what io method is the best?
i belive that MapViewOfFile() because its more hardware related.
And it doesnt require RAM for buffer, i can just work on it as it was memory.
Maybe its slower if i do heavy work on file (instead of memory it use mapped hardware), but if for example im sending a file via network - there is no need for RAM buffer.

"Best" in what way?

What are your goals? What are your requirements? It all depends upon what you are doing.

send file over network

TransmitFile.

When dealing with files, memory is used. When mapping a hardware device into memory space things are different, but this isn't really something you do with CreateFileMapping+friends.

There's - possibly - slightly more overhead when dealing with memory-mapped files than normal Read/WriteFile calls, since you'll get a pagefault (including ring3->ring0->ring3) for every 4kb accessed. Probably not something you'll be able to seriously benchmark unless you get hold of a really old CPU, though

Be more specific. What sort of network? 10Gb Ethernet or 1Mb Internet (it makes a big difference)? "Best" in what way? Less work for you to code? Less work for the OS to do? Less CPU overhead? Most network speed? Least SDRAM used? What?

please more info about it.
in paging table memory mapped device has propably supervisor access bit set to 0, meaning that ring3 can access this memory. Its OS thing to allow or disallow ring3 access to memory, why would it restrict it to cpl < 3 when i havepermission to create mapping handle? Where do i have page fault?
I doubt that OS will swap out memory wich is used for memory mapped device, it must keep track of it.
Im not sure how it works, its my theory about it.


fast, simple, better, and without pointless resource waste.

You can't have all together, not possible.

TransmitFile seems to provide all that (although I'm not sure about the "pointless resource waste")

Is TransmitFile actually "simple"? I imagine the OS has a lot of work to do internally. The OP has not stated what is meant by "best" or "better" so how can we know what is wanted?

forget about TransmitFile, what are the flaws of memory mapped io?

Four years ago, I tested memory-mapped files (MMF) vs. ReadFile for large files on HDD under Win XP. The results: for uncached data (first-time read), the speed was similar, but if you read the same file again, MMF will be ­≈2 times faster, because it does not have to copy the memory, just remap the memory pages from cache. MMF is harder to use: there are more functions to call, and you cannot grow a file once it's created.

ok how does mmf work?
does it instruct OS to map portion of HDD space to address space?
or is it something else, like creating page with no access, upon access calling page fault handler, redirecting to normal page, comparing 8 bytes, and using WriteFile/ReadFile to actually write data there?

a115433,

CPU usually can't directly address HDD space (PIO/DMA vs. memory-mapped I/O). x86 OS usually uses demand-paging to achieve mapping effect. WriteFile/ReadFile use cache manager which, in turn, uses memory manager to map sections of file into system virtual address space.