Archive

Posts Tagged ‘computation’

Building a harvester rig

I decided to build a workstation computer, for several reasons. I have access to many different types of computing resources, with differing advantages and disadvantages. The Nectar Grid for instance provides virtual machines with many cores. Very good for coarse-grained parallel processing, like the case when I have 132 million semigroups in a file and I would like to classify them according to their eggbox pictures. I can divide the file into smaller chunks and using GNU Parallel I can easily utilize all available processors. But for single thread calculations these virtual machines are not ideal since they tend to have older/power efficient slower processors. I’m also fed up with laptops, they are fine for writing code, but once under some load they get hot and burn the battery in an hour. Also, PC building is a dying art and I wanted to have one more go, before it disappears.

For most semigroup and group algorithms the bottleneck is the memory. Orbit calculations, or more generally search trees are not trivially parallelizable. So, the requirements for the new rig:

  1. good single-thread performance
  2. large memory bank with error correction
  3. 24/7 energy efficient operation

One fast processor with huge amount of memory. This is of course totally against the post-Moore’s law era’s trends: many slow cores with little local memory.

Why error correction? RAM modules are quite reliable nowadays, but they do fail sometimes. In most cases the buggy software is blamed, as software failures are more frequent and there is no way to tell the difference between software and hardware failure. But they are rare. So, unless you need very precise calculations that run for days, you don’t need ECC RAM. But that’s exactly what we are doing here, sometimes a week long computation where absolute mathematical precision is required, so not even a single bit-flip caused by a malevolent cosmic ray is allowed. It freaks you out if you start thinking about undetected memory failures. It is a bit like the question whether you want to drive with or without the safety belt.

Choices for the components:

Motherboard: For the Intel 22nm Haswell processors, ASUS P9D WS with the Intel C226 chipset which is meant to be a server chipset (so no overclocking), but the board has some fancy features, like the UEFI BIOS.

DSC_3426

Memory: 4x8GB low-voltage, 1600MHz ECC modules. 32GB in total, 16GB per channel.

DSC_3440

Power supply: Fanless 400W Platinum (Seasonic). The thought of wasting energy when producing power is simply outrageous. That’s why the premium PSU.

Graphics card: No gaming here.

Storage: Another non-issue. Just use whatever drive I have available at the moment.

Processor: Intel Pentium G3420

DSC_3445

No pins here. You have to worry about the motherboard, not the CPU.

DSC_3437

What?!?!? A budget processor in ‘high-end’ workstation?!? Well, it is enough for the current purpose. For single thread calculations you don’t need many cores/hyper-threading. This Pentium does support ECC, therefore it is more like a crippled Xeon than a dumbed-down i7. The name ‘Pentium’ is very nostalgic. There is an upgrade path to E3 Xeons, so the build is somewhat future-proof.

After reading Inside Intel this company would be my last choice to buy from, but their products are actually superior. Faster and more power efficient. The full system power consumption is 23W when idling. Not that this value matters, since the machine  is only switched on when there is something to calculate. Half-steam takes 40W, while full blast (both cores using all 4 memory modules) requires 51W. Cooling is not an issue. Unlike my desktop with an AMD FX 8320.

Harvester?!? Well, the expression ‘mining rig’ has been taken by the cryptocurrency folks. Not a big problem since an agricultural metaphor seems better fitting anyway. In most search algorithms we grow trees with enormous foliage in order to get the solutions, the fruits. The truth is that I got the word from the science-fiction movie Moon.

Naming is an important issue. Once we named a subversion server ‘mordor’, and it killed three hard drives in a row. Since the motherboard-memory combination exhibits the Aussie national colors (my current place of residence), the name mollymook came as an obvious choice.

DSC_3450

My original idea was not to have a case. Just have the pieces on the table. So I ended up with this:

DSC_3456This of course was a very stupid idea. First, an ATX board is a bit bigger than a Raspberry Pi, second the cables are very rigid, so the above configuration was metastable. A little touch at one point and the components started jumping around.

So, a case was needed. As for housing I’m ok with Scandinavian design. No, ikea hasn’t started selling computers cases yet, but Fractal Design has. It was meant to be a joke, but the furniture company is indeed mentioned in the manual.

Impressive 140mm silent cooling fans.

DSC03362In a modern case, installing the components and cable-management is easy.

DSC03381And you end up with a clean build.

DSC03383

For the OS, Ubuntu 14.04 was installed, but this is just an arbitrary choice. For running GAP all you need is a solid UNIX base system and a C-compiler, the details don’t matter.

It is all good, HR mollymook is happily cutting its way through the fabric of the mathematical universe, exploring the unknown, while physically residing in a stylish black&white high airflow box.

Yep, this pseudo-poetic-techno-rubbish metaphor is a good way of ending this buildlog.

Advertisements

GAP on the Pi

August 1, 2013 6 comments

GAP is a system for computational discrete algebra… but I probably should not spend too much time introducing it. Either you know it very well, or you can get all the information from here. I work with it and develop for it every day.  One good thing about it that its kernel is written in standard C, so it is easy to get it working on many different systems.

Raspberry Pi is a little computer that invokes 10% of the feeling of the good old days in the 8-bit era. If you think 10% is a low value, well, today’s computers give 0.1% of the inspiration we experienced back then in 80’s early 90’s, so comparatively Raspberry Pi is doing very well. Again, no need to introduce it. In case you haven’t heard of it, then it is time to catch up. Here or here.

I ordered two of these without knowing what to do with them. One I gave away as a gift, the second one I tried it as a media center. Spectacular success, though my friends were not so excited. The general opinion was that it did not matter on what sort of gadget we watched the movie… Hm, I simply don’t understand people… 🙂

So as a next step I put Arch Linux on it. When I got to the command line (few seconds) I realized that GAP was missing. It takes 106 minutes to compile it form source (mainly because of the GMP library) but then, of course, it works nicely.

Later I got carried away and installed the GUI packages as well.

DSC02844

Here is the Raspberry Pi calculating the holonomy decomposition of the full transformation semigroup on 5 points. As a nontrivial but very simple exercise.

DSC02843Yes, nontrivial calculations require all the processor power, but the chip does not get too hot. Without any cooling…  It is not fast, but requires little power.

DSC02845

Why doing this? Certainly not for the performance gain. Heavy mathematical calculations are really slow on the Pi, but that is exactly the point. When testing new algorithms it is easier to detect performance bottlenecks on a slower machine. We get lazy on big rigs and don’t produce quality code, unlike in the 8-bit era, where lousy programming yielded no result at all. So, if your algorithm works on the Pi then it will not waste kilowatts on the research cluster.

Also, we develop mathematical software and produce mathematical results eventually used in proofs. Computer calculated results have to be checked carefully. Ideally, given an algorithm we would implement it on two different platforms, using different programming languages developed by two independent teams. That is a luxury we simply cannot afford. However, Raspberry Pi provides us the possibility to test the code on a different architecture at least.

Now I’m checking the availability of a C compiler for our washing machine… 😉

UPDATE: I sometimes use this system for software development. Due to its minimality, it is distraction-free.

Computing Partition Monoids – 3 weeks down under

July 24, 2013 Leave a comment

First of all, we have some name clashes, simply because now there are many persons named James in semigroup theory. So to get this right, here are the names appearing in this story, in alphabetical order:

  • JE James East, University of Western Sydney
  • ENA Attila Egri-Nagy, University of Western Sydney
  • JDM James D. Mitchell, University of St Andrews

Pictorially:

DSC02631

This is a story of the latest development of the Semigroups package, but let’s put things into perspective.

In 2004 JDM inherited the Monoid package with the task of turning it into a GAP4 package as it was originally developed for GAP3. This was done, but after that he became unsure whether it was worth investing more time into the development. He felt that there was not much interest.

In 2010 ENA went St Andrews, determined to squeeze out a release from JDM with some new features needed for SgpDec. One man crowd with banners WE LIKE, WE NEED MONOID! RELEASE EARLY! RELEASE OFTEN! Apparently, this was convincing. JDM started to revise the code.

Around this time Monoid was renamed to Citrus. Much to the excitement then later disappointment for orange growers in Florida.

In 2011 JDM came to Eger, Hungary for the cute little algebra conference A^3. He started to write code for SgpDec and consequently making more upstream changes in Citrus.

Here comes the key point in this story. When including new code for partial permutations JDM was faced with a huge amount of copypaste-then-change-a-bit code. This did not feel right. There was an obvious need for more abstract code, meaning that most algorithms for transformation semigroups work for other semigroups as well. This is a very important moment for computational mathematics, as it is a prime example showing that math and computing is not a one-way relationship. Insights in software development can lead to more abstract and cleaner mathematical descriptions.

In 2012 the BIOMICS project started and they got the idea that inviting two developers is twice as good as inviting one, so JDM and ENA again had an opportunity to work together  again (hinting the general semigroup action framework),  again and again.

But it is time for JE appear on the scene. He and JDM met in Portugal (not for the first time) and talked about partition monoids. These are very interesting structures but without any efficient transformation representation, so these are objects that can really put the general framework to test. JDM in his energetic way did some coding to implement partition monoids in Semigroups, but more consultation was needed. This was not so easy since JE and JDM normally live at the other ends of the world.

In 2013, at the University of Western Sydney a new research centre for Mathematics (called CRM) was formed. The Centre have a program for inviting distinguished international visitors to work with UWS researchers. The more people the visitor can work with the better. JDM was an obvious candidate, and the visit did indeed happen between June 30-July22.

We took over the small meeting room for two weeks. Algorithms were discussed and tried on the board first. The whiteboard space was very limited, but nowadays you just take a shot with your mobile phone and erase the board.

DSC02616Then the algorithms were immediately turned into GAP functions and methods and tested against the hand-calculated examples and known mathematical results.

DSC02618JDM’s original estimate was one week, but it actually took 2 weeks to get some tangible results. As software development goes, the estimate was surprisingly accurate.

One example of these new computational results was the D-class picture of the full partition/Brauer monoids. Something that human eyes had never seen before. All in good timing for doing some marketing at the GAIA conference further down in Melbourne. JE gave a talk featuring these new images and there was also a backstage software demo for potential `customers’.

At GAIA we also observed something which must be a generational issue. We got the impression that not long ago mathematicians wrote and sang songs at conferences. What do we do now when we go back to the hotel room? We write code. 🙂 Also, the last row in the lecture room is not used for sleeping any more, it is used for software development. This is of course just an observation, not a judgement, but it seems that academic life was a wee bit more relaxed before.

All in all, as one can see in bitbucket commit messages and in the upcoming preprints, this was a very productive research visit and we also had time for some travel and bushwalks. Maybe a bit taxing on the families, but hopefully the 3 weeks of geek-life will be forgiven.