from Wikipedia, the free encyclopedia

MareNostrum ( Latin Mare nostrum 'our sea' , a Roman name for the Mediterranean ) is the name for several generations of supercomputers at the Universitat Politècnica de Catalunya in Barcelona . The facility is operated by the Barcelona Supercomputer Center BSC. The supercomputer is intended for research in the life sciences , meteorology , and environmental sciences , and for commercial applications in the pharmaceutical , automotive, and aeronautical fields . The complete technical installation was built into the former chapel of Torre Girona and is surrounded by 5 m high glass walls. The supercomputer has been replaced several times by more up-to-date technology since it was first put into operation in 2004.


MareNostrum fulfills a variety of tasks in science, research and teaching. The working areas include big data , bioinformatics , biomechanics , climatology , cloud computing , cognitive science , computer architecture and code design, distributed systems , training , technical simulations, fusion energy , genome research , geophysics , software development for supercomputers, materials science , molecular modeling, operations infrastructure, performance analysis, programming models , Social simulations, computational earth sciences, extreme math problems and algorithms, quantum information.

MareNostrum 1

When it was put into operation in November 2004, MareNostrum achieved a continuous computing power of 20 teraflops using 3,564 PPC 970, 2.2 GHz processors and thus came 4th in the TOP500 in November 2004 .

On 13 April 2005, the computer for the first time at 4800 PPC 970 was 2.2 GHz processors and Myrinet with its full capacity of 27.9 Teraflops booted . In this new configuration, it achieved fifth place in the June 2005 edition of the comparison list with a total of 4,812 processors.

MareNostrum 2

After a conversion in 2006 to PPC 970, 2.3 GHz processors and Myrinet, MareNostrum 2 had 10,240 cores and 20 TB RAM and thus achieved 62.6 teraflops. The system had approximately 300 TB of disk space. That was enough in November 2006 for fifth place on the list. In 2008 it came in 41st place worldwide with a performance of 63.8 TFLOPS as Spain's fastest system. In June 2012 it still occupied place 465. When the system was dismantled, the remaining parts were divided into smaller clusters with 256 and 512 computing nodes, which continued to be operated in various Spanish universities and institutes.

MareNostrum 3

MareNostrum 3

For the operation of MareNostrum 3, extensive construction work to strengthen the power supply and the cooling system was necessary.

Two systems were in operation between 2011 and 2013. One was Bullx B505, a system of 5,544 Xeon E5649 6C 2.53 GHz processors, InfiniBand QDR, NVIDIA 2090 processors and 3024 GB of memory. This system managed 103.2 teraflops and was in operation until 2013.

MareNostrum 3 went into operation between 2012 and 2013 and initially used 33,664 DX360M4, Xeon E5-2670 8-core, 2,600GHz processors and InfiniBand FDR for the connection. The performance was 636.9 teraflops. As of 2013, the computer had 48,896 Intel Sandy Bridge processors in 3,056 nodes, plus 84 Xeon Phi 5110P in 42 nodes, with more than 115 TB main memory and 2 PB of GPFS disk storage . In total it reached 925.1 teraflop and 1.1 petaflop peak. The system reached 29th place in the Top500 ranking in June 2013.

MareNostrum 4

MareNostrum 4

MareNostrum 3 was replaced by MareNostrum 4 from mid-2017. The new system surpasses MareNostrum 3 by a factor of 10 to 12. It initially had 11.1 petaflops peak computing capacity and achieved 13.7 petaflops with the additional cluster of IBM Power9 and Nvidia Volta processors. According to the Top 500 ranking from June 19, 2017, it was the third strongest cluster in Europe and the thirteenth worldwide. MareNostrum 4 is connected to the Big Data facilities of the Barcelona Supercomputer Center BSC, which have a storage capacity of 24.6 petabytes, and is connected to the European universities via the RedIris and GÉANT networks.

The heterogeneous architecture is remarkable. There is the general block that does the main computing work and an additional block for researching newly developed technologies. Five storage units (Elastic Storage) manage 14 petabytes of data, an Intel Omni-Path high-speed network and an Ethernet link the components.

General-purpose cluster

  • The general block initially consisted of 48 racks of 3,456 nodes. Each node has two Intel Xeon Platinum chips, each with 24 processors, for a total of 165,888 processors and a main memory of 290 terabytes. Although the output has increased by a factor of 10, the energy requirement only increased by 30% to 1.3 MW. In mid-2018 the system consisted of 2x Intel Xeon Platinum 8160 24C with 2.1 GHz, 216 nodes with 12x32 GB DDR4-2667 DIMMS (8 GB / core) and 3,240 nodes with 12x8 GB DDR4-2667 DIMMS (2 GB / core). SUSE Linux Enterprise Server 12 SP2 serves as the operating system .

Emerging Technologies Blocks

The block with the newly developed technologies contains clusters of three different technologies that are integrated and updated as they become available on the market. New processors and software can thus be operated, tested and optimized even before the next generation of computers is fully developed. Specialized chips e.g. B. Graphics processors can optimize and accelerate corresponding tasks to a particular degree. The transition to future new technologies can thus take place smoothly.

  • A cluster of IBM POWER9 and Nvidia Volta GPUs with a computing power of over 1.5 petaflops. The cluster with the name CTE - Power consists of 52 nodes. Each node consists of 2 x IBM Power9 8335-GTG @ 3.00 GHz (2 x 20 cores and 4 threads / core, a total of 160 threads per node), 512 GB main memory distributed over 16 DIMMs x 32 GB @ 2666 MHz, 2 x SSD 1.9 TB as local storage, 2 x 3.2 TB NVME, 4 x GPU NVIDIA V100 (Volta) with 16 GB high bandwidth memory 2, single port Mellanox EDR, GPFS over fiber optic 10 GBit, operating system Red Hat Enterprise Linux Server 7.4. The cluster went into operation in May 2018 and already exceeded the performance of MareNostrum 3 by 50%. This cluster is particularly energy efficient and placed the system in 9th place in the Green500. This cluster contains 19,440 processors and a total of 27,648 GB RAM and came in 255th place in June 2018. MareNostrum was thus represented in the top 500 with two systems at the same time.
  • A cluster of Intel Knights Hill (KNH) processors with a computing capacity of 0.5 petaflops was originally planned. After Intel discontinued the Knight Hill processors, plans went in the direction of another Intel-based system, possibly the Aurora A21 for the year 2021. Then it was decided to use a cluster of AMD Rome processors and AMD Radeon Instinct MI50 graphics processors. The system is said to reach 0.52 petaflops.
  • A cluster of 64Bit ARMv8 processors with a computing capacity of 0.5 petaflops.


BSC also operates other larger computing nodes. The second largest cluster under the name MinoTauro unites 39 servers each with in 2019

  • 2 Intel Xeon E5-2630 v3 8-core processors, 2.4 GHz
  • 2 K80 NVIDIA GPU cards
  • 128 GB main memory
  • 120 GB solid state disk as local storage
  • 1 PCIe 3.0 x8 8GT / s, Mellanox ConnectX-3FDR 56 Gbit
  • 4 Gigabit Ethernet ports

The system achieved a total of 250.94 Tflops Peak, of which 226.98 Tflops from the graphics processors and 23.96 Tflops from the main processors. Red Hat Enterprise Server is used as the operating system


Barcelona is planned as the location for one of the precursor computers of the European supercomputers, which are funded as part of the European High-Performance Computing Joint Undertaking - EuroHPC. This computer is expected to have over 200 petaflops peak computing capacity and go into operation on December 31, 2020. MareNostrum 5 is to have a budget of 223 million euros, which includes the purchase, installation and operation for five years. Half of the budget is provided by the EU, the other half comes from the states of Spain, Portugal, Turkey, Croatia and Ireland, which form a consortium for this purpose. Clusters at eight different locations in eight different European member states are planned for EuroHPV. The selected locations are Sofia (Bulgaria), Ostrava (Czech Republic), Kajaani (Finland), Bologna (Italy), Bissen (Luxembourg), Minho (Portugal), Maribor (Slovenia) and Barcelona (Spain). 19 of the 28 EU members are involved in the entire project, as well as some countries that are not part of the EU. The project has a total budget of around € 840 million. There are said to be three forerunner machines with more than 150 petaflops, which will later be supplemented by 5 units in Exa-scale and five petascale machines with 4 petaflops. The predecessor machines should provide roughly four times more computing power than the current systems of the Partnership for Advanced Computing in Europe (PRACE). Another goal is the development and integration of a European processor technology, which should eliminate the dependence on non-European technology. The commissioning of MareNostrum5 is planned for the end of 2020. MareNostrum5 will take up more space than MareNostrum4, the space in the Torre Girona chapel will no longer be sufficient, so that part of it will have to be installed in the neighboring BSC building.

Web links

Commons : MareNostrum  - collection of images, videos and audio files

Individual evidence

  1. MareNostrum - eServer BladeCenter JS20 (PowerPC970 2.2 GHz), Myrinet | TOP500 supercomputer sites. Retrieved March 18, 2018 .
  2. MareNostrum - JS20 Cluster, PPC 970, 2.2 GHz, Myrinet | TOP500 supercomputer sites. Retrieved March 18, 2018 .
  3. MareNostrum - BladeCenter JS21 Cluster, PPC 970, 2.3 GHz, Myrinet | TOP500 supercomputer sites. Retrieved March 18, 2018 .
  4. Bullx B505, Xeon E5649 6C 2.53GHz, Infiniband QDR, NVIDIA 2090 | TOP500 supercomputer sites. Retrieved March 18, 2018 .
  5. a b MareNostrum | BSC-CNS. Retrieved March 17, 2018 .
  6. Barcelona Supercomputing Center (Ed.): MareNostrum III User's Guide . ( [PDF]).
  7. ^ MareNostrum 3 | BSC-CNS. Retrieved March 17, 2018 .
  8. Top500 List - June 2013 | TOP500 supercomputer sites. Retrieved March 18, 2018 .
  9. Top500 List - June 2017 | TOP500 supercomputer sites. Retrieved April 2, 2018 .
  10. a b c d e f MareNostrum 4 begins operation. Retrieved March 17, 2018 .
  11. Technical Information | BSC-CNS. Retrieved July 28, 2018 .
  12. Support Knowledge Center @ BSC-CNS. Retrieved July 28, 2018 .
  13. MareNostrum 4 POWER9 racks begin operation with high expectation for AI - based research | BSC-CNS. Retrieved July 28, 2018 .
  14. The new BSC machine is Europe's “greenest” supercomputer | BSC-CNS. Retrieved July 28, 2018 .
  15. MareNostrum P9 CTE - IBM Power System AC922, IBM POWER9 22C 3.1GHz, Dual-rail Mellanox EDR Infiniband, NVIDIA Tesla V100 | TOP500 supercomputer sites. Retrieved November 9, 2018 .
  16. BSC Fires Up Power9-V100 Hybrid Compute On MareNostrum 4 . June 13, 2018 ( [accessed July 31, 2018]).
  17. MinoTauro. Retrieved February 8, 2019 .
  18. MareNostrum 5 will host an experimental platform to create supercomputing technologies “made in Europe”. Accessed August 11, 2019 .
  19. European Commission - PRESS RELEASES - Press release - Digital Single Market: Europe announces eight locations for new world-class supercomputers. Retrieved August 13, 2019 .