Re: Asynchronous (overlapping) file access
Claus wrote:
Hi all!
I have a question concerning the asynchronous file access when using the
Platform SDK functions CreateFile/WriteFile. I am not completely sure what
"asynchronous" means. I thought, that the function WriteFile would return as
soon as possible after calling, and the file system does the job of writing
the data to disk in the background. But when I did some benchmarking, I
realized that synchronous and asynchronous file access are equally fast, when
measuring the time it takes for the WriteFile function to return. It could
well be, that I messed up some flags at the CreateFile / WriteFile
functions... Maybe some code is helpful to describe my benchmarking:
<code snipped>
Right now the code is for asynchronous IO. When benchmarking synchronous IO
I also comment out the "HasOverlappedIoCompleted" line.
The thing is, that timer_1 and timer_2 show almost identical results, only 2
milliseconds or so apart. The filesize is 150MB which takes (about) 0.85
seconds. Both timers show this result (only 2 ms difference). There is also
no difference in using WriteFileEx instead of WriteFile, or if I use
"overlapped.hEvent = NULL;" instead of "overlapped.hEvent = h_event;"
Is my understanding of asynchronous IO wrong, or is there a bug in my code?
Regards,
Claus
Yes, "asynchronous" means WriteFile returns before the write is
complete. The hoped-for advantage of doing this is that you can do
useful work on the CPU while I/O hardware performs the write.
But since you make only one call to WriteFile, and time the completion
of the write, all you have done is measure the speed of the disk without
taking any advantage of the overlapped capability. You wasted it by
immediately calling HasOverlappedIoCompleted. You could have used that
wasted processor time to prepare more data and/or queue several more writes.
Asynchronous writing does not speed up the disk, but it does let you
speed up the overall operation by preparing and queueing data
concurrently with writing previous data.
--
Scott McPhillips [MVP VC++]