Re: [Libreoffice] Calc - a better design ...
Hi Michael, I'm hesitant to ask this because I cannot personally promise time toward LO (only on an as-can basis, which is dismally small ATM), but hey, you can easily so no. :-) You mention really need[ing] a whiteboard to elaborate properly. I submit that putting your thoughts together, perhaps in picture form and available on the LO wiki, or put together as a small video to Youtube, would be extremely useful to casual LO coders like myself. As an individual volunteer without a face-to-face LO team member against whom to bounce ideas, I'd thoroughly love an actual-paid-engineer's thoughts on how best to proceed on this front. I'm personally motivated for Calc because in my science career, I really have to bend over backwards to make Calc work effectively for my needs where Excel works just plain better/faster/smaller, yet I philosophically have stuck myself with Free software. To me, one of the biggest areas of weakness for LO, after the various random crashes (which are getting better!), is the memory bloat, and speed. It's not features. I'm happy to mess with Ixion (and indeed have poked at it some already), but I imagine the going would be much more useful with a LO core developer's guiding thoughts, including technical end-goals and migration paths. So, would you have the time to create a screencast or video of yourself in front of a whiteboard? Cheers, Kevin At 9:42am -0400 Wed, 27 Jul 2011, Michael Meeks wrote: Hi Kevin, On Wed, 2011-07-27 at 06:56 -0400, Kevin Hunter wrote: At 6:07am -0400 Wed, 27 Jul 2011, Michael Meeks wrote: but IMHO -the- fundamental design problem with calc, is something quite banal - the concept that a spreadsheet is built from cells: without breaking that basic misconception I don't think we can do any of the really interesting space / time optimisations we need to do. Can you elaborate a little on this fundamental design flaw ? Hah - so, yes - and not easily - I'd really need a whiteboard. As a naive and unfortunately focused elsewhere personality, I don't immediately know of a better model for creating a spreadsheet. So - of course, the first thing to say is that (at a first pass), I'd want the UI to continue to look the same - all that would change would be the underlying model. Is it just a problem of sparsity? Or is there a much more sophisticated method for memory sharing of various similar cell attributes, perhaps analogous to CSS? Here is the thing - since we work on a per-cell basis, all of our type checking, and formula adaptation, and storage, and dependencies etc. are all ultimately per-cell based. This (crudely) means that we have an O(n) where n is the number of cells in vast numbers of operations where we do not want it. It also tends to explode storage and computation time around dependencies - which is a substantial cost. NB. much of this is quite normal for a spreadsheet FWIW, I believe Excel is not completely dis-similar; however I think we can do better with much improved (much more complex / slow) data structure design that will ultimately be far faster to execute. Take a banal example; when we do a SUM() of a million rows containing plain doubles, we would want to (at root) have a function that [ in the ideal case ] does: double sum (double *array, sal_Int64 nItems); Which we can shove onto a gpu, multi-thread if nItems is big, etc. etc. unfortunately approaching this limit is basically impossible in calc. Instead for this case we would do a very slow, pointer-chasing iteration over a million scattered ScCell's - we would do type checking - to ensure that each one was a double, and only then would we add / accumulate them. Of course none of that can be pushed to a GPU, none of it can be SIMD accelerated, and threading it is rather hard. Now ... if we stored contiguous spans of identically typed items [ ie. a column of doubles ], as value runs in our data structures; currently we (amazingly) have arrays of (row/cell*) pairs incidentally. Then we could push a lot of branch-heavy, expensive type checking, and lots of pointer chasing outside the main-loop, we'd also use rather less memory. That's fine for doubles of course; but the really huge wins come from an entirely new model of dependencies and areas containing formulae - I propose only storing a base formula, and an iterator to describe how that formula changes down a row: to compress an entire column of formulae to a single formula. Futhermore on top of that substantially discarding the existing dependency mechanism such that a single-cell change in a contiguous range of doubles would have a shared broadcast on that whole range (with the specific detail of what cell changed), such that that could be turned into a specific row (or row range) to re-compute in any dependent by comparing the base formula, plus it's iterator with the range that changed ;-)
Re: [Libreoffice] Calc - a better design ...
Hi Kevin, On Thu, 2011-10-20 at 02:00 -0400, Kevin Hunter wrote: I'm hesitant to ask this because I cannot personally promise time toward LO (only on an as-can basis, which is dismally small ATM), but hey, you can easily so no. :-) You mention really need[ing] a whiteboard to elaborate properly. That really helps; having said that - I sat down with Eike Kohei to discuss this in Paris, and (I hope) managed to communicate the essence of the idea. I submit that putting your thoughts together, perhaps in picture form and available on the LO wiki, or put together as a small video to Youtube, would be extremely useful to casual LO coders like myself. Sure - so first off, since I'm not actively hacking on calc (as of now), this is really not my call. I tried to persuade Kohei Eike of the intrinsic improvements possible with the new design - if I'm lucky then they agree that I'm not mad might think about that. Of course - I can create a video too, but ... ;-) As an individual volunteer without a face-to-face LO team member against whom to bounce ideas, I'd thoroughly love an actual-paid-engineer's thoughts on how best to proceed on this front. So - the very essence of what I'd like to see happen in calc, and the foundation for it - is to remove the idea that a spreadsheet is a collection of 'Cell' objects. This seems (to me) to be the foundation of our scalability problems. I'm personally motivated for Calc because in my science career, I really have to bend over backwards to make Calc work effectively for my needs where Excel works just plain better/faster/smaller, yet I philosophically have stuck myself with Free software. To me, one of the biggest areas of weakness for LO, after the various random crashes (which are getting better!), is the memory bloat, and speed. It's not features. Right. So the biggest piece (I see) that need tackling here before we can take advantage of the new code is to start restricting the scope of 'ScBaseCell' pointers in LibreOffice calc. Last I looked (which was a while ago) we use ScBaseCell pointers all around the place for things like undo/redo, change tracking, copy/paste, document construction etc. If you wanted to re-start the effort to remove ScBaseCell's mpNote pointer (which is very infrequently used) - that'd be a great place to see some of the problems: ultimately I think we want to remove ScBaseCell (and it's derivatives) entirely - leaving a (numeric) cell as a single 'double' inside a fixed column-array of entries of the same type. Of course, even without the grand vision coming to fruition, saving 4 (or 8) bytes per cell would be worthwhile, and improving the above areas to handle storage of ranges of cell contents in a better encapsulated way would be rather valuable - I think. But of course, you really want to talk to Eike / Kohei / Markus. I'm happy to mess with Ixion (and indeed have poked at it some already) Right - IMHO, the real problem we have is not so much Ixion (which is great), but massaging the existing code into a good shape to be ready for it's heart transplant ;-) The above would be a great step in that direction. Of course, if the calc developers don't object, I'm happy to create a video of me making a fool of myself with a whiteboard too if you think it helps :-) All the best, Michael. -- michael.me...@suse.com , Pseudo Engineer, itinerant idiot ___ LibreOffice mailing list LibreOffice@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/libreoffice
Re: [Libreoffice] Calc - a better design ...
Hi Sounds like a good idea would be to create a master tracking bug in bugzilla around this plan, then split off the different changes into blocking sub-bugs, and mark some of the easier ones so that other people can start doing them. Regards, Noel. Michael Meeks wrote: Hi Kevin, On Thu, 2011-10-20 at 02:00 -0400, Kevin Hunter wrote: I'm hesitant to ask this because I cannot personally promise time toward LO (only on an as-can basis, which is dismally small ATM), but hey, you can easily so no. :-) You mention really need[ing] a whiteboard to elaborate properly. That really helps; having said that - I sat down with Eike Kohei to discuss this in Paris, and (I hope) managed to communicate the essence of the idea. I submit that putting your thoughts together, perhaps in picture form and available on the LO wiki, or put together as a small video to Youtube, would be extremely useful to casual LO coders like myself. Sure - so first off, since I'm not actively hacking on calc (as of now), this is really not my call. I tried to persuade Kohei Eike of the intrinsic improvements possible with the new design - if I'm lucky then they agree that I'm not mad might think about that. Of course - I can create a video too, but ... ;-) As an individual volunteer without a face-to-face LO team member against whom to bounce ideas, I'd thoroughly love an actual-paid-engineer's thoughts on how best to proceed on this front. So - the very essence of what I'd like to see happen in calc, and the foundation for it - is to remove the idea that a spreadsheet is a collection of 'Cell' objects. This seems (to me) to be the foundation of our scalability problems. I'm personally motivated for Calc because in my science career, I really have to bend over backwards to make Calc work effectively for my needs where Excel works just plain better/faster/smaller, yet I philosophically have stuck myself with Free software. To me, one of the biggest areas of weakness for LO, after the various random crashes (which are getting better!), is the memory bloat, and speed. It's not features. Right. So the biggest piece (I see) that need tackling here before we can take advantage of the new code is to start restricting the scope of 'ScBaseCell' pointers in LibreOffice calc. Last I looked (which was a while ago) we use ScBaseCell pointers all around the place for things like undo/redo, change tracking, copy/paste, document construction etc. If you wanted to re-start the effort to remove ScBaseCell's mpNote pointer (which is very infrequently used) - that'd be a great place to see some of the problems: ultimately I think we want to remove ScBaseCell (and it's derivatives) entirely - leaving a (numeric) cell as a single 'double' inside a fixed column-array of entries of the same type. Of course, even without the grand vision coming to fruition, saving 4 (or 8) bytes per cell would be worthwhile, and improving the above areas to handle storage of ranges of cell contents in a better encapsulated way would be rather valuable - I think. But of course, you really want to talk to Eike / Kohei / Markus. I'm happy to mess with Ixion (and indeed have poked at it some already) Right - IMHO, the real problem we have is not so much Ixion (which is great), but massaging the existing code into a good shape to be ready for it's heart transplant ;-) The above would be a great step in that direction. Of course, if the calc developers don't object, I'm happy to create a video of me making a fool of myself with a whiteboard too if you think it helps :-) All the best, Michael. Disclaimer: http://www.peralex.com/disclaimer.html ___ LibreOffice mailing list LibreOffice@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/libreoffice
Re: [Libreoffice] Calc - a better design ...
At 9:42am -0400 Wed, 27 Jul 2011, Michael Meeks wrote: Anyhow - as text, this is not terribly convincing; drawing on a whiteboard would be more so, and with lots of working code - even more so ;-) I keep hoping to have time to play with ixion with Kohei in this regard. Actually, as text it is convincing. So my original analogy to CSS works, but your example of sum(*array) vs for(i = 0...) and making use of SIMD is also telling. Who is working on these particular internals, if anybody? Do I take from your last sentence that the answer is partially nobody? I would /love/ to get my hands dirty on exactly this kind of restructuring and code work ... if only my hands weren't currently tied to a graduate degree that I also enjoy. Sigh. Cheers (and thanks for the extensive reply!), Kevin ___ LibreOffice mailing list LibreOffice@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/libreoffice