Exascale by 2018: Crazy ...or possible?

I recently saw some estimates that show we should hit exascale supercomputer performance by around 2018. That seems a bit ambitious – if not stunningly optimistic – and the search to get some perspective led me on an hours-long meander through supercomputing history, plus what I like to call “Fun With Spreadsheets.” Right now …

[Mr Finance]

Is this relevant anymore or is it just superpowers comparing cock sizes? what, if any, is the maximum performance available on a cloud system like EC2? And are there any real world applications that actually need to run on a single system rather than a data center running a decent cloudOS (I do understand security will prevent running nuclear weapons modelling etc on externally networked cloud systems)

[JDX]

Well I'm no expert but isn't a lot of the work that goes on about how quickly the parts of the system talk to each other as well as how fast each bit works? If so, then distributed processing introduces a massive paradigm shift in how important it is to partition problems to support such methods.

But in global terms... you're probably right for many types of problem.

[pawatan]

Relevant

It's relevant. The oil industry uses dedicated supercomputers (the modeling is very, very compute intensive) and due to data security issues (and efficiency issues) that sort of computing and storage will never be on the cloud.

[Anonymous Coward]

all this

Maximum performance, max speed and mines bigger than yours talk.

And yet what do we apply it too?

Watching porn and playing games!

Wonderful.

[Ian Bush]

There are a number of reasons why clouds can't meet all needs, at least in the scientific area. One is that many scientific codes need the low latencies that these machines provide to scale to

anything like the number cores they contain, and the network in clouds just don't cut it. There are also I/O and other issues. For a recent report take a look at

https://www.nersc.gov/assets/StaffPublications/2012/MagellanFinalReport.pdf

there's a key findings section of about a page and a half near the beginning.

[Michael H.F. Wilkinson]

I could use such power

We are looking at terapixel scale image processing. One aim is compute the entire aerial photography data set of the Haiti earthquake for damage assessment. We aim to do this in a few hours. This can still be done at the multi gigaflop rates provided by decent size clusters. For the 2004 tsunami, we would need orders of magnitude more compute power to handle all measurements required (we are talking petapixel range). These are computations that cannot possibly wait. Though exaflop processing might not be needed yet, as image resolution increases, so do the data rates. Compute power will have to keep up. Real-time emergency response requires real-time processing of vast amounts of data. The size of each compute node of the EC2 and similar cloud systems is far too small for our purposes (we need to have at least 32 cores (preferably 64) cores per node and between 128 GB and 1 TB of RAM per node).

So, yes, I can see real-world, life and death applications for these kinds of compute powers. Other applications include large scale scientific simulations (think evolution of galaxies, or the cosmic web, but also earthquakes, ocean currents, or even the blood flow through a sizeable chunk of the vascular system).

[Mr Finance]

Re: I could use such power

Very interesting examples, thanks. Still not sure I understand what prevents these tasks from being more granularly handled by vast nunbers of the smaller commercial compute nodes. Is that a historical artifact of the way software has been written for these high performance real time tasks, or is it something more fundamental that make granular computing inherently inefficient (eg high calculation ci-dependencies that result in low compute utilisation while waiting for outcome of dependant calculation running on another (even physically co-located) server. Assuming in the main it is a historical software artifact that will get re-written over time, are there un fact any sorts of application that have sufficiently granular calculation co-dependency to outweigh cost advantage of cots hardware per flop? Appreciate your thoughts.

[Michael H.F. Wilkinson]

Re: Re: I could use such power

Many image processing tools are easily broken up into small loads. The tools we develop are so-called connected filters in which data from all parts of the image may influence all other parts. Often you want to avoid this, but e.g. for analysis of complete road networks, you cannot predict which parts need to communicate with which other parts. To cut communication overhead down you need comparatively coarse grained parallellism.

We have been able to device a scheme (which we are now implementing) which cuts down local memory requirements from O(N) to O(N/Np + sqrt(N)), with N the number of pixels and Np the number of processor nodes. Normally you just require O(N/Np) per node. Communication overhead can be dropped from O(N) to O(sqrt(N) log(N)). This has moved the problem from the impossible to the "possible, but you need a coarse-grained cluster."

Of course, the first (shared memory) parallel algorithm for this filters dates from 2008, so we still have much to learn. Other problems in science can also require global interaction between particles (gravity has infinite reach). A lot of work is done cutting communication down, but this is often at the expense of accuracy.

[Mr Finance]

Re: Re: Re: I could use such power

Thanks Michael, helpful and interesting

[Boris the Cockroach]

How many flops

does the BOINC distributed computering thing deliver as total?

[Anonymous Coward]

Re: How many flops

The oracle that isn't Oracle says: "The entire BOINC network averages about 5.3 PFLOPS as of July 25, 2011"

[Anonymous Coward]

Accuracy is important when measuring your epeen

Then at least do it right.

Before we exceeded 1 megaflops, guess what:

there was a point in time when computing achieved a kiloflops, and before that, somewhere along the line, the effective equivalent of a flops. Both of these I'm pretty sure came sometime some time between 1940 and 1964 - a rather more relevant period than your psuedo-random-dynasty.

And tell me, what is a megaflop? a million FLoating point OPerations that's what, now how fast would you like them dished up, hmm?

What? You want them in the 1/60th of the period between Dan-Olds-equivalent births; I'm sure there's a proper reg unit for it, but what i'm getting at is it's megaFLOPS you fool. With an S. for per Second. Gah - it's as bad as people writing about energy who don't know the difference between a Watt and a Watt-Hour.

And then, finally, after all that, here's a little thought experiment that doesn't even need a spreadsheet. If we surpassed 1 Mega (10^6) FLOP/s in 1964, and 1 Peta (10^15) FLOP/s in 2008, then how many FLOP/s do we need in this context to reach the next milestone?

Here's a hint: it's not 100 PETAFLOPS. I swear people who failed high-school maths, and somehow still got into computing, called themselves consultants, and managed to blag themselves some sort of profile ... were put on this earth just to piss me off.

El Reg - really this much basic fact/comprehension/math fail in one article, and I bet you'll still post the author a cheque. Bah!

[Shannon Jacobs]

Re: Accuracy is important when measuring your epeen

990 <> 90, eh? I was wondering how he did that, so I expected to see prior comments on it...

[danolds]

Re: Accuracy is important when measuring your epeen

In the original version of the story, there was a typo near the top saying that we needed an addition 90 pflops to hit exascale - it should have been 990 pflops. It was corrected as soon as we saw it. I think that covers at least part, if not all, of your rant.

For the rest of it, I agree, we certainly did have 'flops' before we had mega, giga, tera and petaflops....and we had 'flopm' (Floating Point Operations Per Minute) too, right? Although those were probably measurements of humans with pencils or chalk, I guess I should have factored those in to be complete.

Oh, and you misspelled 'cheque' in your last line. The correct spelling is 'check', as in 'the check is in the mail' or 'you should check your watch to make sure you're on time for your upcoming court-mandated anger management session.'

[Anonymous Coward]

Completely ignorant question

How does this compare to the stuff the rest of us are using?

How long will it be before my mums buying a laptop computer/tablet from Currys with the power that was available in super computers of yesteryear?

[Ian Bush]

Re: Completely ignorant question

Depends on what you have (obviously), but in the region of 10 GFlops will be the right sort of ball park.

[Anonymous Coward]

my sister-in-law...

...is working on an exascale project the Edinburgh Uni are involved in. My brother insists on calling it 'giggle-flaps' - much to my amusement.