Lockless uidinfo.

[Pawel Jakub Dawidek]

On Sat, Aug 18, 2007 at 04:35:42PM -0700, Jeff Roberson wrote:
> On Sun, 19 Aug 2007, Pawel Jakub Dawidek wrote:
> >Ok, after implementing atomic_fetchadd_long() on amd64, we get additional
> >6% of performance improvement:
> >
> >x ./uidinfo_lockfree.txt (atomic_cmpset_long loop)
> >+ ./uidinfo_waitfree.txt (atomic_fetchadd_long)
> >+------------------------------------------------------------------------------+
> >|                                                                          
> >+|
> >|                                                                          
> >+|
> >|x   xx    xx                                                              
> >+ ++|
> >|  |__MA___|                                                               
> >|AM|
> >+------------------------------------------------------------------------------+
> >   N           Min           Max        Median           Avg        Stddev
> >x   5       1561566       1575987       1568964       1569767     5853.1399
> >+   5       1662362       1665936       1665810     1664881.8     1541.2693
> >Difference at 95.0% confidence
> >       95114.8 +/- 6241.96
> >       6.05917% +/- 0.397636%
> >       (Student's t, pooled s = 4279.88)
> 
> How does this effect the single-threaded performance?  Do you attribute 
> this to atomic fetchadd being cheaper than atomic cmpset?  What is your 
> processor?

CPU: Intel(R) Xeon(R) CPU           E5310  @ 1.60GHz (1597.65-MHz
K8-class CPU)
  Origin = "GenuineIntel"  Id = 0x6f7  Stepping = 7
  Features=0xbfebfbff<FPU,VME,DE,PSE,TSC,MSR,PAE,MCE,CX8,APIC,SEP,MTRR,PGE,MCA,CMOV,PAT,PSE36,CLFLUSH,DTS,ACPI,MMX,FXSR,SSE,SSE2,SS,HTT,TM,PBE>
  Features2=0x4e33d<SSE3,RSVD2,MON,DS_CPL,VMX,TM2,SSSE3,CX16,xTPR,PDCM,DCA>
  AMD Features=0x20100800<SYSCALL,NX,LM>
  AMD Features2=0x1<LAHF>
  Cores per package: 4

Ok, I changed the code to something like this:

	long old;
	int diff, loops;

	atomic_add_int(&uidinfo_cnt1, 1);
	if (diff > 0) {
		loops = 0;
		do {
			loops++;
			old = uip->ui_sbsize;
			if (old + diff > max)
				return (0);
		} while (atomic_cmpset_long(&uip->ui_sbsize, old, old + diff) == 0);
		if (loops > 1)
			atomic_add_int(&uidinfo_cnt2, loops);
	} else {
		atomic_add_long(&uip->ui_sbsize, (long)diff);
	}

This allows me to see how many additional loops I do, because with
lock-free version we still can have contention and loop, that's why
wait-free version is superior.

Actually I was a bit surprised with the results:

debug.uidinfo.cnt1: 88746008
debug.uidinfo.cnt2: 31296304

(Running 8 processes.)

Which means, because of contention, we do 31296304 additional atomic
operations, which is about 30% more.

-- 
Pawel Jakub Dawidek                       http://www.wheel.pl
pjd at FreeBSD.org                           http://www.FreeBSD.org
FreeBSD committer                         Am I Evil? Yes, I Am!
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 187 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-arch/attachments/20070819/187a8939/attachment.pgp