~pvital Blog

Last Friday, October 24th, IBM released the IBM SDK for Multicore Acceleration Version 3.1 (a.k.a Cell SDK 3.1) with support to two different Linux® distributions (RHEL 5.2 and Fedora 9) and in three different package bundles: Product, Developer, and Extras.

As reported here,  about 1 month ago I was replaced to the Cell IDE project, and the IDE is one of the great number of SDK’s packages.

After 46 days away, I returned yesterday to work. On August, 2nd I broke my left arm playing soccer at the IBM Cup and I could not work.

During this time offline, I was moved to the Cell IDE project. Cell IDE is a set of Eclipse plug-ins that integrate the Cell Broadband Engine tool chain and enable rapid building of Cell Broadband Engine applications, and it is included in the IBM SDK for Multicore Acceleration.

This change is good for me, once in my master’s project I’m working with transactional memories applied to the Cell processor and other multicore processors. I guess this new project gives me a lot of opportunity and changes to grow up.

Today a new list of the 500 world’s most powerful supercomputers (Top500) was presented at the 23rd International Supercomputing Conference in Dresden, Germany. This list has a particular issue: the petaflops barrier was broken. The responsible by this fact? Roadrunner, the most powerful supercomputer of the world.

Roadrunner, named after the New Mexico state bird, cost about US $100 million, and was designed and build by IBM at Los Alamos National Laboratory to the Department of Energy’s National Nuclear Security Administration. It is the world’s first hybrid supercomputer, designed to Cell Broadband Engine® works with the Opteron® processors from AMD. A few numbers of Roadrunner:

• connects 6,562 dual-core AMD Opteron® chips as well as 12,240 Cell chips (on IBM Model QS22 blade servers),
• has 98 terabytes of memory,
  
• is housed in 278 refrigerator-sized, IBM BladeCenter® racks occupying 5,200 square feet,
• has 10,000 connections – both Infiniband and Gigabit Ethernet — that required 55 miles of fiber optic cable,
• weighs 500,000 lbs.

But other important thing about Roadrunner is its operating system: a Linux version from Red Hat. All the Top10 most powerful supercomputers have Linux. Four of them have only Linux running: #1 Roadrunner (IBM), #4 Ranger (Sun Microsystems), #5 Jaguar (Cray Inc.), #7 Encanto (SGI), #8 EKA (Hewlett-Packard) and #10 Total’s SGI Altix (SGI). The rest of them have a kind of mixed system (Linux+something): #2 BlueGene/L (IBM), #3 Argonne Blue Gene/P Solution (IBM), #6 JUGENE Blue Gene/P Solution (IBM) and #9 IDRIS Blue Gene/P Solution (IBM) have SLES9+CNK.

The number of supercomputers with Linux have increased by the years. From the 500 supercomputers of the last list, 427 have Linux and 40 have a mixed system with Linux (UNICOS/Linux, CNK/SLES 9, UNICOS/SUSE Linux and UNICOS/lc). See the official numbers here. The first time that a supercomputer with Linux appeared in the Top500 list was on June,1998. The development of the Linux usage by the years can be visualized here.

Hoje uma nova lista dos 500 mais poderosos supercompudores do mundo (Top500) foi apresentada na 23a International Supercomputing Conference em Dresden, Alemanha. Esta lista tem um tópico particular: a barreira do petaflop foi quebrada. O responsável por isso? Roadrunner, o supercomputador mais poderoso do mundo.

Roadrunner, nome de uma ave do estado do Novo Mexico nos EUA, custou cerca de 100 milhões de dólares, e foi projetado e construido pela IBM no Laboratório Nacional Los Alamos para a Departamento de Energia. Ele é o primeiro supercomputador híbrido do mundo, projetado para processador Cell Broadband Engine® trabalhar em conjunto com o processador Opteron® da AMD. Alguns números do Roadrunner:

• conecta 6562 chips dual-core AMD Opteron® bem como 12.240 chips Cell (em servidores blades IBM Modelo QS22),
• tem 98 terabytes de memória,
  
• está em 278 IBM BladeCenter® racks ocupando 483.095 m²t,
• tem 10.000 conexões – ambas Infiniband e Gigabit Ethernet — que requiseram 88,5 Km de cabos de fibra óptica,
• pesa cerca de 227 toneladas.

Mas um fato interessante sobre o Roadrunner é seu sistema operacional: uma versão de Linux da Red Hat. Todos os Top10 supercomputadores mais poderosos usam Linux. Quatro deles tem somente Linux: #1 Roadrunner (IBM), #4 Ranger (Sun Microsystems), #5 Jaguar (Cray Inc.), #7 Encanto (SGI), #8 EKA (Hewlett-Packard) e #10 SGI Altix da Total (SGI). O resto possuem um tipo de sistema misto (Linux+alguma coisa): #2 BlueGene/L (IBM), #3 Argonne Blue Gene/P Solution (IBM), #6 JUGENE Blue Gene/P Solution (IBM) e #9 IDRIS Blue Gene/P Solution (IBM) todos com SLES9+CNK.

O número de supercomputadores com Linux tem crescido através dos anos. Dos 500 supercomputadores da última lista, 427 rodam Linux e 40 tem um sistema misto com Linux (UNICOS/Linux, CNK/SLES 9, UNICOS/SUSE Linux and UNICOS/lc). Veja os números oficiais aqui. A primeira vez que um supercomputador com Linux apareceu na lista Top500 foi em Junho de 1998. O crescimento do uso de Linux através dos anos em supercomputadores pode ser visto aqui.

Update: Curiosamente, Roadrunner é nome americano do personagem Papa-Léguas da animação Papa-Léguas e Coiote. Os nomes dos personagens da animação são baseados em nome de animais reais nativos dos desertos do sudoeste americano, o galo-corredor e o coiote.

Last year I attended the 19th International Symposium on Computer Architecture and High Performance Computing (SBAC-PAD) in Gramado, RS, Brazil. The web page of the event is here.

There were very nice presentations like the “Computational Characteristics of Production Seismic Migration and its Performance on Novel Processor Architectures” by Jairo Panetta (Petróleo Brasileiro SA, Brazil), “Impacts of Multiprocessor Configurations on Workloads in Bioinformatics” by Mauricio Breternitz (Intel Corporation, USA) and the Google Keynote presentation.

The event had many parallel Workshops and the most famous of them are the WSCAD (Workshop em Sistemas Computacionais de Alto Desempenho – Workshop in High Performance Computing Systems). This is a portuguese event, so I think that is not interesting point the presentations here, but there was a good CELL programming short course.

Summarizing the event I can point three things discussed in the keynotes and panels, to solve the problem of the new challenges of parallel and multi-core programming:

1. Use high level programming languages;
2. Use tools developed to do the work of parallelling the code;
3. Raise the level of abstraction of the problem.

These three points of view are defended by Kunle Olukotun (Stanford University), Mauricio Bareternitz (Intel) and Laxmikant (Sanjay) Kale (University of Illinois at Urbana-Champaign), respectively.