Hi Team, Please find the attached diff with the latest changes.
The diff file changes that were done for memset and spinlock, are modified
on top of the previous patch.
New changes for review: 0001-AIX-support.tas.memset.diffs
Previous patch: 0001-AIX-support.v8.patch
>> diff --git a/src/include/storage/s_lock.h b/src/include/storage/s_lock.h
>
We have replaced the asm code with the gcc specific routines __sync_lock_test_and_set ()
and ran the pgbench. On AIX both have the similar performance, On PPcle we see a slight
variation in the performance. It seems the asm code is performing better on PPCle. Please
let us know, would it be better to continue to use the PPC assemble code over gcc routines.
Attached are the complete stats. (spinlock.stats.log)
Below is the summary…
AIX With asm AIX With __sync PPCle With asm PPCle With __sync
number of transactions "226554/ "230810/ "356945/ "364346/
above the 10.0 ms 556329 567936 918647 863937
latency limit (40.723%)" (40.640%)" (38.856%)" (42.173%)"
latency average 16.160 ms 15.821 ms 9.796 ms 10.416 ms
initial connection time 230.786 ms 249.668 ms 19.606 ms 20.090 ms
tps 3086.209412 3158.14915 5103.024267 4799.621896
> diff --git a/src/makefiles/Makefile.aix b/src/makefiles/Makefile.aix
>
> At least it needs to be updated to match what MemSet() looks like
> nowadays. The changes may be just cosmetic, but better check. Should
> also check the effect on MemSetAligned(). That might matter more for
> performance in practice.
As per the stats in the previous mail wrt to memset, both the loop and the native
memset are performing similar after optimization “-O2”. So for now we removed the
native memset changes.
We will setup the memset performance on different platforms and will post the
details in a different thread.
>>> +# -blibpath must contain ALL directories where we should look for libraries
>>> +libpath := $(shell echo $(subst -L,:,$(filter -L/%,$(LDFLAGS))) | sed -e's/ //g'):/usr/lib:/lib
>
Added additional details about the -blibpath usage. These details are available in the below link
https://download.boulder.ibm.com/ibmdl/pub/software/dw/aix/es-aix_ll.pdf
> 21 -- test overflow/underflow handling
>
> 22 SELECT gamma(float8 '-infinity');
>
> 23 ERROR: value out of range: overflow
This testcase is resolved from AIX libm side. Below is the result.
postgres=#
SELECT x, gamma(x),lgamma(x)
FROM (VALUES (0.5), (1), (2), (3), (4), (5),
(float8 'infinity'), (float8 'nan')) AS t(x);
0.5 | 1.772453850905516 | 0.5723649429247001
1 | 1 | 0
2 | 1 | 0
3 | 2 | 0.6931471805599453
4 | 6 | 1.791759469228055
5 | 24 | 3.1780538303479458
Infinity | Infinity | Infinity
NaN | NaN | NaN
Kindly let us know your feedback on the diffs.
I will commit the diffs to the patch(0001-AIX-support.v8.patch) once you are fine.
Warm regards,
Sriram.
