Re: PAE Slowdown

At 2:00 PM -0500 10/8/07, Richard Todd wrote:
Jeff Kramer <jeffk@xxxxxxxx> writes:

Hey all,

I know that AMD64's the preferred way to run >4 gig systems, but I'm
having a weird situation with 6.2-RELEASE-p8 and 6-STABLE as of last
night. When I compile the PAE kernel, my system performance drops
like a rock. It still boots and everything still runs, but for
instance, running the Flops port my megaflops drop from the 950 MFLOPS
range to 4 MFLOPS. It feels about as fast as a 486.

This may not be a PAE-related problem. I say this because I noticed you
have the same MB I have:

ACPI APIC Table: <INTEL DP965LT >

Several Intel MBs, including the DP965LT, have a BIOS bug that rears
its head when you have 4G (or more) of memory installed, where the
BIOS sets the cache control registers incorrectly. This cause a chunk
of your main memory (on my system, the chunk between 448MB and 512MB)
to be labeled uncachable, with the result being random slowdowns
whenever the kernel or user processes happen to touch memory in that
chunk. This problem drove me crazy trying to figure out what the problem
was until I stumbled on this report on a Linux users' forum explaining
the situation.

http://forums.fedoraforum.org/showthread.php?t=157232

Fortunately, the workaround is fairly straightforward, adding an rc.d script
to twiddle the MTRRs. Assuming this is your problem, if you could post the
output of "memcontrol list" it should be possible to id which of the entires
is bogus and needs to be removed.

Sweet, I just downgraded to the 1669 bios rev and it looks like it's running at full speed. I'm compiling an SMP/PAE kernel now, but it looks like this was the fix! Thanks, Richard!

--

Jeff Kramer
jeffk@xxxxxxxx
http://www.jeffkramer.org/
_______________________________________________
freebsd-stable@xxxxxxxxxxx mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-stable
To unsubscribe, send any mail to "freebsd-stable-unsubscribe@xxxxxxxxxxx"

Relevant Pages

  • [ANNOUNCE] Linux 2.4.32-hf32.7
    ... I finally put together my the remaining fixes for kernel 2.4.32. ... would read kernel memory beyond the allocation size, ... If SCTP receives a badly formatted HB-ACK chunk, ... same inode on ext3 filesystem. ...
    (Linux-Kernel)
  • Re: Directories > 2GB
    ... because it needs to map the buffer into a single contiguous chunk ... This was placing a lot of pressure on the memory ... code to deal with non contig memory was looking like a major ... work to well in the linux kernel. ...
    (Linux-Kernel)
  • Re: malloc and Active memory
    ... >> that explain why the Active memory is not released again as soon as the ... > isn't a SINGLE chunk available of that size. ... > there isn't a single 7GB chunk available and malloc() will fail. ... See also the kernel Documentation directory, ...
    (Fedora)
  • Re: Reclaiming freed memory back from running process
    ... Process freed that memory ... If system start getting out of memory will kernel ... it back only after process exits?If a memory page is not used by a process, the kernel will swap it out ... I think what he said is regarding the glibc cache for freed chunk. ...
    (comp.os.linux.development.apps)