0

For, e.g. determining an amount of memory that is safe to allocate for processing file or device with this type of I/O loop:

 HANDLE hFile /* = file open with GENERIC_READ */;
 LARGE_INTEGER liSize;
 SIZE_T cbAlloc;
 BYTE *pBuffer = NULL;

 GetFileSizeEx(hFile, &liSize);
 while(liSize.QuadPart)
 {
     cbAlloc = min(ALLOC_MAX, liSize.QuadPart);
     /* Allocate pBuffer with cbAlloc bytes size, 
        Use pBuffer, then Free pBuffer... */
     liSize.QuadPart -= cbAlloc;
 }

Right now my ALLOC_MAX is a macro constant.

However, when researching a way to determine a safe amount to allocate dynamically, with e.g. GlobalMemoryStatusEx, my next question is, how does even this information (including total physical memory available) even help me determine what the right amount is for ALLOC_MAX? It would not be wise to allocate the exact amount of available memory, would it? What is a reasonable amount to try to allocate, three-quarters of the available amount? Half of it?

Improve this question
1
  • 1
    Why would you do such a huge allocation? On the contrary, you should try to allocate something small, around the size of page, maybe more, say 10kb, and stream the data. Commented Sep 17, 2024 at 3:46

2 Answers 2

Reset to default
4

You should consider

  • which kind of target system(s) you are aiming for, and how much real memory they will have at minimum

  • what other kind of programs and processes will run at the same machine at the same time, or if your program is running on a box which is dedicated for your process exclusively

  • what you are hoping to gain from allocating 1GB at a whole - will there be any significant performance difference when you choose a much smaller buffer, lets say 1k, 10k, maybe 100k? (And don't guess - measure!)

  • why ALLOC_MAX is a constant. Depending on the context, this choice might be fine, but maybe you want the memory consumption of the program configurable, for example, to adapt it to the specific machine or environment where the process will be executed.

The ideal choice for ALLOC_MAX is the smallest one which gives you acceptable performance - if you make it configurable, it will be easier to make some benchmarks in different environments without any need to recompile the program.

Improve this answer
9
  • 1
    “Measure” is bold for a reason. Commented Sep 17, 2024 at 7:26
  • 1
    lets say 1k, 10k, maybe 100k? IME page-sized multiples will probably perform best, so 8kb, 128kb, etc. Also IME, unless the IO system is specifically designed for high-performance IO and the code is properly "tuned" to match things like file system block sizes and just about every other detail of the storage system, single-threaded read/writes with a buffer that's just a handful of pages gets about 95% or better of the performance from huge buffers such as 1 GB. IMO it wouldn't be unreasonable to just hardcode a 16 KB buffer size and forget about it. Commented Sep 17, 2024 at 12:37
  • @AndrewHenle: all correct, but what we don't know is what the OP's code is actually doing with the buffer content, in the sections they did not show us, which can have a significant influence on choosing the size. Commented Sep 17, 2024 at 12:43
  • @AndrewHenle Have you measured it? Commented Oct 2, 2024 at 14:16
  • 1
    @gnasher729: I think for explaning my point it makes no difference whether I wrote "k" or "kb" - to be honest, today I have no idea whether this was intentional or a typo at the time when I was writing it ;-) Commented Oct 23, 2024 at 14:37
1

You measure. You hope that one 1GB read is faster than 1000 1MB reads. You are missing that you are using memory that might be needed elsewhere (so another app on your computer starts swapping), and your data doesn’t fit into L2 cache anymore, which can make a massive difference. So there are plenty reasons why a big size might be slower.

Improve this answer
1
  • 1
    Re: "you are missing..." not really, that's why I'm asking this question... Commented Sep 17, 2024 at 14:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.