SHMLBA on linux-ia64

SHMLBA on linux-ia64

[Umut Aymakoglu]

Hi -
  The file /usr/include/sys/shm.h defines SHMLBA  as  __getpagesize() which
is 16K. whereas in /usr/include/asm/shmparam.h thereis another define of
SHMLBA as (1024*1024) and a comment saying on IA-64 architecture  due to
some performance problems it is better to use this alignment.

  We want to know what the recommended SHMLBA should be.

thanks in advance,
Umut

Re: SHMLBA on linux-ia64

[David Mosberger]

>>>>> On Tue, 2 Sep 2003 17:33:15 -0700, Umut Aymakoglu <umuta@us.ibm.com> said:

  Umut> Hi - The file /usr/include/sys/shm.h defines SHMLBA as
  Umut> __getpagesize() which is 16K.

Must be an old libc.  For example, with glibc v2.3.2, I see:

 $ fgrep 'e SHMLBA' /usr/include/bits/shm.h
 #define SHMLBA          (1024 * 1024)

  Umut> whereas in /usr/include/asm/shmparam.h thereis another define
  Umut> of SHMLBA as (1024*1024) and a comment saying on IA-64
  Umut> architecture due to some performance problems it is better to
  Umut> use this alignment.

Not "performance problem" but "_potential_ performance degradation".
On existing ia64 processors, you won't see a performance degradation,
but it is nevertheless recommend to align shared mappings to a 1MB
boundary because the architecture reserves the right that some future
CPU might run slower when there is a virtual alias that is not
congruent modulo 1MB.

	--david