Knocking on the doors of the Petaflop/s Erfahrungen mit 31 Listen und Ausblick auf die nächsten Jahre mit den TOP500 Hans Werner Meuer hans@meuer.de Universität.

Knocking on the doors of the Petaflop/s Erfahrungen mit 31 Listen und Ausblick auf die nächsten Jahre mit den TOP500 Hans Werner Meuer Universität Mannheim & Prometeus GmbH 4th Erlangen High-End-Computing Symposium, Monday, June 23, 2008

31th List: The TOP10 Manufacturer Computer Rmax [TF/s]
Installation Site Country Power [MW] #Cores 1 IBM Roadrunner BladeCenter QS22/LS21 1026 DOE/NNSA/LANL USA 2.35 122,400 2 BlueGene/L eServer Blue Gene Solution 478.2 DOE/NNSA/LLNL 2.33 212,992 3 Intrepid Blue Gene/P Solution 450.3 DOE/ANL 1.26 163,840 4 Sun Ranger SunBlade x6420 326 TACC 2.00 62,976 5 Cray Jaguar Cray XT4 QuadCore 205 DOE/ORNL 1.58 30,976 6 JUGENE 180 Forschungszentrum Juelich (FZJ) Germany 0.50 65,536 7 SGI Encanto SGI Altix ICE 8200 133.2 New Mexico Computing Applications Center 0.86 14,336 8 HP EKA Cluster Platform 3000 BL460c 132.8 Computational Research Laboratories, TATA SONS India 1.60 14,384 9 112.5 IDRIS France 0.32 40,960 10 SGI Altix ICE 8200EX 106.1 Total Exploration Production 0.44 10,240

Outline Mannheim Supercomputer Statistics & Top500 Project Start in 1993 Competition between Manufacturers, Countries and Sites My Supercomputer Favorite in the Top500 Lists The 31st List as of June 2008 Performance Development and Projection Bell‘s Law Supercomputing, quo vadis? Top500, quo vadis? - My first topic are the so called Mannheim Supercomputer Statistics introduced by me at the very firtst Mannheim Supercomputer Conference in 1986 before we started with the TOP500 Project in I will present to you the results but also the deficiencies of these statistics which has been the reason to start with the TOP500 project in I will present to you the rationals and people behind this project. - Secondly I will focus on an extremely important topic in my point of view: COMPETITION. The TOP500 are so successful mainly because they promote and encourage COMPETITION: between manufacturers ,between countries and sites. - Then I am going to discuss with you my favorite Supercomputer out of ! -Our last list was published at tSC In Tampa last Novemer and I will give you a couple of results - In the next section we will focus on performance development and even performance projections - Finally I will address shortly our future plans

Mannheim Supercomputer Statistics & Top500 Project Start in 1993
Presented at ISC’86 – ISC’92/Mannheim Supercomputer Seminars Counting of Supercomputers in the World 530 systems in the year 1992 I started the so called Mannheim Supercomputer Statistics at our very first Mannheim Supercomputer Conference in 1986. - Our approach was more than simple: We just counted the Supercomputers (exclusively Vector Computers) in the different continents by the help of the manufacturers . - in 1992 we ended up with our 7th version of the Mannheim Statistics with 530 systems in the world.

Manufacturer Share 100% NEC Hitachi 80% Fujitsu 60% CDC Cray 40% Here ist the result of the Mannheim statistics as far as the Manufacturers are concerned: -Cray is the clear leader with 60% share - CDC with a bit less than 10% until they went out of Supercomputer businees and - the three Japanese vector computer manufacturers Fujitsu, Hitachi and NEC 20% 0% 1986 1987 1988 1989 1990 1991 1992

Countries Share 100% Others 90% Others Europe 80% Germany 70% 60% France 50% UK 40% Japan 30% USA And here is the share of systems per country: - it seems to be a competition betweem Americas and Japan - The US has a decreasing portion , Japan an increasing portion - in 1992 we counted appr. 35 % system in the US but 40% in Japan. Please remember when looking at the 1st Top500 list 20% 10% 0% 1986 1987 1988 1989 1990 1991 1992

Deficits of the Mannheim Supercomputer Statistics Not very reliable Vector Supercomputers only – „numerical“/ mainly „scientific and engineering“ applications Definition of „Supercomputer“ necessary The deficits of the Mannheim Supercomputer Statistics are obvious: - they have been not very reliable / data in Germany and also in Europe could be checked by us more or less easily but not in the US and especially not in Japan with a lot of entry level systems like VP50 /VP 30 from Fujitsu. Is such an entry level system a supercomputer ? Is a Convex C1 a supercomputer ? Or a Cray 1 in the Deutsches Museum/ German Museum ? - at the beginning of the nineties of course Vector Computer were not alone but already massively parallel systems from e.g. TMC - what we really needed was the definition of SUPERCOMPUTER

Top500 Procedure Listing the 500 most powerful computers in the world Yardstick: Rmax of Linpack Solve Ax=b, dense problem, matrix is random Update twice a year: • ISC’xy in June in Germany • SC’xy in November in the U.S. All information available from the TOP500 webserver at: And here is the TOP500 Procedure we have introduced in spring 1993: - Listing of the 500 most powerful computers in the world < OK; why 500 and not 600/ well we ended with 530 systems in 1992 with Mannheim Supercomputer Statistics> - 'most powerful' was defined by a benchmark and we chose the best linpack performance Rmax <OK why linpack? we did this not since linpack was the best but it was the only one at this time and available for all systems in question> - Update this list biannually in June in Germany at ISC and in November at SC in the US < of course other solutions are possible e.g. an update continously > I will discuss pros and cons of Linpack in the final section of this talk

Top500 Authors Project was started in spring 1993 by: Hans W. Meuer Erich Strohmaier Jack Dongarra Authors since: O6/1993 Horst Simon 11/2000 The project was started by my young colleague Erich Strohmaier and by me at the University of Mannheim in Germany. I had engaged Erich after finishing his PhD in Theoretical Physics at Heidelberg University in 1990. - Erich and I wanted to make the TOP500 Project a success. But I myself was convinced that it was impossible for two German greenhorns to make the project a success. - What we needed was a worldwide renowned man in the HPC field. An American would best suit our purpose of course.I asked the ideal man in the field, Jack Dongarra from the University of Tennessee. Since we had decided in Mannheim for linpack as the yardstick for our ranking, Jack as the father of linpack accepted to become the third author. - Our fourth author joining our team in 2000 was Horst Simon from LBNL.

Dongarra's TOP500 List of World's Fastest Supercomputers Released at Mannheim Conference „CRPC researcher Jack Dongarra of the University of Tennessee and Oak Ridge National Laboratory is one of three renowned computer scientists who assemble the legendary TOP500 List of the world's fastest supercomputers. Released twice a year since 1993, the list features sites with the most powerful computer systems, determined with information from a questionnaire sent to high-performance computer (HPC) experts, computational scientists, manufacturers, and the Internet community at large. Dongarra, Hans Meuer of the University of Mannheim, and Erich Strohmaier of the University of Tennessee released their June 1999 TOP500 List at the 14th Mannheim Supercomputing Conference and Seminar, held June in Mannheim, Germany.“ Source: The Center for Research on Parallel Computation at Rice University To do something together with a famous man might be dangerous, not for the famous man but for the underdogs so to say. Here is the proof for my statement Very soon in the US the TOP500 list was Jack Dongarra's list not Erich's and my list. OK, that's life!

Outline Mannheim Supercomputer Statistics & Top500 Project Start in 1993  Competition between Manufacturers, Countries and Sites My Supercomputer Favorite in the Top500 Lists The 31st List as of June 2008 Performance Development and Projection Bell‘s Law Supercomputing, quo vadis? Top500, quo vadis? - My first topic are the so called Mannheim Supercomputer Statistics introduced by me at the very firtst Mannheim Supercomputer Conference in 1986 before we started with the TOP500 Project in I will present to you the results but also the deficiencies of these statistics which has been the reason to start with the TOP500 project in I will present to you the rationals and people behind this project. - Secondly I will focus on an extremely important topic in my point of view: COMPETITION. The TOP500 are so successful mainly because they promote and encourage COMPETITION: between manufacturers ,between countries and sites. - Then I am going to discuss with you my favorite Supercomputer out of ! -Our last list was published at tSC In Tampa last Novemer and I will give you a couple of results - In the next section we will focus on performance development and even performance projections - Finally I will address shortly our future plans

Top500 Status Competition between Manufacturers, Countries and Sites
1st Top500 List in June 1993 at ISC’93 in Mannheim • 30th Top500 List on November 13, 2007 at SC07 in Reno 31st Top500 List on June 18, 2008 at ISC’08 in Dresden 32nd Top500 List on November 18, 2008 at SC08 in Austin 33rd Top500 List on June 24, 2009 at ISC’09 in Hamburg Acknowledged by HPC-users, manufacturers and media And my first slide in this section gives you the TOP500 status.

Competition between Manufacturers, Countries and Sites
1st List as of 06/1993 31st List as of 06/2008 Countries Count Share % USA 225 45.00 % Japan 111 22.20 % Germany 59 11.80 % France 26 5.20 % United Kingdom 25 5.00 % Australia 9 1.80 % Italy 6 1.20 % Netherlands Switzerland 4 0.80 % Canada 3 0.60 % Denmark Korea Others 20 4.00 % Totals 500 100 % 100 % 500 Totals 9.80 % 49 Others 1.20 % 6 India 2.40 % 12 China 0.40 % 2 Korea — Denmark Canada Switzerland 1.00 % 5 Netherlands Italy 0.20 % 1 Australia 10.60 % 53 United Kingdom 6.80 % 34 France 9.20 % 46 Germany 4.40 % 22 Japan 51.40% 257 USA Share % Count Countries Our 7th Mannheim Supercomputer Statistics indicated that the US and Japan have a competition with a neck-and-neck race. But the 1st TOP500 list demonstrated that the rumor about the Japanese danger was overstated; since the US were leading clearly with 45% of all TOP500 installations versus just a share of 22 % of Japan,i.e. Japan was far behind with just half of the US installations. If we look at the latest list as of November 2006 and published at the SC conference in Tampa we realize that the dominance of Americas is even much bigger than 14 years ago: - Now they have a share of 62 % of the systems installed and Japan has just one tenth of this share, i.e. 6 % as UK in Europe also has. You certainly all here agree that something must be done to change this really horrible situation.

Countries/Systems If we look at the different countries involved in HPC we have the following slide: - USA increasing - Japan decreasing - Germany decreasing - UK increasing slightly

1st List as of 06/1993 31st List as of 06/2008 Manufacturers Count Share % Cray Research 205 41.00 % Fujitsu 69 13.80 % Thinking Machines 54 10.80 % Intel 44 8.80 % Convex 36 7.20 % NEC 32 6.40 % Kendall Square Research 21 4.20 % MasPar 18 3.60 % Meiko 9 1.80 % Hitachi 6 1.20 % Parsytec 3 0.60 % nCube Totals 500 100 % 7.00 % 35 Others 100 % 500 Totals 5.40 % 27 Dell 4.40 % 22 SGI 42.00 % 210 IBM — nCube Parsytec 0.20 % 1 Hitachi/Fujitsu Meiko MasPar Kendall Square Research NEC 36.60% 183 Hewlett-Packard Intel Thinking Machines 0.80 % 4 Fujitsu 3.20 % 16 Cray Inc. Share % Count Manufacturers If we focus at the manufacturers Cray Research was the clear leader with 41 % of the share, ashead of Fujitsu with 14 %. At position 3 already TMC is showing up 11 % of the Share ahead of Intel with 9 %. At this time Intel stil has had its Supercomputer Division. Surprisingly IBM was not represented in the 1. TOP500 list at all. In the 28th list IBM has the clear lead with appr. half of the share, exactly 47 %. The second position is held by Hewlett Packard... ad Cray Inc. is down with only 3 % of the share.

Manufacturer/Systems The HPC market is extremely dynamic: in only 14 years and 28 lists we have experienced a complete transformation of this HPC market, as shown in this slide.

Top Sites through 31 Lists Rank Site Country % over time 1 Lawrence Livermore National Laboratory United States 5.4 2 Sandia National Laboratories 3.63 3 Los Alamos National Laboratory 3.59 4 Government 3.23 5 The Earth Simulator Center Japan 1.93 6 National Aerospace Laboratory of Japan 1.65 7 Oak Ridge National Laboratory 1.43 8 NCSA 1.3 9 NASA/Ames Research Center/NAS 1.24 10 NERSC/LBNL 1.18 11 University of Tokyo 12 Pittsburgh Supercomputing Center 1.11 13 Semiconductor Company (C) 1.1 14 Naval Oceanographic Office (NAVOCEANO) 1.06 15 ECMWF United Kingdom 1.02 16 ERDC MSRC 0.91 17 Forschungszentrum Juelich (FZJ) Germany 0.87 18 IBM Thomas J. Watson Research Center 0.86 19 Japan Atomic Energy Research Institute 0.8 20 Minnesota Supercomputer Center 0.72 In this slide we list the 20 most powerful sites in 28 TOP500 Lists. The percentage in the right column indicates which portion of the Rmax sum of all systems in a list the site has in the average list of 28.

Outline Mannheim Supercomputer Statistics & Top500 Project Start in 1993 Competition between Manufacturers, Countries and Sites  My Supercomputer Favorite in the Top500 Lists The 31st List as of June 2008 Performance Development and Projection Bell‘s Law Supercomputing, quo vadis? Top500, quo vadis? - My first topic are the so called Mannheim Supercomputer Statistics introduced by me at the very firtst Mannheim Supercomputer Conference in 1986 before we started with the TOP500 Project in I will present to you the results but also the deficiencies of these statistics which has been the reason to start with the TOP500 project in I will present to you the rationals and people behind this project. - Secondly I will focus on an extremely important topic in my point of view: COMPETITION. The TOP500 are so successful mainly because they promote and encourage COMPETITION: between manufacturers ,between countries and sites. - Then I am going to discuss with you my favorite Supercomputer out of ! -Our last list was published at tSC In Tampa last Novemer and I will give you a couple of results - In the next section we will focus on performance development and even performance projections - Finally I will address shortly our future plans

My Supercomputer Favorite in the Top500 Lists
Top10 in the 06/1997 list Manufacturer Computer Rmax [GF/s] Installation Site Country Year #Proc 1 Intel ASCI Red 1068 Sandia National Laboratories USA 1996 7264 2 Hitachi CP-PACS/2048 368.2 Center for Computational Science Japan 2048 3 Fujitsu Numerical Wind Tunnel 229 National Aerospace Laboratory 167 4 SR2201/1024 220.4 University of Tokyo 1024 5 Cray T3E 176 Forschungszentrum Jülich Germany 512 6 Government 1997 7 Max-Planck-Gesell. MPI/IPP 8 NASA/Goddard Space Flight Center 9 Pittsburgh Supercomp. Center 10 University Stuttgart And here it is: Intel ASCI Red the first Linpack Teraflops System in the world entering our list at position first of course in the 9th TOP500 list published at SC'97 in Mannheim. This system was the last Supercomputer built by Intel. They did it for the DOE - Department of Energy- at a tme they had already closed their Supercomputer division. ASCI Red was installed at Sandia National Laoratories and the acceptance test included that ASCI Red was capable of performing the Linpack benchmark above 1 Teraflops.

My Supercomputer Favorite in the Top500 Lists
Performance [GFlop/s] 0,1 1 10 100 1000 10000 100000 Jun 93 Jun 94 Jun 95 Jun 96 Jun 97 Jun 98 Jun 99 Jun 00 Jun 01 Jun 02 Jun 03 Jun 04 Jun 05 N=1 N=500 Sum 1 TFlop/s 1 PFlop/s Intel ASCI Red Jun 05 Let's come back to our 12 th Supercomputer Conference in Mannheim where in the Opening Session I was introducing our 9th TOP500 list with the number one spot Intel ASCI Red the first Teraflops Computer in the World. Click. Based on the the current and already measured TOP500 data which cover the last 4 1/2 years and under the assuption that the current performance development will continue for some time to come we can now extrapolate the observed performance values using linear regression on the logarithmic scale. This means we fit exponential growth to these three levels of performance . Based on the extrapolation of these fits we could claim the following: In the year 2005, exactly in June, there are Teraflops Systems only in the TOP500. Many of our attendants at this time believed that Meuer is becoming grazy... with a forecast over exactly 8 years in such a dynamic area ! Well, I have to admit that my feeling was also not too comfortable at this time since I knew that nothing is harder to forecast than the future.

Number of Teraflop/s Systems in the Top500
My Supercomputer Favorite in the Top500 Lists Number of Teraflop/s Systems in the Top500 500 100 10 1 You certainly can imagine that I was expecting the next TOP500 lists with extreme curiousity. - Nov 97 nothing, still ASCI Red the only one - June " " - Nov " " I became really nervous ! - June 99 thanks heaven, ASCI Blue Mountain from SGI entered the scene at Los Alamos National Laboratory. It was the second system on our planet making the Linpack Teraflops barrier and it was ranked #2 in the 06/99 list with 1. 6 Teraflops since ASCI Read now was slightly above 2 Teraflops since they managed to increase the number of processors to more than 9000. - now we experieced movement as far as Teraflops systems are concerned in the list.... - in June 2005 as forecasted we have had Teraflops systems only in the list. Jun-97 Jun-98 Jun-99 Jun-00 Jun-01 Jun-02 Jun-03 Jun-04 Jun-05

My Supercomputer Favorite
in the Top500 Lists

Outline Mannheim Supercomputer Statistics & Top500 Project Start in 1993 Competition between Manufacturers, Countries and Sites My Supercomputer Favorite in the Top500 Lists  The 31st List as of June 2008 Performance Development and Projection Bell‘s Law Supercomputing, quo vadis? Top500, quo vadis? - My first topic are the so called Mannheim Supercomputer Statistics introduced by me at the very firtst Mannheim Supercomputer Conference in 1986 before we started with the TOP500 Project in I will present to you the results but also the deficiencies of these statistics which has been the reason to start with the TOP500 project in I will present to you the rationals and people behind this project. - Secondly I will focus on an extremely important topic in my point of view: COMPETITION. The TOP500 are so successful mainly because they promote and encourage COMPETITION: between manufacturers ,between countries and sites. - Then I am going to discuss with you my favorite Supercomputer out of ! -Our last list was published at tSC In Tampa last Novemer and I will give you a couple of results - In the next section we will focus on performance development and even performance projections - Finally I will address shortly our future plans

The 31st List as of June 2008 Processor Architecture/Systems

The 31st List as of June 2008 Operating Systems/Systems

The 31st List as of June 2008 Processor Generations/Systems (June 2008) Others 12% POWER5+ 2% Xeon E54xx (Harpertown) 24% PowerPC 450 2% Opteron Quad Core 3% PowerPC 440 4% Xeon X54xx (Harpertown) 8% Xeon 53xx (Clovertown) 19% Opteron Dual Core 8% Xeon 51xx (Woodcrest) 18%

The 31st List as of June 2008 Interconnect Family/Systems

The 31st List as of June 2008 Architectures / Systems

Outline Mannheim Supercomputer Statistics & Top500 Project Start in 1993 Competition between Manufacturers, Countries and Sites My Supercomputer Favorite in the Top500 Lists The 31st List as of June 2008  Performance Development and Projection Bell‘s Law Supercomputing, quo vadis? Top500, quo vadis? - My first topic are the so called Mannheim Supercomputer Statistics introduced by me at the very firtst Mannheim Supercomputer Conference in 1986 before we started with the TOP500 Project in I will present to you the results but also the deficiencies of these statistics which has been the reason to start with the TOP500 project in I will present to you the rationals and people behind this project. - Secondly I will focus on an extremely important topic in my point of view: COMPETITION. The TOP500 are so successful mainly because they promote and encourage COMPETITION: between manufacturers ,between countries and sites. - Then I am going to discuss with you my favorite Supercomputer out of ! -Our last list was published at tSC In Tampa last Novemer and I will give you a couple of results - In the next section we will focus on performance development and even performance projections - Finally I will address shortly our future plans

Performance Development and Projection
Jack‘s Notebook - Here we have plotted the performance overe the last 14 years at position (entry level performance), at the leading position #1 as well as the total accumulated performance, the sum. - it is not difficult to realize that all thes figures show an exponential growth, the scale left hand side is a logarithmic scale. - If we compare these grow rates with Moore's law we have the following: Moore's laww states for Microprocessors a doubling of performance in 18 months. The growth here is larger, for the summ we have a doubling in 13.9 years, for position #1 a doubling in 13 months and for position #500 already in months. - The reason for this larger performance growth comes from two dimensions: processor performance and number of processors used. - Btw, the curves at position #1 and #500 are quite different: at #1 we typically see a step function, once a system has made the number one spot it remains there for a couple of TOP500 editions: That was true for NWT, Intel ASCI Red, and so on, and for ES being the #1 from 06/ /04. < Jack: We have a Computenik on hand>. Also the current #1, IBM BlueGene/L is holding that position in different stages of expansion since 5 lists. - <Click> For ASCI Red - holding the #1 position in 7 lists it took 6 years to move from #1 to #500. <Cick><Click> Currently Jack's Notebook - he is always operating the most powerful notebook of all the TOP500 authors- has a performance of appr. 10 Gflops. Jack‘s Notebook

Performance Development and Projection
Performance Projection 1 Pflop/s Jack‘s Notebook 6-8 years Jack‘s Notebook 8-10 years Jack‘s Notebook - Again as already used when discussing ASCI Intel Red we are doing a projection- based on 28 real data - into the future by a least square fit on the logarithmic scale. - Let's start with Jack' notebook, in 2014 Jack will use a Teraflops Notebook, that means in less than 20 years after the first Teraflops system has entered our planet. -<CLICK> It will take years to move from #1 to #500. <CLICK> It will take years to move fom position #500 to notebook level. <CLICK> We will reach the Linpack Petaflops threshold most likely in 2008, one of the hot candidates will be RaodRunner at Los Alamos National Lab which has already shown up in our last list in 16 portions of 5 Teraflops each. <CLICK> And in 2015 we will have Petaflops systems only in the TOP500 list.

RoadRunner – the First Petaflop/s System in the World
and its Impact on Supercomputing Chair: Prof. Dr. Hans Meuer, Prometeus & University of Mannheim, Germany 1 Gflop/s 1 Tflop/s 100 Mflop/s 100 Gflop/s 100 Tflop/s 10 Gflop/s 10 Tflop/s 1 Pflop/s 10 Pflop/s 1 Eflop/s 100 Pflop/s 2019 ?? June 2008 RoadRunner Wile E. Coyote ?? June 1993 TOP500 Start June 1997 Intel ASCI Red 1986 Cray 2 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 32

Outline Mannheim Supercomputer Statistics & Top500 Project Start in 1993 Competition between Manufacturers, Countries and Sites My Supercomputer Favorite in the Top500 Lists The 31st List as of June 2008 Performance Development and Projection  Bell‘s Law Supercomputing, quo vadis? Top500, quo vadis? - My first topic are the so called Mannheim Supercomputer Statistics introduced by me at the very firtst Mannheim Supercomputer Conference in 1986 before we started with the TOP500 Project in I will present to you the results but also the deficiencies of these statistics which has been the reason to start with the TOP500 project in I will present to you the rationals and people behind this project. - Secondly I will focus on an extremely important topic in my point of view: COMPETITION. The TOP500 are so successful mainly because they promote and encourage COMPETITION: between manufacturers ,between countries and sites. - Then I am going to discuss with you my favorite Supercomputer out of ! -Our last list was published at tSC In Tampa last Novemer and I will give you a couple of results - In the next section we will focus on performance development and even performance projections - Finally I will address shortly our future plans

Bell‘s Law Bell's Law of Computer Class formation was discovered about 1972. It states that technology advances in semiconductors, storage, user interface and networking advance every decade enable a new, usually lower priced computing platform to form. Once formed, each class is maintained as a quite independent industry structure. This explains mainframes, minicomputers, workstations and Personal computers, the web, emerging web services, palm and mobile devices, and ubiquitous interconnected networks. We can expect home and body area networks to follow this path. From Gordon Bell (2007),

Bell‘s Law Bell’s Law states, that:
Important classes of computer architectures come in cycles of about 10 years. It takes about a decade for each phase : Early research Early adoption and maturation Prime usage Phase out past its prime Can we use Bell’s Law to classify computer architectures in the TOP500?

Bell‘s Law Gordon Bell (1972): 10 year cycles for computer classes
Computer classes in HPC based on the TOP500: Data Parallel Systems: Vector (Cray Y-MP and X1, NEC SX, …) SIMD (CM-2, …) Custom Scalar Systems: MPP (Cray T3E and XT3, IBM SP, …) Scalar SMPs and Constellations (Cluster of big SMPs) Commodity Cluster: NOW, PC cluster, Blades, … Power-Efficient Systems (BG/L or BG/P as first example of low-power / embedded systems = potential new class ?) Tsubame with Clearspeed, Roadrunner with Cell ?

Bell‘s Law Computer Classes / Systems

Bell‘s Law Computer Classes - refined / Systems

Bell‘s Law HPC Computer Classes Class Early Adoption starts:
Prime Use starts: Past Prime Usage starts: Data Parallel Systems Mid 70’s Mid 80’s Mid 90’s Custom Scalar Systems Mid 2000’s Commodity Cluster Mid 2010’s ??? BG/L or BG/P Mid 2020’s ???

Outline Mannheim Supercomputer Statistics & Top500 Project Start in 1993 Competition between Manufacturers, Countries and Sites My Supercomputer Favorite in the Top500 Lists The 31st List as of June 2008 Performance Development and Projection Bell‘s Law  Supercomputing, quo vadis? Top500, quo vadis? - My first topic are the so called Mannheim Supercomputer Statistics introduced by me at the very firtst Mannheim Supercomputer Conference in 1986 before we started with the TOP500 Project in I will present to you the results but also the deficiencies of these statistics which has been the reason to start with the TOP500 project in I will present to you the rationals and people behind this project. - Secondly I will focus on an extremely important topic in my point of view: COMPETITION. The TOP500 are so successful mainly because they promote and encourage COMPETITION: between manufacturers ,between countries and sites. - Then I am going to discuss with you my favorite Supercomputer out of ! -Our last list was published at tSC In Tampa last Novemer and I will give you a couple of results - In the next section we will focus on performance development and even performance projections - Finally I will address shortly our future plans

Supercomputing, quo vadis?
European Countries / Systems

Countries/Systems (June 2008) United States 52% Others 15% China 2% Japan 4% United Kingdom 11% France 7% Germany 9%

Countries/Performance (June 2008) United States 60% Others 15% Sweden 2% Japan 5% France 6% Germany 8% United Kingdom 7%

Countries/Systems (June 2008) Top 50 India 2% United States 52% Spain 2% Russia 2% Poland 2% Norway 2% United Kingdom 4% Sweden 6% Germany 8% Japan 10% France 10%

Continents/Systems

Asian Countries / Systems

Producing Regions / Systems

Outline Mannheim Supercomputer Statistics & Top500 Project Start in 1993 Competition between Manufacturers, Countries and Sites My Supercomputer Favorite in the Top500 Lists The 31st List as of June 2008 Performance Development and Projection Bell‘s Law Supercomputing, quo vadis?  Top500, quo vadis? - My first topic are the so called Mannheim Supercomputer Statistics introduced by me at the very firtst Mannheim Supercomputer Conference in 1986 before we started with the TOP500 Project in I will present to you the results but also the deficiencies of these statistics which has been the reason to start with the TOP500 project in I will present to you the rationals and people behind this project. - Secondly I will focus on an extremely important topic in my point of view: COMPETITION. The TOP500 are so successful mainly because they promote and encourage COMPETITION: between manufacturers ,between countries and sites. - Then I am going to discuss with you my favorite Supercomputer out of ! -Our last list was published at tSC In Tampa last Novemer and I will give you a couple of results - In the next section we will focus on performance development and even performance projections - Finally I will address shortly our future plans

Top500, quo vadis? www.top500.org
The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access.

Top500, quo vadis? Summary after Fifteen Years of Experience
TOP500 proofed itself by correcting the Mannheim Supercomputer Statistics. It is simplistic, but (or because of it) it gets trends right. It does not easily allow to track market size It’s inventory based. This smoothes seasonal fluctuation. Turn-over very high, so it still reflects recent developments. Replacement Rate To sum up after 14 years of experience: - the TOP500 proofed itself by correcting the old Mannheim Supercomputing Statistics which we used 7 years before we started with the TOP500 project. -the TOP500 is a simplistic approach with the very often attacked Linpach benchmark, but it gets trends right, e.g. with architectures, manufacturers, countries and sites. Extremey correct are the performance predictions are which we have seen during my talk.Therefore even predictions over 8 years are possible in a reliable way. - if you use the TOP500 that way you certainly can benefit al lot, much more as by grandmother's predictions from e.g. IDC, Diebold, etc. - but you should not use the TOP500 claiming that a system at position say 254 is better suited for your application than a system at position 354, this definitely is a misuse and we are always warning to use it tha wayt. The TOP500 are not suited for selecting a system with specific applications. You have to run your own benchmarks and application sets to select your system and not the TOP500. - with the TOP500 it is not possible to track the market size, in e.g. US $. We do not know how much are the systems at the different positions, of course we are ofen asked.. but do not record these informations. - In the TOP500 we have figured out the following opposite structures of entries: older systems especially in the upper part of the list remaining there for a couple of lits, they are smoothing seasonal fluctuations. And we have the short time entries- very often for only one list - since the turn-over is very high. In the small graph you see that appr. 180 systems are dropping out halfyearly since they are not powerful enough to make a new list.

Motivation for Additional Benchmarks
Top500, quo vadis? Motivation for Additional Benchmarks Linpack Benchmark Pros One number Simple to define and rank Allows problem size to change with machine and over time Allowing Competitions Clearly need something more than Linpack HPC Challenge Benchmark and others Cons Emphasizes only “peak” CPU speed and number of CPUs Does not stress local bandwidth Does not stress the network No single number can reflect overall performance Here I am discussing the more or less Pros and Cons of the Linpack Benchmark.

Top500, quo vadis? Presented at ISC’06 June 27-30 2006
Last year at ISC'06 in Dresden I have had invited my friend Jack to deliver the Friday Keynote presentation on the topic "HPC Challenge Benchmark and the TOP500". Jack gave an excellent talk on this topic and was rated by our attendants amongst the two best presentations at ISC'06.

Top500, quo vadis? For those who are not familiar with HPC Challenge Benchmark: - It consists of basicly 7 benchmarks, each of this benchmarks are stressing different parts of a computer sytem. - Of course HPL, the High Performance Linpack, is also part of these 7 benchmarks for stressing the CPU. - Those who are interested in the slides of Jack's 45 minutes speech last year should drp me an and I will send. - 7 results means of course that we no longer have the advantage of a singlke number, meaning that the results are presented by so called Kiviar Charts rathe tha a single number. - Btw, it is also Jack's meaning that we will see still Linpack as a yardstick in the TOP500 for a while but that on the other hand the TOP500 should be expanded to produce lists using anothrer yardstick.

Vielen Dank für Ihr Aufmerksamkeit! Die Folien sind hier erhältlich:

The 31st List as of June 2008 Manufacturers/Systems (June 2008)
IBM 43% Others 8% Dell 5% SGI 4% Cray Inc.3% Hewlett-Packard 37%

The 31st List as of June 2008 Manufacturers/Performance (June 2008)
IBM 48% Others 8% Sun Microsystems 3% Dell 6% SGI 6% Cray Inc.7% Hewlett-Packard 22%

The 31st List as of June 2008 Countries/Systems (June 2008)
United States 52% Others 15% China 2% Japan 4% United Kingdom 11% France 7% Germany 9%

The 31st List as of June 2008 Countries/Performance (June 2008)
United States 60% Others 15% Sweden 2% Japan 5% France 6% Germany 8% United Kingdom 7%

The 31st List as of June 2008 Processor Generations/Systems (June 2008) Others 12% POWER5+ 2% Xeon E54xx (Harpertown) 24% PowerPC 450 2% Opteron Quad Core 3% PowerPC 440 4% Xeon X54xx (Harpertown) 8% Xeon 53xx (Clovertown) 19% Opteron Dual Core 8% Xeon 51xx (Woodcrest) 18%

The 31st List as of June 2008 Processor Generations/Performance (June 2008) Others 9% Xeon 53xx (Clovertown) 13% Itanium2 Montecito Dual Core 2% POWER6 3% Xeon X54xx (Harpertown) 5% Xeon E54xx (Harpertown) 14% Opteron Dual Core 8% Opteron Quad Core 8% Xeon 51xx (Woodcrest) 11% PowerPC 440 9% PowerXCell 8i 9% PowerPC 450 9%

The 31st List as of June 2008 Multi-Core Level/Systems (June 2008)
Many Core 1% Single Core 2% Quad Core 56% Dual Core 41%

The 31st List as of June 2008 Multi-Core Level/Performance (June 2008)
Many Core 9% Single Core 2% Quad Core 51% Dual Core 38%

The 31st List as of June 2008 Cores per Socket /Systems

The 31st List as of June 2008 Cores per Socket /Performance [Tflop/s]

The 31st List as of June 2008 Concurrency Levels
If we look at the different countries involved in HPC we have the following slide: - USA increasing - Japan decreasing - Germany decreasing - UK increasing slightly

RoadRunner Der Roadrunner ist das erste System, das die Petaflops-Grenze durchbricht. Fotos: IBM The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: computerzeitung.de 13. Juni 2008

RoadRunner Im Roadrunner verbaut sind QS22-Cell-Blades, die in Deutschland entwickelt wurden. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: computerzeitung.de 13. Juni 2008

Der Cell-Prozessor hat neun Cores: acht rechnen, einer koordiniert.
RoadRunner Der Cell-Prozessor hat neun Cores: acht rechnen, einer koordiniert. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: computerzeitung.de 13. Juni 2008

RoadRunner IBM's Tri-blade
At the heart of Roadrunner is a special server node that IBM calls a triblade. This specially designed server is made up of two QS22 blades and an LS21 blade. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: eweek.com Juni 2008

RoadRunner The Cell-Based QS22
What makes Roadrunner unique is that IBM will use its Cell processor, originally designed for game consoles, as a way to accelerate the more difficult scientific calculations that the super-computer will perform during its time in Los Alamos. This is the inside of the QS22 blade, which holds the Cell processors. Roadrunner uses two QS22 blades with a total of four cell processors. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: eweek.com Juni 2008

RoadRunner The AMD-Based LS21
The LS21 blades use dual-core processors from AMD. This part of the machine handles the more basic computing, while the Cell-based servers are used to speed up the applications when needed. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: eweek.com Juni 2008

A Typical Roadrunner Rack
IBM turned to mostly commodity parts to create Roadrunner. Here, the blades are arranged inside a typical BladeCenter H chassis, which helps support 180 of the triblades in each rack. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: eweek.com Juni 2008

RoadRunner The Voltaire Switch
In keeping with its focus on commodity parts, the engineers at Los Alamos requested that IBM use Voltaire InfiniBand switches in the racks. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: eweek.com Juni 2008

Miles and Miles of Cabling
RoadRunner Miles and Miles of Cabling In putting together Roadrunner, IBM engineers needed to run miles and miles of fiber-optic cabling to connect the various pieces into one supercomputer. All together, IBM estimated that it used 57 miles of fiber-optic cabling in Roadrunner. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: eweek.com Juni 2008

The Roadrunner Console
Each section of Roadrunner contains a console used to interface with the supercomputer. In this case, each resembles an older ThinkPad model; an icon for Red Hat Linux, the operating system, is visible in the corner. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: eweek.com Juni 2008

Bringing It All Together
RoadRunner Bringing It All Together IBM is currently testing Roadrunner at its facility in upstate New York. The company is preparing to disassemble it in a few weeks and ship the system out to New Mexico. There, sometime in August, engineers will reassemble the supercomputer. The results of ALL 28 lists so far together with a variety of additional informations are available fro our website - This website has a remarkable traffic: 21 K PI/day (viewed traffic - not including traffic generated by robots, worms, etc.) - most recently we have put on our website a generator for the most powerful sites, please visit and adjust to your demands besides all the other informations waiting for your access. Quelle: eweek.com Juni 2008

Knocking on the doors of the Petaflop/s Erfahrungen mit 31 Listen und Ausblick auf die nächsten Jahre mit den TOP500 Hans Werner Meuer hans@meuer.de Universität.

Similar presentations

Presentation on theme: "Knocking on the doors of the Petaflop/s Erfahrungen mit 31 Listen und Ausblick auf die nächsten Jahre mit den TOP500 Hans Werner Meuer hans@meuer.de Universität."— Presentation transcript:

Similar presentations

About project

Feedback

Log in

Auth with social network:

Knocking on the doors of the Petaflop/s Erfahrungen mit 31 Listen und Ausblick auf die nächsten Jahre mit den TOP500 Hans Werner Meuer hans@meuer.de Universität.

Similar presentations

Presentation on theme: "Knocking on the doors of the Petaflop/s Erfahrungen mit 31 Listen und Ausblick auf die nächsten Jahre mit den TOP500 Hans Werner Meuer hans@meuer.de Universität."— Presentation transcript:

Similar presentations

About project

Feedback