[Linux-ia64] Re: Confused about fsystemcalls

[Linux-ia64] Re: Confused about fsystemcalls

[David Mosberger]

>>>>> On Thu, 30 Jan 2003 11:07:51 +1100, Peter Chubb <peter@chubb.wattle.id.au> said:

  Peter> The mygetpid() call appears to return the correct value.

  Peter> Our figures show over an average of ten runs: --- glibc
  Peter> getpid() takes 579 cycles, stddev 24 --- mygetpid takes 1468
  Peter> cycles, stddev 286

Have you tried on a McKinley?  The numbers you quote for the standard
getpid() look like for Merced.  I never tried on Merced, though I
don't know of why it shouldn't work fast, off hand.

  Peter> I was also a little confused about the entry put into the ELF
  Peter> AUX vector -- was the value or the user-space address of
  Peter> syscall_via_epc meant to go into the vector?  Currently it
  Peter> says in include/asm-ia64/elf.h

Oh, it was broken.  I just fixed it today (since I got glibc to work
with the new syscall stub) and it's in the lia64 bk tree now.

	--david

[Linux-ia64] Re: Confused about fsystemcalls

[Peter Chubb]

>>>>> "David" = David Mosberger <davidm@napali.hpl.hp.com> writes:

>>>>> On Thu, 30 Jan 2003 11:07:51 +1100, Peter Chubb <peter@chubb.wattle.id.au> said:
Peter> The mygetpid() call appears to return the correct value.

Peter> Our figures show over an average of ten runs: --- glibc
Peter> getpid() takes 579 cycles, stddev 24 --- mygetpid takes 1468
Peter> cycles, stddev 286

David> Have you tried on a McKinley?  The numbers you quote for the
David> standard getpid() look like for Merced.  I never tried on
David> Merced, though I don't know of why it shouldn't work fast, off
David> hand.


Worked it out -- one needs to call the mygetpid() function twice, once
before the timing stuff, and once during, so we don't count the 
time to get the gate page into the TLB and into the cache.

Timings are now 177 cycles for the mygetpid() routine as opposed to
570 for glibc (all on Itanium 1), and we're now working on
gettimeofday().


Peter C

[Linux-ia64] Re: Confused about fsystemcalls

[David Mosberger]

>>>>> On Thu, 30 Jan 2003 15:31:19 +1100, Peter Chubb <peter@chubb.wattle.id.au> said:

  Peter> Worked it out -- one needs to call the mygetpid() function
  Peter> twice, once before the timing stuff, and once during, so we
  Peter> don't count the time to get the gate page into the TLB and
  Peter> into the cache.

Yes, that makes a big difference.  Here is what I see for getppid() on
a McKinley:

	getppid cycles: 1410 79 56 54 54 54 54 54 54 54

(I just added getppid() and I also added code to patch over the
McKinley E9 workaround on CPUs that don't need it.)

  Peter> Timings are now 177 cycles for the mygetpid() routine as
  Peter> opposed to 570 for glibc (all on Itanium 1),

Ah, that's much better.

  Peter> and we're now working on gettimeofday().

Cool.  I assume you're doing this based on the lock-free version that
Andrew Morton has in his latest patch?

	--david