lower-price components and open community software is making 
Linux superclusters not only a viable alternative, but 
increasingly a preferred one for academic and industrial 
supercomputing. 

Today's announcement by the University of New Mexico that it is 
building a Linux cluster using IBM's Netfinity thin servers 
that will be the world's fastest Linux HPC box and Number 24 on 
the Top500 supercomputer list lends credence to the view that 
Linux will eventually rule HPC and that it won't be too long 
before the world's fastest supercomputer will be a Linux 
machine running on IA-64 processors. 

The Linux cluster being built by UNM, which is nicknamed Los 
Lobos, is comprised of 256 of the two-way Netfinity 4500R thin 
servers. Netfinity 4500R thin servers appear to be one and the 
same as the forthcoming "Illinois" and "Assault" servers that 
IBM is set to ship in a few weeks (CI No 3,863). Illinois is a 
3U, two-way server that offers more I/O capacity in terms of 
multiple PCI buses as well as more PCI slots, plus more memory 
capacity and more disk capacity than the current 4000R 
"Intimidator" thin server. Illinois also includes hot swap 
disks and various redundant components to increase the 
availability of individual units. Assault will offer more I/O, 
memory and disk capacity than Illinois as well as ActivePCI and 
integrated chipkill memory. Illinois and Assault thin servers 
are expected to support the latest Intel processors running at 
733MHz or even 800MHz. UNM is getting 4500R servers equipped 
with 733MHz Pentium IIIs, 256 of them, in fact, for a total of 
512 processors. 

Back in December, IBM had said that it was putting together a 
deal for a 512-node follow-on to UNM's 64-way "Road Runner" 
cluster that would use four-way Intel PC server motherboards 
for a total of 2,048 processors (CI No 3,812). Someone got 
their wires crossed in speaking about it and probably meant 
that IBM would propose a 512-processor cluster using either 
two-way or four-way Intel motherboards. The two-way Netfinities 
thin servers offer a much smaller form factor than the four-way 
servers, which have memory, disk and I/O capacities that are 
overkill at least as far as parallel computing is concerned. 
Processor clock speed and keeping costs low is much more 
important than having a large number of processors in an SMP 
configuration because with parallel computing you can always 
just add more processors. That's why Los Lobos is built using 
what appears to be the Illinois thin servers. We reckoned last 
December that a 2,048 processor Linux cluster with 512 nodes 
would have about 1.2 teraflops of peak number-crunching 
capacity and about 900 gigaflops running real Fortran and C 
applications. The 256-node, 512-processor Los Lobos cluster is 
estimated by IBM to provide about 375 gigaflops of peak power 
and about 285 gigaflops where the compiler hits the chip. 

Los Lobos will be configured with 1Tb of disk capacity, 1.5Tb 
of tape capacity, 512Mb of main memory per Pentium III 
processor and 256Kb of L2 cache memory per processor. These 
memory specs are very modest, and that is one of the reasons 
why the National Science Foundation only had to kick in $1.5m 
to help build the Los Lobos cluster. The total cost is more 
than that, and UNM and IBM is not saying what the total price 
tag for the Linux cluster will be when it goes operational this 
summer. It will very likely be in the $2.5m range, which comes 
to about $10,000 per node. 

Frank Gilfeather, executive director of HPC at the university, 
says that the real compelling argument for Linux clusters is 
not just their low cost - which tends to be anywhere from a 
third to half of the cost of a conventional supercomputer - but 
the openness of Linux. Researchers in various fields can have 
Linux running on their desktops, use open community code, tweak 
it for their needs, test it on a local baby cluster and then 
run their jobs on the production superclusters like Los Lobos, 
all without leaving the Linux environment. When you add those 
two factors together, and multiply it by the fact that 
university and commercial researchers do not want to specialize 
in operating systems, interoperability or performance tuning - 
they just want to run their applications on the fastest 
computers they can afford - it isn't hard to see that Linux 
will very likely take over in HPC long before it does in the 
commercial world at large. 

The industry lowdown has it that within the next 12 to 18 
months, a Linux hypercluster using Intel's IA-64 processors 
will unseat Intel's ASCI Red (the current number one on the Top 
500 list) and IBM's ASCI White (projected to be number one by 
2001) supercomputers as the fastest production supercomputer in 
the world.