Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Prof. Sussman's Reading List

aurellem.org

Post reply on HN
Print view View original

71–78 of 78 posts

Re: Prof. Sussman's Reading List

#71

Post by sedachv »

abecedarius wrote: post #24

On The Connection Machine : "Beautiful thesis, though it doesn't tell you anything you can really do today." I don't understand: is data-parallel computing on a GPU much worse somehow? Or is it that there are better sources to read about data-parallel algorithms?

> is data-parallel computing on a GPU much worse somehow?

OpenCL is a very awkward way to do vector processing - everything is hard-coded to an abstract model of a typical consumer GPU memory hierarchy. CUDA is even worse with a ton of versions all having different limitations according to what the Nvidia chips can do.

It's awkward to do a lot of SIMD tasks on GPUs. The Connection Machine was a general-purpose SIMD originally designed for parallel graph algorithms.

OpenCL looks like what the Connection Machine C* language might get macroexpanded into prior to compilation: http://people.csail.mit.edu/bradley/cm5docs/CStarProgramming...

Re: Prof. Sussman's Reading List

#72

Post by lrc »

tjl wrote: post #55
kjak wrote: post #9

I'm happy to see The Variational Principles of Mechanics (by the great Lanczos) on that list. It really is deep, and I think it does a fantastic job of explaining classical mechanics. I've read the vast majority of the book multiple times. Since we're on the topic of Sussman, has anyone here read through SICM? I've heard that the code is difficult to get to work, but does anyone have an opinion on the rest? I haven't…

I referred to the Lanczos book quite often when I was working on my PhD. I've read the first part of SICM and skimmed the rest. Based on that, it looks like it does an excellent job in building up an understanding of how the math works. By the time I looked at SICM, I had implemented my own code to do something similar (coded in Maple). It looked like the progression of the code was well handled. On the whole, I thin…

Clojure is underway: http://github.com/littlredcomputer/math

Re: Prof. Sussman's Reading List

#73

Post by abecedarius »

sedachv wrote: post #71
abecedarius wrote: post #24

On The Connection Machine : "Beautiful thesis, though it doesn't tell you anything you can really do today." I don't understand: is data-parallel computing on a GPU much worse somehow? Or is it that there are better sources to read about data-parallel algorithms?

> is data-parallel computing on a GPU much worse somehow? OpenCL is a very awkward way to do vector processing - everything is hard-coded to an abstract model of a typical consumer GPU memory hierarchy. CUDA is even worse with a ton of versions all having different limitations according to what the Nvidia chips can do. It's awkward to do a lot of SIMD tasks on GPUs. The Connection Machine was a general-purpose SIMD o…

Thanks. I've read Hillis's book but not studied the modern stuff; I'd gotten the impression the hardware was capable of about as much, and faster now, even if organized differently -- shared memory instead of a network -- and with less-pleasant languages.

Re: Prof. Sussman's Reading List

#74

Post by rurban »

KAM: Automatic Planning and Interpretation of Numerical Experiments Using Geometrical Methods

  * By Kenneth Man-Kam Yip, 1989
  * Coolest PhD thesis ever!
  * Solve problems using graphs.
  * So cool!
This caught my interest: http://dspace.mit.edu/handle/1721.1/7025

It is indeed one of the coolest papers and programs ever. KAM is a smart ODE solver, written in ZetaLisp on a Symbolics. It analyzes 2D pointsets created by any 2d equations, esp. non-linear ones. Typically a system of ordinary or partial differential equations, with a set of boundary and initial conditions. A typical non-linear physical system. It creates MST's (Minimal spanning trees) of the calculated points to get the shape and number of curves, to see the number of clusters (checking the distance of the curves), and if the curves are linear or space filling. Then the phase space is searched for initial states and end conditions, and to get useful summaries. It cannot do shape matching though, so repetitions and mirroring are not detected as such.

The goal is to get high-level descriptions of the model and the numerical dataset, and at which parameter ranges and conditions the system falls into chaos. Chaotic systems are bad for predictability but mostly good for engineering purposes.

Re: Prof. Sussman's Reading List

#75

Post by sedachv »

abecedarius wrote: post #73
sedachv wrote: post #71

Earlier quoted context omitted.

> is data-parallel computing on a GPU much worse somehow? OpenCL is a very awkward way to do vector processing - everything is hard-coded to an abstract model of a typical consumer GPU memory hierarchy. CUDA is even worse with a ton of versions all having different limitations according to what the Nvidia chips can do. It's awkward to do a lot of SIMD tasks on GPUs. The Connection Machine was a general-purpose SIMD o…

Thanks. I've read Hillis's book but not studied the modern stuff; I'd gotten the impression the hardware was capable of about as much, and faster now, even if organized differently -- shared memory instead of a network -- and with less-pleasant languages.

That was pretty much where Thinking Machines was headed before bankruptcy - their third model (CM5) was a SPARC cluster primarily for running Connection Machine Fortran and C*.

Re: Prof. Sussman's Reading List

#76

Post by tjl »

lrc wrote: post #72
tjl wrote: post #55

Earlier quoted context omitted.

I referred to the Lanczos book quite often when I was working on my PhD. I've read the first part of SICM and skimmed the rest. Based on that, it looks like it does an excellent job in building up an understanding of how the math works. By the time I looked at SICM, I had implemented my own code to do something similar (coded in Maple). It looked like the progression of the code was well handled. On the whole, I thin…

Clojure is underway: http://github.com/littlredcomputer/math

That's interesting, but I get a 404 for that link.

Re: Prof. Sussman's Reading List

#78

Post by sir-pinecone »

abc_lisper wrote: post #6
selimthegrim wrote: post #5

Earlier quoted context omitted.

It's not any heavier than GEB, and they will probably get more out of it. My picks in there would be Lanczos, Heinlein, and Levy, and SICP and SICM.

Umm.. I could read GEB; it was tough but more understandable. QCD is way over my head :), but may be because I am dumb.

Hah, I'm sure a great number of people would say being able to read GEB makes you not dumb.
Post reply on HN