> Peng Du <imdup...@gmail.com> wrote: > > > Hello, experts > > > > I have a question about syscall "madvise" > (http://docs.sun.com/app/docs/doc/816-5168/madvise-3c? > a=view). I want to know whether MADV_WILLNEED and > MADV_DONTNEED will affect pageout decisions. In my > intuition, page scanner should try to keep > MADV_WILLNEED pages in physical memory, and prefer > MADV_DONTNEED pages for swapping out. > > I tried to use this 10 years ago in order to make > star faster to no avail. > I would like to see a way to tell Solaris to do read > aheads but to remove > anything from the cache after it was read once.
As I'd said, I've seen MADV_DONTNEED work, at least in some situations. But I think you'd have to look at the kernel code keeping in mind just how you were trying to use MADV_DONTNEED. Or do a search of opensolaris.org for MADV_DONTNEED or perhaps MADV_SEQUENTIAL, to see how other programs use them. For example, I think they will _not_ behave as expected if applied to the buffer passed to read(2), but only to pages of memory obtained from mmap(2). That would be because read(2) does not necessarily read directly into your buffer, but rather pages in enough pages of the file (if mmap-able rather than a STREAMS device or pipe or somesuch) and copies from the mapped in pages to your buffer. Pages in your read() buffer are probably part of stack or heap (anonymous memory) or private mapping of your data segment, in which case they're also backed by swap rather than by the file you're reading. (Maybe if one were using directio(3c) in a situation where it was supported, and read()ing into a page-multiple-sized, page-aligned buffer, it might work to use madvise() directly on that buffer; I haven't investigated that.) (I'm not saying that you made that particular mistake; I'm only saying that it's not always as obvious as one might like just how madvise(3c)/memcntl(2) actually work.) -- This message posted from opensolaris.org _______________________________________________ opensolaris-code mailing list opensolaris-code@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/opensolaris-code
