VSC Seminar Talk: Modern Multi-Core Architectures for Supercomputing

On December 11, 2015 I gave a two hour tutorial talk on “Modern Multi-Core Architectures for Supercomputing” as part of the VSC School Seminar lecture series. This series’ goal is to push the efficiency of the codes running on the VSC clusters as well as improve the skills of the VSC School students.

My talk sets the stage for high performance computing for science in general and parallel computing in particular. The audience is supposed to have basic programming experience with C, C++ or Fortran as such it is perfectly suitable for (under)graduate students.

I start from the hardware basics where I cover the principal building blocks of modern processors, discuss memory hierarchies, pipelining and multi-core as well as multi-threaded processors. Based on those basics I move on to software aspects (frequently back-referencing to relevant hardware peculiarities), in particular discussing most common and important techniques to actually make use of the available parallel computing resources (which is the big problem we – the computational scientists – face today). Software-wise, I focus on OpenMP, vectorisation and distributed computing. Checkout the slides below.

Download (VSC_School_Seminar_2015_12_11_Weinbub_multi_core_architectures.pdf)

After my talk, Karl(i) Rupp continued with another two hour talk titled “Modern Many-Core Architectures for Supercomputing”. This talk covers hardware and software aspects of graphics processing units and co-processors as well as performance modelling. You can get the talk slides here.

On January 13, 2016 Jesper Larsson Träff (TU Wien) will give a talk on “Effective MPI Programming: Concepts, Advanced Features, Dos and Don’ts“.

 

 

VSC School Kick-Off

Our proposal for a research project within the VSC School got accepted and we got funding for one PhD student for three years. (Actually it got accepted some months back, but we were only recently able to make actually use of it.) Within this project, our future PhD student will conduct research in computational science on top of our flagship free open source linear algebra library: ViennaCL.

vsc_research_school_students

Accepted/Prospect VSC Research School PhD students/Postdocs

This much needed and highly important program for Austria’s HPC landscape enables us, the researchers, to not just fund research in computational science but also to conduct much required code development and code optimization of parallel code for large-scale compute clusters (such as our excellent VSC-based supercomputers – the brand new one being the VSC-3, which starts regular operation on April 1, 2015).

The VSC Research School’s importance for Austria’s research infrastructure stems from the fact that software-related funding is usually not possible with conventional funding institutions (e.g. FWF), as it does not represent basic research. The fact of the matter is, however, that the lion’s share of today’s science and engineering research just won’t work without specifically tailored research software, especially when the problems at hand are so computational intensive that they require highly parallel processing (or do you want to wait months to investigate simulation results?!) – hence, the absolute need for supercomputers and a program like the VSC Research School and its excellent support team to aid the researchers in making efficient use out of a behemoth like the VSC-3. (Let me tell you that developing code, which efficiently utilizes supercomputers, is no small feat.)

Ok, back to topic: On March 5, 2015 we had the kick-off meeting, organized by the highly motivated VSC team. After the meeting, where all projects got introduced and some networking was done, we got the chance to visit VSC-2 (the by now old supercomputer) and the new one, VSC-3, which are both located at the Arsenal.

vsc_3_plaque

The VSC-3 plaque, mounted right next to the cluster.

We took part of a guided tour led by the technical manager of the VSC, Ernst Haunschmid.

vsc_3_box

Ernst Haunschmid shows us one of the many units, containing a bunch of compute blades submerged in, well, oil. Also, lot’s of cables.

As the computers are oil cooled, there is no need for an AC cooling down the entire room (as is the case for the conventional VSC-2 cluster). Although the oil cooling method is superior with respect to efficiency and energy consumption, the room gets rather warm: it had around 35 degrees Celsius.

vsc_3_switches

VSC-3’s QDR-80 dual-link high-speed InfiniBand spine switches (on the back) and core switches (on the front). I so infinitely not envy the guy responsible for installing the network cables. Also, Florian.

We were allowed to take a closer look at the insides of the big compute tubs and let me tell you, yes everything (the compute blades that is) is submerged in oil. In a response to a question regarding what the crew does with the used/old oil (yes, it wears out over time), the team responded, that the oil is so well-filtered and biological that in principle it could be sold to Viennese pastry shops for shining-up the surface of the famous Austrian pastry: Punschkrapferl. Well, obviously this was a joke, but, if you ask me, there is a business model, right there.

Much to my surprise, I found this in the midst of all the equipment to keep the VSC-3 up and running.

vsc_3_cats

No, there are no cats based at the VSC facility.

Turns out that the VSC crew uses cat grit to tackle spilled oil. This is what I call: out-of-the-box thinking.

In conclusion, the kick-off meeting was great. The other projects are very interesting and I can’t wait to see how the entire program evolves.

To the VSC team: Thank you for your efforts in not only organizing the kick-off meeting and the guided tour, but also the program itself and the support you provide to the participating universities!