END OF AN ERA, FRACTALFORUMS.COM IS CONTINUED ON FRACTALFORUMS.ORG

I have 3 GiB of RAM and I think thats causing me problems with 1680x1050 8x8 supersampled pictures.
Watching the task manager when the mandelbrot starts computing 1,100,000 (ish) k bytes is used.
Then the CPU usage drops to 50 % (ish) so I assume Mandel Machine is doing something single threaded.
Soon the initialising series approximation message shows on the status bar and the memory increases to 2,300,000 (ish).
The render is then much slower than before.
Perhaps Mandel Machine is having to search through 1GiB of memory assigned by the series approximation?
Another round of approximation is needed to fill in the details and thats where I give up.
The memory keeps climbing up to a maximum of 2,385,000 k bytes then the CPU falls to 0 and the memory falls back to 2,315,000.
Over and over, perhaps Javaw.exe is hitting an allocation limit, has to garbage collect to free memory and keeps going through this cycle?
I've attached the mmf parameter file is you're interested, removed the .txt extension to load into Mandel Machine.

Hi Botond

I have a linux machine (running CentOS 6.5) and I copied Mandel Machine to it with the intention of getting MM running under Wine.  Surprisingly, it didn't run (I think it is a problem with the Java Machine you supply) but out of curiosity I installed Sun's Java Runtime and then clicked the jar file, and It burst into life.

There seem to be some issues with the system not detecting all glitches, but I can live with that.

Is the program working perfectly?  Not sure.  But so far all signs are positive.

Which location?

Oh, I remember, that is a tough location! I could only raise the iteration count to 250 millions (I have 10 GB RAM), and it is still too low to display anything but a flat area.

But stay tuned, the next release of Mandel Machine will be focused around a cleverer series approximation, and one pleasant side-effect will be reduced memory footprint!

Writing the reference orbit to disk could help, but it is unnecessary. In the next version of MM, only the iterations after series approximation will be kept in memory, reducing the memory consumption by at least 90% on average.

Efficient series approximation is key to fast perturbation calculation, even more than I previously thought. Take this location for example:
http://stardust4ever.deviantart.com/art/Magnum-Opus-Ex-3132-6-Zooms-274474754

This was rendered at 3200x3200 resolution in 28 days using Fractal Extreme, using 4-8 core AMD machines. The next version of MM renders it on my dual-core Ivy Bridge laptop in just under 30 seconds. Taking into account the differences in hardware, the speedup is more than 100000x 

That sounds amazing! 

The location has minimum 377372 and maximum 378486 iterations.
The difference is 1114, so I guess you need to calculate at least that amount?

KF can skip 310508 iterations, but renders this in 640x360 in 6 minutes, which is "only" 150 times faster than FX, not considering hardware...
Long double is used for this location, which breaks KF if the maximum number of iterations is set lower than 310508

I made Mandel Machine v1.2 available. It is still in development, so this is a beta version. Some of the promised features (reduced memory footprint, true unlimited zoom depth or Pauldelbrot's new glitch correction method) are still missing, and I also had to disable some previously working features temporarily. Those will be available in the final release.

This version is focused around two key areas: maximum zoom depth is increased to 3700 (this required the implementation of extended precision routines), and series approximation is adjustable. It can be made more aggressive than before, using up to 130 terms. The cost of initializing SA can expressed as L_ref*N_terms², where L_ref is the length of the reference orbit and N_terms is the number of terms (coefficients) used in the approximation. As this cost is independent of image size, increasing the number of terms above a certain limit makes sense only by huge images where the savings outweigh the increased cost. At small to medium image sizes a right balance has to be found. It might also be automatable in the future.

A full list of changes and the download link can be found on my Mandel Machine site:
http://web.t-online.hu/kbotond/mandelmachine/

The number of coefficients is adjustable between 6 and 130. The higher this number is, the more iterations can be skipped. It also depends largely on the actual location.

Errors are no longer measured. The number of skipped iterations is determined based on the coefficients only. The basic idea is that if we have N coefficients, X iterations can be skipped safely if the Xth iterations of the coefficients can be divided into two groups that fulfill the following criteria:

• The first group contains the first M coefficients (A_1,X, A_2,X ... A_M,X), the second group the remaining N-M coefficients (A_M+1,X, A_M+2,X ... A_N,X)
• Every term based on the coefficients of the first group is several orders of magnitude larger than any term based on the coefficients of the second group. A term for the coefficient A_i,X is A_i,X * delta₀ⁱ

As long as the coefficients can be divided into these two groups, the perturbed orbit can be approximated with the first M terms safely.