- garbage collection - 1 Update
- decltype( a )::value_type - 2 Updates
- What is computing for me ? - 1 Update
- I correct some typos because i write fast, read again.. - 1 Update
| Tim Rentsch <txr@alumni.caltech.edu>: Aug 05 07:45AM -0700 >> disadvantageous properties: it will not reclaim cyclic structures >> unless they explicitly have the cycles removed; > Using std::weak_ptr might provide some mitigation here. Sure. But that starts to move into the domain of doing manual management. One needs to think about where to use the weak pointers: not enough, and cycles don't get reclaimed; too many, and the data structure falls apart too early. That doesn't mean reference counting is hopeless -- any non-trivial application does some level of manual management, and even full-scale mark/sweep environments need weak pointers in some circumstances. The important thing is that making sure everything gets reclaimed incurs a mental cost, and that cost is higher for reference counting than it is for mark/sweep-style collectors. > heavyweight because the reference-count update is required to be > multithread-safe. A single-thread only smartpointer is much faster in > multithreaded programs, but its usage obviously requires more care. (Disclaimer: I'm not a unique_ptr expert. I did look into the issues here, and believe my comments below are correct, but I very well may have missed or misunderstood something. All feedback welcome.) First, my comment was about reference counting, and unique_ptr is not reference counting. Or it may be thought of as 1-bit RC -- it either has a non-null pointer or it doesn't (and the "count" is in the unique_pointer itself, not in shared per-object memory). In any case it is definitely less automatic than full RC (which I take as being equivalent to shared_ptr, with a similar disclaimer). As such it requires more mental effort than using shared_ptr or some other more complete RC scheme. Second, regarding resource use. The idea that unique_ptr has no space or time overhead is, I believe, incorrect. In particular, when updating a unique_ptr from a different unique_ptr (temporary), there is a space cost in the form of code space needed for a value check, and a time cost in checking the value to see if a destructor needs to be called. Those may not be large costs, but they are not zero. Compare to a mark/sweep-style environment, where all that needs to happen is copying the pointer. Third, besides needing more manual management, using unique_ptrs carries an additional mental cost in having to understand the more elaborate language semantics that they rely on. In particular, unique_ptr has move semantics but not copy semantics. The kinds of code patterns used for "normal" objects (ie, that do have copy semantics) in many cases simply don't apply to unique_ptrs. To use unique_ptrs requires a new kind of thinking, which means more mental cost. I'm not saying these things can't be learned -- obviously they can be. But there is an initial cost in learning, and more importantly an ongoing cost in using (cost here meaning mental effort) things like unique_ptr that don't use the usual semantic constructs. If that use could be tucked away underneath a layer of abstraction that would be one thing, but with unique_ptr it's right there in your face and there's no getting around having to deal with it. Is that a good tradeoff to make? I expect in some cases it is. I believe though that in many cases a different tradeoff would be better. >> lead to hard-to-understand behavior in the presence of exception >> processing. > Not sure what you want to say here? [...] Invariants maintained locally and incrementally are often harder to get right and harder to understand than invariants maintained centrally. >> big increase in productivity, which has been measured as a >> factor somewhere between 1.5 and 2. > Compared to what? Compared to doing all memory management manually. > To the proper C++ code using std::vector, > std::string, std::make_unique, std::make_shared, or to the C or > C-style C++ code calling malloc/free or new/delete manually? Even just using unique_ptr and shared_ptr, C++ provides a form of reference counting, which is more automatic management than having to manage all memory (de-)allocation manually. I don't know of any studies that compare the productivity results of using reference counting versus a more centralized form of garbage collection such as mark/sweep. Based on personal experience I expect that some level of increased productivity would be found, but it would be nice to get some more objective results on that. > I'm sure GC suits fine some programs. Yes, and the question is which programs (and which classes of programs). I would never claim that any approach is right for all applications. But I think many people, or maybe even most people, form opinions about that based on pre-conceptions that turn out not to be right if examined more closely, so I like to encourage people to look at things from different points of view, and see if their perceptions might change on further investigations. |
| Juha Nieminen <nospam@thanks.invalid>: Aug 05 07:00AM >>> #include <ostream> >> Rather redundant, don't you think? > With C++11 and later, yes. Given that the code is C++11, it is thus literally redundant. |
| "Alf P. Steinbach" <alf.p.steinbach+usenet@gmail.com>: Aug 05 10:45AM +0200 On 05.08.2018 09:00, Juha Nieminen wrote: >>> Rather redundant, don't you think? >> With C++11 and later, yes. > Given that the code is C++11, it is thus literally redundant. That's a good point. I was thinking mainly that Stefan's conventions for his course materials possibly predate C++11. Cheers!, - Alf |
| Sky89 <Sky89@sky68.com>: Aug 05 01:35AM -0400 Hello.... What is computing for me ? It is first finding an efficient "abstraction" and being able to hide or reduce "complexity", and also Loose coupling is also like hiding or reducing complexity, i am always doing it on my software projects, and to be able to do it more efficiently you have to be more smart, also i am always looking to reduce efficiently the serial part of the parallel program, i am calculating it by CPU cycles and this part of my work is like "assembler", and this to be able to "predict" more correctly scalability, look for example at my Parallel Compression Library or my Parallel archiver here: https://sites.google.com/site/scalable68/parallel-compression-library https://sites.google.com/site/scalable68/parallel-archiver Notice that it says: - It's NUMA-aware and NUMA efficient on windows (it parallelizes the memory reads and writes on NUMA nodes) and - It minimizes efficiently the contention so that it scales well and Now it uses only two threads that do the IO (and they are not contending) so that it reduces at best the contention, so that it scales well. and - Now it supports processor groups on windows, so that it can use more than 64 logical processors and it scales well. And notice that it says the following: I have done a quick calculation of the scalability prediction for my Parallel Compression Library, and i think it's good: it can scale beyond 100X on NUMA systems. The Dynamic Link Libraries for Windows and Dynamic shared libraries for Linux of the compression and decompression algorithms of my Parallel Compression Library and for my Parallel archiver were compiled from C with the optimization level 2 enabled, so they are very fast. So as you noticed i am making my Parallel Compression Library and Parallel archiver efficient. Also computing for me is also finding a good and more efficient way to ensure reliability and stability, and this part of my work has been enhanced more by my "experience" in computing and by implementing more serious softwares. Also computing for me is also making my software portable, this why i said before the following: About portability of my software projects I have thought more, and as you have noticed i have written Intel assembler routines for 32 bit and 64 bit for atomically incrementing and and for atomically CompareExchange etc. so now they are working with x86 AMD and Intel processors for 32 bit and 64 bit, but i will soon make my Delphi and FreePascal and C++ libraries portable to the other CPUs like ARM(for Android) etc. for that i will use the following Delphi methods for Delphi: http://docwiki.embarcadero.com/Libraries/XE8/en/System.SyncObjs.TInterlocked.CompareExchange and http://docwiki.embarcadero.com/Libraries/Tokyo/en/System.SyncObjs.TInterlocked.Exchange And I will use the same functions that you find inside FreePascal, here they are: https://www.freepascal.org/docs-html/rtl/system/interlockedexchange64.html and https://www.freepascal.org/docs-html/rtl/system/interlockedexchange.html and https://www.freepascal.org/docs-html/rtl/system/interlockedcompareexchange.html and https://www.freepascal.org/docs-html/rtl/system/interlockedcompareexchange64.html I will use them inside my scalable lock that is called scalable MLock that i have "invented", so that it will be portable, here it is: https://sites.google.com/site/scalable68/scalable-mlock And when my scalable MLock will become portable on Delphi and FreePascal i will port with it all my other libraries that uses atomically increment and decrement etc., so my libraries will become portable to the other CPUs like ARM for Android etc., so i think you will be happy with my work. Also what is computing for me ? Read the rest to understand better my work: What about the today computing ? You have to know me more.. I am not thinking becoming an expert of "coding".. I am not like that.. Because I am an "inventor", and i have invented many scalable algorithms and there implementations to do better HPC(high performance computing), i am thinking NUMA systems, and i am thinking "scalability" on manycores and multicores and on NUMA systems etc. this is my way of "thinking", and as a proof look at my new scalable reference counting with efficient support for weak references, here it is: https://sites.google.com/site/scalable68/scalable-reference-counting-with-efficient-support-for-weak-references As you have noticed it is "fully" scalable reference counting, so this is HPC(high performance computing) , and i have implemented this scalable algorithm that i have "invented" in Delphi and on the Delphi mode of FreePascal , so that to make Delphi and FreePascal "much" better, and notice with me that you will not find it on C++ or Rust. This is my way of thinking , i am "inventing" scalable algorithms and there implementations. And i said the following: "I think that this Parallel ForEach and ParallelFor are like futulities, because they don't bring "enough" high level abstraction to consider them interesting, because i think my Threadpool with priorities that scales very well is capable of easily emulating Parallel ForEach with "priorities" and ParallelFor with "priorities" that scale very well, so no need to implement Parallel ForEach or Parallel For." But to be "nicer", i think i will soon implement both Parallel ForEach with "priorities" that scales very well and ParallelFor with "priorities" that scales very well using my Threadpool with priorities that scales very well, and they will be integrated as methods with my Threadpool with priorities that scales very well, so that you will be happy. So i will ask you ? where will you find my Threadpool with priorities that scales very well? and where you will find my Parallel ForEach and Parallel For with priorities that scales very well ? You will not find them on C++ and you will not find them on Rust, because i have "invented" them, because i am an "inventor", and this is my way of thinking. Here is my powerful Threadpool with priorities that scales very well, read about it and download it from here: https://sites.google.com/site/scalable68/an-efficient-threadpool-engine-with-priorities-that-scales-very-well It is a very powerful Threadpool, because: More precision about my efficient Threadpool that scales very well, my Threadpool is much more scalable than the one of Microsoft, in the workers side i am using scalable counting networks to distribute on the many queues or stacks, so it is scalable on the workers side, on the consumers side i am also using lock striping to be able to scale very well, so it is scalable on those parts, on the other part that is work stealing, i am using scalable counting networks, so globally it scales very well, and since work stealing is "rare" so i think that my efficient Threadpool that scales very well is really powerful, and it is much more optimized and the scalable counting networks eliminate false sharing, and it works with Windows and Linux. Read the rest: Read the rest: You have to understand my work.. I have invented many scalable algorithms and there implementations, here is some of them that i have "invented": 1- Scalable Threadpools that are powerful 2- Scalable RWLocks of different sorts. 3- Scalable reference counting with efficient support for weak references 4- Scalable FIFO queues that are node-based and array-based. 5- My Scalable Varfiler 6- Scalable Parallel implementation of Conjugate Gradient Dense Linear System Solver library that is NUMA-aware and cache-aware, and also a Scalable Parallel implementation of Conjugate Gradient Sparse Linear System Solver library that is cache-aware. 7- Scalable MLock that is a scalable Lock. 8- Scalable SeqlockX And there is also "many" other scalable algorithms that i have "invented". You can find some of my scalable algorithms and there implementations in Delphi and FreePascal and C++ on my website here: https://sites.google.com/site/scalable68/ What i am doing by "inventing" many scalable algorithms and there implementations, is wanting to make "Delphi" much better and making FreePascal on the "Delphi" mode much better, my scalable algorithms and there implementations are like HPC(high performance computing, and as you have noticed i said also: You will ask why have i invented many scalable algorithms and there implementations? because also my work will permit us also to "revolutionise" science and technology because it is HPC(high performance computing), this is why i will also sell some of my scalable algorithms and there implementations to companies such as Google or Microsoft or Embarcadero. Also HPC has revolutionised the way science is performed. Supercomputing is needed for processing sophisticated computational models able to simulate the cellular structure and functionalities of the brain. This should enable us to better understand how our brain works and how we can cope with diseases such as those linked to ageing and to understand more about HPC, read more here: https://ec.europa.eu/digital-single-market/en/blog/why-do-supercomputers-matter-your-everyday-life So i will "sell" some of my scalable algorithms and there implementations to Google or to Microsoft or to Embarcadero. I will also enhance my Parallel archiver and my Parallel compression Library that are powerful and that work with both C++Builder and Delphi and to perhaps sell them to Embarcadero that sells Delphi and C++Builder. Also I will implement soon a "scalable" Parallel For and a Parallel ForEach.. This why i said before that: "I think that this Parallel ForEach and ParallelFor are like futulities, because they don't bring "enough" high level abstraction to consider them interesting, because i think my Threadpool with priorities that scales very well is capable of easily emulating Parallel ForEach with "priorities" and ParallelFor with "priorities" that scale very well, so no need to implement Parallel ForEach or Parallel For." But to be "nicer", i think i will soon implement both Parallel ForEach with "priorities" that scales very well and ParallelFor with "priorities" that scales very well using my Threadpool with priorities that scales very well, and they will be integrated as methods with my Threadpool with priorities that scales very well, so that you will be happy. And my next step soon is also to make my Delphi and FreePascal and C++ Libraries portable to other CPUs like ARM etc. because currently they work on x86 AMD and Intel CPUs. And my next step soon is also to make my "scalable" RWLocks NUMA-aware and efficient on NUMA. This is was some parts of my everyday computing.. Thank you, Amine Moulay Ramdane. |
| Mr Flibble <flibbleREMOVETHISBIT@i42.co.uk>: Aug 05 12:33AM +0100 On 05/08/2018 03:50, Sky89 wrote: > Hello.... > I correct some typos because i write fast, read again.. Fuck off you egregious fuckwit of a cunt. AND. TAKE. YOUR. MEDICATION. /Flibble -- "Suppose it's all true, and you walk up to the pearly gates, and are confronted by God," Bryne asked on his show The Meaning of Life. "What will Stephen Fry say to him, her, or it?" "I'd say, bone cancer in children? What's that about?" Fry replied. "How dare you? How dare you create a world to which there is such misery that is not our fault. It's not right, it's utterly, utterly evil." "Why should I respect a capricious, mean-minded, stupid God who creates a world that is so full of injustice and pain. That's what I would say." |
| You received this digest because you're subscribed to updates for this group. You can change your settings on the group membership page. To unsubscribe from this group and stop receiving emails from it send an email to comp.lang.c+++unsubscribe@googlegroups.com. |
No comments:
Post a Comment