Presentation of our results at the LSST All Hands Meeting

LSST All Hands Meeting brings together people who work on the LSST project, so that they can exchange the latest information on the work of their individual groups and talk about the next collaborative projects. The main aim is to push LSST project toward realization.

This year our GPU group from Split attended the meeting. Mario Juric and Kresimir Cosic were invited to present during a panel session their experiences with GPUs and show how GPUs could be used for the benefit of LSST.

New job scheduler on Hybrid

Hybrid cluster has undergone various system upgrades, including the job scheduler. Now it uses Sun Grid Engine (SGE) to handle job submissions.

More information is available on the project's wiki pages.

The final report to the Croatian Science Foundation

Our project of building a hybrid GPU-CPU cluster is financed by the Croatian Science Foundation. The project has been funded for one year and here you can find the final report to the Foundation.

The report describes what we have done and the problems we encountered during that period. It also provides details on the financial aspects of the project.

Content of the report:
1. Project activities in the second half-year period
2. Project timetable
3. Examples of research activities on Hybrid
4. Publications and presentations
5. List of costs

Collaboration with the LSST project

The Large Synoptic Survey Telescope project (LSST) is the most ambitious ongoing project in the US astronomy. For the first time, the LSST will provide time-lapse digital imaging of faint astronomical objects across the entire sky. The LSST has been identified as a US scientific priority in reports by diverse national panels.

In the end of 2009 we received a sub-award from the LSST, under which they co-finance one programmer for a year. The work on Hybrid that led to this sub-award started already during the testing phase of our cluster. This initial work focused on speeding-up the code that calculates realistic distribution of stars in our Milky Way galaxy and creates a catalogue of stars that a telescope would see. The work was initiated by Mario Jurić (Harvard University), the co-PI on the Hybrid project. He worked with Krešmir Ćosić, our project assistant. The total runtime was shortened by a factor of 200 thanks to GPUs.

This success attracted the interest of LSST, because GPUs could be of a great help in their efforts to produce simulated images for the future LSST telescope. Currently it takes a week to run one LSST simulation on CPUs. This led to an official collaboration agreement with the LSST Corporation to work on adapting parts of their code to work on GPUs. These first results on improving the Milky Way simulation code are going to be presented at the 215th American Astronomical Society Meeting in Washington, DC, in January 2010.

Slides from our first open lectures

Here you can find some slides and a couple of code examples from our first open lectures on programming in CUDA: www.gpuhybrid.org/wiki.
The lectures where given by:
- Jurica Teklić (he presented an introduction to programming in CUDA),
- Krešimir Ćosić (he showed the source code of an example of using GPUs to speedup a 3D visualization of a galaxy with more than 20 million of stars)
- Dejan Vinković (he showed an example of calculating a 2D configuration of electrostatic potential on GUPs)

Photos from the event are below.

Our first introductory lecture on using GPUs

We are organizing our first open lecture on using GPUs for general purpose computing. It is taking place at the Physics Department, University of Split, on 3. December 2009. A two hour lecture is going to be in Croatian.

---

RAČUNANJE NA GRAFIČKIM KARTICAMA

3.12.2009 u 18-20h, I kat, Kopilica 2, Split,
Odjel za fiziku, Prirodoslovno-matematički fakultet, Sveučilište u Splitu
(kako doći do nas: ucrtano na karti grada)

Iako su grafičke kartice poznate po generiranju i prikazu slike, odnedavno višejezgrena arhitektura grafičkog procesora omogućava ubrzanja nekih opće računalnih aplikacija za gotovo 100 puta.

U sklopu projekta izrade CPU/GPU klastera pri Odjelu za fiziku PMF-a u Splitu (http://gpuhybrid.org), organiziramo uvodnu radionicu na temu programabilnih grafičkih kartica. Cilj ove uvodne radionice je upoznavanje s novim mogučnostima grafičkih procesora kroz programiranje, korištenjem CUDA programskog modela u C/C++.

Kroz dvosatno druženje govorit ćemo o sljedećim temama:
- GPU arhitektura
- CUDA programski model
- osnove programiranja za CUDA-u
- primjeri i primjene
Potrebno predznanje: poznavanje osnova C/C++ programskih jezika.
Slobodno ponesite prijenosno racunalo, ako sadrži jednu od ovih grafičkih kartica: http://www.nvidia.com/object/cuda_learn_products.html

Molimo sve zainteresirane da se jave na e-mail adresu teklic_at_pmfst.hr

The very first test of the visualization equipment

Yesterday evening we performed the very first test of our visualization equipment. There are still details to tweak, but the very first view convinced us that it is worth the money invested.

Croatian Science Foundation approved the second payment installment

After evaluating our half-year report the Croatian Science Foundation (NZZ) approved the second payment installment for our project.

This way NZZ will invest 100,000 euros in total into our project. The grant can be used only for purchasing computer hardware and other equipment.

Our project's domain name and Wiki pages introduced

Our project now has its own domain name: gpuhybrid.org.
This is now the new location of the project's homepage.

We also started with Wiki pages that will be used by users of our GPU/CPU cluster to write about various practical topics related to the cluster usage and programing on GPUs. The address is: www.gpuhybrid.org/wiki

Programming on GPUs using CUDA (a manual written in Croatian)

Thanks to Matija Piškorec, a student from the Faculty of Electrical Engineering and Computing, University of Zagreb, we now have a Croatian manual on writing codes in CUDA.

His booklet is written in Croatian and contains step by step explanations and examples of codes in CUDA. He used our GPU/CPU cluster to run and test the codes.

You can download it from here.

1st report to the Croatian Science Foundation

Our project of building a hybrid GPU-CPU cluster is financed by the Croatian Science Foundation. The project lasts for one year and here you can find the first half-year report to the Foundation.

The report describes what we have done so far and the problems we encountered during that period. It also provides details on the financial aspects of the project.

Content of the report:
1. Project background
2. Project timetable for the first 6 months
3. Project activities in the first 6 months
3.1. Infrastructure and equipment
3.2. Personnel support
3.3. Fundraising and expansion of the collaborative network
4. Cluster configuration
5. Dissemination efforts

HYBRID in the local news

Local daily newspaper Slobodna Dalmacija has published a short story on April 3rd, 2009, about our cluster (in Croatian).

A wider context to the project Hybrid

It is my pleasure to inform you that the hybrid GPU-CPU cluster is finally assembled and the first "burn-out" tests of the equipment are underway!

Since the scheduling script is still to be designed, for now the cluster will be opened for logins on individual nodes that can access two GPUs, so that project collaborators can start playing around with the coding and code performance.

We were slowed down considerably due to various problems in purchasing the equipment and assembling the room for the cluster (power supply, acoustic isolation). The good news is that we have a dedicated room for our equipment and a dedicated classroom that will follow in the second part of the project by the end of the summer. I have to mention that the cluster would not be assembled without the hard work of our two project assistants: Jurica Teklić and Dubravko Balić.


Another very good news is that, after a long wait, the Ministry of Science and Education approved the novice position (PhD position) for one graduate student working on the Hybrid.

As an effort to expand this project and get some money for additional personnel and travel support, we applied for a new research grant and now we are waiting for the results of the first round of evaluations. The important news regarding this application is that we have two companies involved as project partners and technical support to our project. One is VEUS d.o.o. from Zagreb and the other is Mirriad Limited from UK. The project partner is also the Department of Astronomy at the University of Washington as the host of our collaboration with the LSST project (http://www.lsst.org). This collaboration with LSST is especially interesting due to the scale and importance of the LSST project.

In general, the project proposal has four main Themes that will receive a special attention if we get the funds:
1. Large scale data mining (handling large data sets such as the LSST database)
2. Computational microscope (simulation of macromolecules and materials on atomic scale in general)
3. Particle dynamics (N-body problems and dynamics of individual "particles" in various environments)
4. Weather Forecasting for Split region (improving weather forecasting models for Split region, which has a specific microclimate)

First Node and Tesla Unit Active

Yesterday, March 10th 2008, we successfully booted and provisionally set up the first two CPU nodes (2 quad core CPUs each) paired with one quad-GPU Tesla S1070 unit. Present at the happy event was the entire team (Dubravko, Jurica, Dejan & Mario -- see the photos), but the credit for setting everything up goes entirely to Jurica & Dubravko.

Three other Tesla units are already at the site, while the remaining six nodes are expected to arrive later this week. If everything goes according to plan, we should have the hardware portion of the cluster assembled next week.

HYBRID's rack with the UPS, front-end node, two slave nodes, and a Tesla S1070.

Zoom-in on Tesla S1070.

Very bottom to top: UPS, front-end node, two slave nodes in a single 1U case, and a Tesla S1070 (the top 1U case).


First diagnostics.

Prof. Vinkovic (PI) with his graduate student Jurica Teklic.

Co-I Mario Juric, visiting from the Institute for Advanced Study.

The local team (CW from center: Dejan Vinkovic, Jurica Teklic, Kresimir Cosic and Dubravko Balic).

The Blog of GPU/CPU Hybrid Computer Cluster Project

Dear readers,

Welcome to the blog of the Hybrid Computer Cluster (HYBRID)project, lead by prof. Dejan Vinkovic at the University of Split. HYBRID is an experimental cluster that explores the use of Graphical Processing Units (GPUs) for low-cost general-purpose computing. It combines 16 quad core CPUs in 8 dual-processor nodes, with 16 NVIDIA TESLA 10 series GPUs in 4 Tesla S1070 units, into a cluster capable of 16 TFLOP of single (~2TFLOP double) precision computing power.

We will use this blog to communicate the status and new developments of the project both to our research groups, as well as the public at large.