TOP500 Supercomputer Sites
Total Page:16
File Type:pdf, Size:1020Kb
TOP500 Sup ercomputer Sites Jack J. Dongarra Computer Science Department UniversityofTennessee Knoxville, TN 37996-1301 and Mathematical Science Section Oak Ridge National Lab oratory Oak Ridge, TN 37831-6367 [email protected] Hans W. Meuer Computing Center University of Mannheim D-68131 Mannheim Germany [email protected] Erich Strohmaier Computer Science Department UniversityofTennessee Knoxville, TN 37996-1301 [email protected] RUM 48/96 Novemb er 18, 1996 TOP500 Sup ercomputer Sites Jack J. Dongarra, Hans W. Meuer, and Erich Strohmaier November 18, 1996 Abstract To provide a b etter basis for statistics on high-p erformance comput- ers, we list the sites that have the 500 most p owerful computer systems installed. The b est Linpack b enchmark p erformance achieved is used as a p erformance measure in ranking the computers. 1 Intro duction and Ob jectives Statistics on high-p erformance computers are of ma jor interest to manufactur- ers, users, and p otential users. These p eople wish to know not only the number of systems installed, but also the lo cation of the various sup ercomputers within the high-p erformance computing community and the applications for which a computer system is b eing used. Such statistics can facilitate the establishment of collab orations, the exchange of data and software, and provide a b etter un- derstanding of the high-p erformance computer market. Statistical lists of sup ercomputers are not new. Every year since 1986 Hans Meuer [1] has published system counts of the ma jor vector computer manufac- turers, based principally on those at the Mannheim Sup ercomputer Seminar. Statistics based merely on the name of the manufacturer are no longer useful, however. New statistics are required that re ect the diversi cation of sup er- computers, the enormous p erformance di erence b etween low-end and high-end mo dels, the increasing availability of massively parallel pro cessing MPP sys- tems, and the strong increase in computing power of the high-end mo dels of workstation suppliers SMP. To provide this new statistical foundation, wehave decided in 1993 to assem- ble and maintain a list of the 500 most p owerful computer systems. Our list has b een compiled twice a year since June 1993 with the help of high-p erformance computer exp erts, computational scientists, manufacturers, and the Internet community in general who resp onded to a questionnaire we sent out; we thank all the contributors for their co op eration. In the present list whichwe call the Top500, we list computers ranked by their p erformance on the Linpack Benchmark. While we makeevery attempt to verify the results obtained from users and vendors, errors are b ound to exist and should b e brought to our attention. Weintend to continue to up date this list half-yearly and, in this way,tokeep track with the evolution of computers. Hence, wewelcome any comments and information; please send electronic mail to [email protected]. The list is freely available by anonymous ftp to 1 ftp.uni-mannheim.de/top500/ or to www.netlib.org/b enchmark/top500.ps. The interested reader can additionally create sublists out of the Top500 database and can make statistics on his own by using the WWW interface at http://parallel.rz.uni-mannheim.de/top500.html or http://www.netlib.org/b enchmark/top500.html. Here you also have access to p ostscript versions of slides dealing with the inter- pretation of the present situation as well as with the evolution over time since we started this pro ject. 2 The LINPACK Benchmark Asayardstick of p erformance we are using the \b est" p erformance as measured by the Linpack Benchmark [2]. Linpack was chosen b ecause it is widely used and p erformance numb ers are available for almost all relevant systems. The Linpack Benchmark was intro duced by Jack Dongarra. A detailed description as well as a list of p erformance results on a wide variety of machines is available in p ostscript form from netlib. To retrieve a copy send electronic mail to [email protected] and bytyping the message send performancefrom benchmark or from any machine on the internet typ e: rcp [email protected]:benchmark/performance performance. The b enchmark used in the Linpack Benchmark is to solve a dense system of linear equations. For the Top500, we used that version of the b enchmark that allows the user to scale the size of the problem and to optimize the software in order to achieve the b est p erformance for a given machine. This p erformance do es not re ect the overal l performance of a given system, as no single number ever can. It do es, however, re ect the performance of a dedicated system for solving a dense system of linear equations. Since the problem is very regular, the p erformance achieved is quite high, and the p erformance numb ers give a go o d correction of p eak p erformance. By measuring the actual p erformance for di erent problem sizes n, a user can get not only the maximal achieved p erformance R for the problem size N max max but also the problem size N where half of the p erformance R is achieved. max 1=2 These numb ers together with the theoretical p eak p erformance R are the peak numb ers given in the Top500. In an attempt to obtain uniformity across all computers in p erformance rep orting, the algorithm used in solving the system of equations in the b enchmark pro cedure must con rm to the standard op eration count for LU factorization with partial pivoting. In particular, the op eration 3 2 count for the algorithm must b e 2=3n + O n oating p oint op erations. This excludes the use of a fast matrix multiply algorithm like \Strassian's Metho d". This is done to provide a comparable set of p erformance numb ers across all computers. If in the future a more realistic metric nds widespread usage, so that numb ers for all systems in question are available, wemay convert to that p erformance measure. 2 3 The TOP500 List Table 1 shows the 500 most p owerful commercially available computer systems known to us. Tokeep the list as compact as p ossible, we show only a part of our information here: N Position within the Top500 ranking world Manufacturer Manufacturer or vendor Computer Typ e indicated by manufacturer or vendor Installation Site Customer Lo cation Lo cation and country Year Year of installation/last ma jor up date Field of Application 1 Pro c. Numb er of pro cessors R Maximal Linpack p erformance achieved max R Theoretical p eak p erformance peak N Problemsize for achieving R max max N Problemsize for achieving half of R max 1=2 If R from Table 3 of the Linpack Rep ort [2] is not available, we use the TPP max p erformance given in Table 1 of the Linpack Rep ort [2] for solving a system of 1000 equations. To use a consistent yardstick for all systemwewe do not use results achieved by advanced parallel algorithm as de ned in [2]. In case of the Cray T90, C90 and J90 systems we had to use older able 3 or Table 1 results. In a few cases weinterp olated b etween two measured system sizes. For mo dels where we did not receive the requested data, the p erformance of the next smaller system measured is used. If there should b e anychanges in the p erformances given in Table 1 we will up date them. In addition to cross checking di erent sources of information, we select ran- domly a statistical representative sample of the rst 500 systems of our database. For these systems we ask the supplier of the information to establish direct con- tact b etween the installation site and us to verify the given information. This gives us basic information ab out the quality of the list in total. As the TOP500 should provide a basis for statistics on the market of high- p erformance computers, we limit the numb er of systems installed at vendor sites. This is done for eachvendor separately by limiting the accumulated p erformance of systems at vendor sites to a maximum of 5 of the total accumulated installed p erformance of this vendor. Rounding is done in favor of the vendor in question. In Table 1, the computers are ordered rst by their R value. In the case max of equal p erformances R value for di erent computers, wehavechosen to max order by R . For sites that have the same computer, the order is by memory peak size and then alphab etically. 3 Top500 Sup ercomputers - Worldwide N Manufacturer Installation Site Field of R N max max world Computer Lo cation/Year Application Pro c. R N peak 1=2 [M op/s] 1 Hitachi/Tsukuba Center for Computational Physics, Univ of Tsukuba Academic 2048 368200 103680 CP-PACS/2048 Tsukuba Japan /1996 614000 30720 2 Fujitsu NAL Research 167 229700 66132 Numerical Wind Tunnel Japan /1996 Aerospace 281000 18018 3 Hitachi UniversityofTokyo Academic 1024 220400 138240 SR2201/1024 Tokyo Japan /1996 307000 34560 4 Intel Sandia National Labs Research 3680 143400 55700 XP/S140 Albuquerque USA /1993 184000 20500 5 Intel Oak Ridge National Lab oratory Research 3072 127100 86000 XP/S-MP 150 Oak Ridge USA /1995 154000 17800 6 Intel Japan Atomic Energy Research Research 2502 103500 .