On 11.01.2008 [11:24:46 +1100], David Gibson wrote: > On Wed, Jan 09, 2008 at 11:31:38PM -0800, Nishanth Aravamudan wrote: > > On 10.01.2008 [16:01:11 +1100], David Gibson wrote: > > > On Tue, Jan 08, 2008 at 09:07:57PM -0800, Nishanth Aravamudan wrote: > > > > For larger databases, users often only want their text to be placed in > > > > hugepages. libhugetlbfs does not currenlty provide this mode, however, > > > > so users have been working around it in ... odd ... ways. Add explicit > > > > > > I still don't really understand how you can get a program with text so > > > enormous it will benefit but data which won't, but there you go. > > > > Well, I *think* it's more that they are sharing the text, so really the > > benefit comes from the hugepages being shared by all the programs. But > > the cost of using hugepage data nullifies it. That's my understanding, > > at least. > > Except that if they were using normal pages, the text would just be > mmap()ed from disk and so should be shared anyway.
Hrm, true. I know of at least one case where the BDT re-linked program failed, but the T re-linked program worked. I will try to look through my notes to get details. > > > > support for T-only relinking via new linker scripts for ppc64, x86 and > > > > x86_64. Unfortunately, I do not think (with my knowledge, at least) that > > > > ppc32 can support T-only, due to the limitation that the PLT must be > > > > within 32M of the text segment, confirmed in testing and based upon the > > > > comment in the elf32ppclinux.xBDT: > > > > > > It might be possible to do this by having a mini-data segment in > > > hugepages near the text segment, containing just the PLT, with the > > > rest of the data segment and BSS back in normal pages. > > > > I think Adam is encountering a problem with glibc where the text, data > > and BSS are not continuous (64-bit B-linker script). This leads to odd > > failures when libraries get mapped in between the segments. I would > > worry that would happen here too... But it's a good idea and I can try > > testing it out. > > Ah. Yuck. Indeed. He's working on a fix, but I don't think it's a particularly pleasant case. And RedHat's prelink exacerbates the situation. Adam, care to fill in the details? -Nish -- Nishanth Aravamudan <[EMAIL PROTECTED]> IBM Linux Technology Center ------------------------------------------------------------------------- Check out the new SourceForge.net Marketplace. It's the best place to buy or sell services for just about anything Open Source. http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace _______________________________________________ Libhugetlbfs-devel mailing list Libhugetlbfs-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/libhugetlbfs-devel
