ansaurus

Question

incomprehensible time consumed in using memory mapped file

Answer 1

+1 A:

How fragmented is the file you are comparing with? You can use FSCTL_GET_RETRIEVAL_POINTERS to get the ranges that the file maps to on disk. I suspect the last 25 MB will have a lot of small ranges to account for the performance you have measured.

MSN 2009-09-03 16:28:19

Thanks for the tip. But it's unlike caused by fragment, as I've test on multiple file ranging from MBs to GBs.

t.g. 2009-09-04 02:07:13

Answer 2

+1 A:

I wonder if mmap behaves strangely when a segment isn't an even number of pages in size? Maybe you can try handling the last parts of the file by progressively halving your segment sizes until you get to a size that's less than mapped_file_source::alignment() and handling that last little bit specially.

Also, you say you're doing 512MB blocks, but your code sets the size to 8<<10. It then multiplies that by mapped_file_source::alignment(). Is mapped_file_source::alignment() really 65536?

I would recommend, to be more portable and cause less confusion, that you simply use the size as given in the template parameter and simply require that it be an even multiple of mapped_file_source::alignment() in your code. Or have people pass in the power of two to start at for the block size, or something. Having the block size passed in as a template parameter then be multiplied by some strange implementation defined constant seems a little odd.

Omnifarious 2009-09-03 21:18:21

Regarding alignment(), mapped_file::alignment() is a static member function Returning the operating system's virtual memory allocation granularity, which is not implementation-defined but platform defined. When using mmap, "the offset must be a multiple of the allocation granularity" (http://msdn.microsoft.com/en-us/library/aa366763%28VS.85%29.aspx). Using a FACTOR(VIEW_SIZE_FACTOR), instead of SIZE(VIEW_SIZE), as a template parameter would guaranteed such a rule would not be offended, also easing the user the labor of finding the granularity.

t.g. 2009-09-04 02:03:23

Also (8<<10) is just a way to ease human reading, as the granularity are usually in KBs, and (1<<10) is 1K. looking at the preceding digit(say, 8), would tell approximately how big it is in MB. and because it's const integter. it will be calculated in compilation rather than run-time.

t.g. 2009-09-04 02:04:14

I would change the factor to 1 to test your suggestion. Thanks.

t.g. 2009-09-04 02:05:22

My point about the size is that I think it's better to take the actual size of your segments as the template parameter and just error out if it isn't a multiple of the platform specific value.

Omnifarious 2009-09-04 03:41:02

Answer 3

+2 A:

This behavior looks quite illogical. I wonder what would happen if we tried something stupid. Provided the overall file is larger than 512MB you could compare again a full 512MB for the last part instead of the remaining size.

something like:

        if(remainder > 0)
        {
            cout << "segment size = " 
                 << remainder 
                 << " bytes for the remaining round";
                if (size > segment_size){
                    block_size = segment_size;
                    offset = size - segment_size;
                }
                else{
                    block_size = remainder;
                    offset = segment_size * i
                }
            if(!segment_compare(block_size, offset)) return false;    
        }

It seems a really dumb thing to do because we would be comparing part of the file two times but if your profiling figures are accurate it should be faster.

It won't give us an answer (yet) but if it is indeed faster it means the response we are looking for lies in what your program does for small blocks of data.

kriss 2009-09-04 22:05:43

Hihi, fantastic idea! Thinking outside the memory-mapped box.

Jonas Byström 2009-09-05 22:42:08

it does not work because, in that way, the alignment rule that "the offset must be a multiple of the allocation granularity" is violated, thus failing the map. I did similar things like further splitting the remainder. see my updates in the OP.

t.g. 2009-09-07 03:14:04

Answer 4

+1 A:

I would try it on a Linux or BSD just to see how it acts, out of curiousity.

I have a really rough guess about the problem: I bet that Windows is doing a lot of extra checks to make sure it doesn't map past the end of the file. In the past there have been security problems in some OS's that allowed a mmap user to view filesystem-private data or data from other files in the area just past the end of the map, so being careful here is a good idea for a OS designer. So Windows may be using a much more careful "copy data from disk to kernel, zero out unmapped data, copy data to user" instead of the much faster "copy data from disk to user".

Try mapping to just under the end of the file, excluding the last bytes that don't fit into a 64K block.

Zan Lynx 2009-09-05 05:09:26

I just tried it on another Windows PC. please see my updates in the OP.

t.g. 2009-09-07 05:48:37

Answer 5

+1 A:

I know this isn't an exact answer to your question; but have you tried side-stepping the entire problem - i.e. just map the entire file in one go?

I know little about Win32 memory management; but on Linux you can use the MAP_NORESERVE flag with mmap(), so you don't need to reserve RAM for the entire filesize. Considering you are just reading from both files the OS should be able to throw away pages at any time if it gets short of RAM...

Dave Rigby 2009-09-05 12:02:33

When the file is too big, it will fail to map. That's why I split the the file into parts and map them view by view, which is the recommended way in Win32 as far as I know. When the file is small enough, yes, it can be mapped in one go with no performance problem.

t.g. 2009-09-07 03:09:26

Answer 6

A:

Could it be that a virus scanner is causing these strange results? Have you tried without virus scanner?

Regards,

Sebastiaan

Sebastiaan Megens 2009-09-07 06:47:26

Unlikely, as I tried on a few PC with the same virus scanner installed. When the result only seems related to h/w configuration. I am guess it's some kind of CPU or RAM cache impact, but could not tell how, that's why I posted the h/w config out.

t.g. 2009-09-07 06:53:39

I agree, indeed it seems unlikely. But in these strange situations I want to check all my assumptions so I always try if the problem is still there with the virus scanner disabled (and the firewall, when reading files from a network).Regards,Sebastiaan

Sebastiaan Megens 2009-09-08 14:03:40

ansaurus

tags:

views:

answers:

incomprehensible time consumed in using memory mapped file

More observations after the suggestions by responders

Updates 2

Test Code

Observation:

System Info

related questions