Strindberg

The IBM Scalable POWERparallel (SP) series of supercomputers were part of the IBM RISC System/6000 (RS/6000) family and were also known as RS/6000 SP systems. The SPs consisted of multiple RS/6000-based nodes connected to each other through a device known as a switch.

An IBM SP2 supercomputer was installed at PDC in 1994 and the system was named after August Strindberg , a Swedish author and playwright who was also a painter. The system was significant in the evolution of supercomputing systems for academic research in Sweden for several reasons. The most important aspect of Strindberg was that it was the first cluster system at PDC, that is, it was built of standard workstations (personal computers), as many supercomputers are today. The fact that the system was built as a cluster made it easy to upgrade and adapt Strindberg so the system could grow as the needs of researchers developed towards more and more parallel processing. In fact, the many upgrades to Strindberg, some of which were significant changes, spanned much of the move from high-speed serial processing to the increasing adoption of parallel processing and the development of parallelisation techniques for coding.

The Strindberg system started as a 55-node system that was 97th in the June 1995 Top500 list . Then, early in 1996, the number of nodes was increased to 96 and the existing nodes were upgraded with faster processors and more memory. In addition, the amount of available disk storage was increased to 380 GB. This system had a peak performance of 267 MFLOPS per node, and the peak performance (for the whole system) was 25.6 GFLOPS. The system had a total memory capacity of 18.5 GB.

This system was the largest IBM SP2 outside the USA at the time, which meant that, for the first time in many years, there was a high-performance computer of international class in Sweden. This system was 64th in the June 1996 Top500 list . Later that year, in October 1996, the number of nodes in Strindberg was increased to 110, so the system was capable of a peak performance of 29 GFLOPS and had 20 GB of memory. That system came 80th in the November 1996 Top500 list .

There was a major upgrade of Strindberg in November 1997. The upgrade involved all parts of the system: the peak CPU performance of each batch node was increased by a factor 2.5, the amount of available memory grew by a factor of 2 and the network bandwidth by a factor 3, plus the number of nodes was increased from 110 to 122.

In January 1998, the system was upgraded yet again. The 130 new 160MHz nodes, together with 16 of the old nodes, made a very powerful 146-node supercomputer which came 60th in the June 1998 Top500 list .

As part of a research collaboration between IBM and PDC, it was also decided to add symmetric multiprocessor (SMP) nodes to the system in the spring of 1998. The six new SMP nodes (known as M nodes) that were added to Strindberg were each equipped with four IBM Power PC 604e 332 MHz CPUs. Each SMP node had an aggregated peak floating point performance of 2,656 MFLOPS, which gave an aggregated performance of more than 100 GFLOPS for the system as a whole with an aggregated memory bandwidth of 441 GB/s. The purpose of the collaboration was to investigate the applicability of SMP technology in general and, in particular, to evaluate the benefits of parallel SMP systems for several specific scientific codes.

In March 1999, Strindberg underwent yet another major upgrade of systems and hardware in which sixteen more SMP nodes were added. A further upgrade later in 1999 added 8 Nighthawk Power3 nodes. The resulting system had 178 nodes and about 2,000 GB of disk available to users. The peak performance of the system was 204 GFLOPS with a the total memory of 115 GB. The part of the system with the Nighthawk nodes could be used separately from the rest of the system. It was referred to as SP-NH and came 350th in the November 1999 Top500 list .

Later, the SP-NH part of the system was upgraded to have 11 Nighthawk nodes, each of which contained 8 or 16 IBM Power3 CPUs, giving a total of 112 CPUs. The nodes were interconnected by an IBM SP Switch2. There was over 380 GB of additional disk storage available for user data through the shared parallel file systems (280 GB for projects and 100 GB for scratch storage) and plenty of space in the distributed file systems for user home directories, application codes and so forth. 

Strindberg was used both as a computational server (using the nodes as powerful individual computational processors) and for parallelisation projects. Information for users of the system at that time noted the following.

• “To attain fast throughput, you can also have several nodes work simultaneously on one program. You get exclusive access to a number of processors, for the duration of a job. Parallelization of an existing program, or the development of a parallel program, can be quite cumbersome. Take full advantage of available expertise at PDC. In the context of a project, a PDC support team can work with you on the parallelisation of your software. ”

With all the upgrades and expansions of the system, Strindberg ended up being housed in many cabinets (also referred to as racks or frames). In December 2003, PDC was relocated to a different building at KTH (namely PDC’s current location at Teknikringen 14), and the parts of the Strindberg system with the Power2 and PowerPC nodes were shut down. The three cabinets with the Power3 nodes were moved to the new PDC computer hall in the basement of Teknikringen 14 where they ran for another year or so as a system known as Nighthawk.

File system connected to Strindberg

There was an innovative Parallel I/O File System (GPFS) connected to Strindberg. It was designed for both parallel and serial applications that required large temporary files and high I/O bandwidth. It had support for the creation of files as large as 128 TB that spanned multiple server nodes. This way, with GPFS file partitioning, users were able to parallelise access to their data without the inconvenience and administrative overhead of maintaining multiple data files.