?

Log in

No account? Create an account

Previous Entry | Next Entry

The SuperComputing exhibition opens

Well, I woke up at 6am local time, and decided to try to make a start on my day's noveling, since I knew I wouldn't have much time once the day's events started. I forced myself to write 200 words before I got the shuttle to the conference centre, where the internet awaited me.

It was great to catch up with people on ToothyChat - my 8am was their 4pm, so I got to say hi to people before the day-long tutorial started, on Application Supercomputing Concepts.


The first section was both boring and boggling, as we were given a whistle-stop tour of the hardware architectures of parallel processors. Details of how many different caches are between the "fabric controller" and the "memory controller" are necessary for some people to consider, but I'm not one of them.

The next section covered parallel benchmarks, which I'm already vaguely familar with; parallel debugging, which was quite interesting because it's quite a hard problem, and much less deterministic than debugging a serial program; and parallel profiling, which is something directly relevant to my job.

Then we had a section on parallel programming models - the basic concepts of task parallelism, data parallelism, communication in shared memory (OpenMP) or distributed memory (MPI) - all of which was stuff I was familiar with from the Parallel Computing 101 tutorial I took at SC05, two years ago; but it was interesting to see some sample code for Titanium (a recent parallel language based on Java). A half hour on optimisation for parallel processors later, and it was off to lunch.

After lunch, one of the designers of MPI arrived to explain MPI in mind-numbing detail to us. He clearly hadn't been in the morning's part of the tutorial, as he repeated some stuff from there; and when he asked for a show of hands, only about 2 of the 40-50 people there had never seen MPI before, but he still went through it at glacial speed. It was interesting stuff as far as it went, it's just that it was already familiar and could have been covered in one third of the time.

But then the final two sections of the tutorial were much better. There was a large section describing lots of real-life applications of parallel computing, which I've always found one of the more interesting bits of the Supercomputing conferences. One of the interesting applications shown was in NASA's climate modelling, predicting the path of Hurricane Katrina. Three computations with slightly different algorithms predicted three spots with 50km between them; but one of them was just 14m away from where Katrina actually went. Stats for those who like them: NASA's Columbia machine (from 2004) has 10240 Itanium processors, 20 terabytes total memory, and achieves 51 TFLOPS.

A fun quote from this section: "Things we can do now in a day would have taken 30,000 years on computers from 10 years ago."


I had an hour or so between the tutorial and when the exhibition floor opened, which I spent doing a little writing (littler than I'd like). Then I wandered around the exhibition floor, getting horribly, horribly lost. At least it's all on one floor this year, but what a floor... I kept losing my bearings and finding myself walking back the way I'd come, which is not a state I find myself in often.

The exhibition floor is quite a sight: high-tech computer companies as far as the eye can see, with eager salesmarketing people dying to talk to you about their chips/server/networking/scheduling software. There seem to be fewer "cool things" than I remember from SC05 (like one company had a Porsche taking up half of their booth space). There are at least three companies giving away games consoles if you read details of their products. Intel are doing some silly competition to celebrate their 45 nanometre something-or-other, where they give away 45 prizes in 45 minutes each day. Microsoft appear to have decided to make stage magic their theme this year, with prestidigitators, a live improv troupe, and suchlike. There seem to be even more Japanese / Koreans than there were last year, including many Japanese universities showing off what they've been doing.

Just one more free gift to report, since I haven't bothered doing the silly treasure hunts that Sun or Microsoft send you on to get their T-shirt (and assuming you don't count the couple of free sweets I took from a couple of companies to which I'm grateful but whose names I don't remember): this is a rather funky pen, with a flashing blue LED inside it. It's cool and all, but it's just a pen. 3/5.

Novel-writing has been going pretty slowly. I've introduced a new POV character for the current chapter (Chapter 10), but I'm not particularly expecting to follow him again once this chapter is up (although you never know...) I continue to write along with only about one chapter's worth of plans ahead of where I am. But the big plot is going to kick in soon, which will be fun. (And I want to find ways to tie what's been happening so far into that main plot, otherwise it'll feel a bit superflous. Every scene has had a definite role it's played - introducing a character or bringing an important meeting or revelation - so it's not like any of it is just padding, but I do want to get it all tightly tied together.)

I've just passed 16670 words, so I decided I'd come here to post this, then head back to the hotel, where I won't be able to be distracted by the internet. Wish me luck!

Tags:

Comments

( 8 comments — Leave a comment )
wikivic
Nov. 13th, 2007 08:15 am (UTC)
luck! and congrats on the word count ^^
ex_robhu
Nov. 13th, 2007 09:41 am (UTC)
At least it's all on one floor this year, but what a floor... I kept losing my bearings and finding myself walking back the way I'd come, which is not a state I find myself in often.
I was amazed to discover that the PDF floor plan (which is fuzzy and hard to read at 100%) doesn't improve when you zoom in because the PDF just contains a fuzzy bitmap!

The 45nm thing is moderately exciting to me, I've been following Ars Technica's coverage. Roll on the non-enthusiast chips.

Microsoft appear to have decided to make stage magic their theme this year
Does anyone take them seriously at such a conference? I remember when they launched 'Microsoft Compute Cluster Server' which seemed to be a bit of a joke.

alextfish
Nov. 15th, 2007 02:46 am (UTC)
Microsoft CCS is quite a serious program - we spent a lot of money adding support to it to our distributed computing tools, since a lot of our customers use it. Out of the customers I've spoken to at SC07, a significant proportion of them use it and/or are significantly interested to hear that we integrate with it.

MS are hard not to take seriously, given the amount of money they can throw at anything. But I was just mentioning today how at SC05, when Bill Gates gave the keynote talk, there was a bit of a love-him-or-hate-him feeling.

This year they seem to have decided to encourage people not to take them seriously, though. Today they had an Austin Powers lookalike going around saying that Microsoft are "groovy, baby".
ex_robhu
Nov. 15th, 2007 07:58 am (UTC)
OK, I guess they must have a certain amount of the 'medium end' clustering market. They more or less don't exist at the high end, with only 1.2% share of the Top 500, compared with 85% for Linux.

I suppose I don't really understand why anyone who had a sizable cluster would want to use Windows. From my experience the reasons for using Windows on the server are:
1. The software you want to run is only available for Windows
2. You are a 'Windows only shop'

I encountered a big number of reason 2 companies in my previous job. There were more of them around in the SME market than in the FTSE end of the market. For reasons that should be all too obvious I think that reason 2 is rather stupid.
alextfish
Nov. 15th, 2007 09:49 pm (UTC)
Yeah, they operate in the medium-size cluster market, which is where MathWorks focus as well: most of our customers have clusters of size 8-64. People don't tend to run MATLAB on the BlueGene/L, although we'd quite like it if they did :)

I don't know much about why people choose LSF/PBS/SGE/CCS for their cluster software, or indeed Linux/Windows for their cluster OS. I might ask around.
ex_robhu
Nov. 15th, 2007 09:53 pm (UTC)
Ack! LSF!!!!

LSF is so bad people write their own replacements.
woodpijn
Nov. 13th, 2007 09:57 am (UTC)
"prestidigitators" - what an utterly cool word. I had to look it up. I'm tempted to dare you to use it in your novel, but maybe you already have.
alextfish
Nov. 15th, 2007 06:45 pm (UTC)
Ooh! I haven't, yet, but I should do :)
( 8 comments — Leave a comment )