Mapping the point-spread function over the image

[Darth Takahashi]

So I guess no one is looking to build a PlayStation cluster then?:(

[michael.h.f.wilkinson]

So I guess no one is looking to build a PlayStation cluster then?:(

By all means build one. But why do things the hard way, when anyone with a computer suitable for MS-Office 2007 or above has at least the equivalent processing power of a Cray Y-MP on his desk? As Niklaus Wirth said: "Software is getting slower faster than hardware is getting faster."

A cluster has tons of communication overhead. Shared memory parallel computing (e.g. using multi-core CPUs) is a lot more efficient in many cases. In any case, the images we are using are not big. They fit in memory easily, so distributed memory processing is not really necessary or indeed helpful.

[themos]

They fit in memory easily

Ah, but do they fit in the fast caches?

A cluster has fewer shared resources and should go through this workload faster. There's not much communication needed either as the problem is "embarrassingly parallel". They're just harder to program. The reason we use multi-cores with shared memories is that this is where Moore's law is operating at the moment, not because it's a good idea. Multi-core shared memory processing is a sign of desperation.

[michael.h.f.wilkinson]

Ah, but do they fit in the fast caches?

A cluster has fewer shared resources and should go through this workload faster. There's not much communication needed either as the problem is "embarrassingly parallel". They're just harder to program. The reason we use multi-cores with shared memories is that this is where Moore's law is operating at the moment, not because it's a good idea. Multi-core shared memory processing is a sign of desperation.

The filters I am working on are not embarrassingly parallel. You can get the PDF of our main article on parallel implementation here:

http://www.cs.rug.nl/~michael/pamiparmaxtree.pdf

A variant is given here:

http://www.cs.rug.nl/~michael/OuzounisISMM2007.pdf

Since those papers we have tested it on 24-core machines and get about 14x to 18x speed-up on 24 cores.

[themos]

The filters I am working on are not embarrassingly parallel.

Ok, but the wider workload is, in the sense that one typically wants to examine a lot of images that, data-wise, are completely independent of one another. But I'm not going to teach you what you already know!

[Darth Takahashi]

Because they have this nice Cell CPU and they are relatively cheap and I have never played with Parallel processing before... And I almost forgot that you can install Linux on them, right?

PlayStation 3's Cell CPU achieves a maximum of 230.4 GFLOPS;)

I was thinking about using 3/4 of them in a network?

But if there is no real advantage then Ok I'll forget it.

Yes, please don't get in on my soap box where PC hardware and and the dreaded MS software is concerned!!!:eek: And the average laptop that's is doing NOP's 98% of the time, right!!!

This there a FORTRAN equivalent of a NOP?

[michael.h.f.wilkinson]

Ok, but the wider workload is, in the sense that one typically wants to examine a lot of images that, data-wise, are completely independent of one another. But I'm not going to teach you what you already know!

That is true. Spreading things out over a cluster works fine (no complex programming needed, just script the lot). We do that with scanned images of pages of manuscripts. The university has a cluster with a total of 3280 cores, and a Blue Gene L with some 12,000 cores, but both are often rather full up with guys folding protein molecules. A few colleagues and I are putting a mini cluster of 2-4 48 core machines together which only we have access to

[NickK]

The filters I am working on are not embarrassingly parallel. You can get the PDF of our main article on parallel implementation here:

http://www.cs.rug.nl/~michael/pamiparmaxtree.pdf

Reading that brings back memories

[michael.h.f.wilkinson]

Reading that brings back memories

??

[NickK]

??

Lol, not the subject matter but the language!

[Darth Takahashi]

So how does this move forward? Are there any adaptive filters that can already be used in a suitable environment (IDE) etc...

[michael.h.f.wilkinson]

So how does this move forward? Are there any adaptive filters that can already be used in a suitable environment (IDE) etc...

In early February, our 3rd year students must find a Bachelor thesis topic. I am drawing up a proposal which could be chosen. Regarding code: we have quite a bit of C code for fast attribute filters. Our students will build on that.

[NickK]

Did anything come of this Michael?

I'm currently just prodding the single PSF deconv on the GPU, so I just wondered how the image wide PSF estimation was going.

[michael.h.f.wilkinson]

Two students did make a prototype that could estimate aberrations over the image (coma, astigmatism, pinched optics), but that is not yet in a format that I would publish. I hope a few students will pick up the project and continue the development.