Last call for 2.4.3 problems

[Ulrich Schmid]

on 17/1/06 18:51 Russ Allbery wrote:
> Ulrich Schmid <schmid at switch.ch> writes:
> 
>>on 13/1/06 20:02 Russ Allbery wrote:
> 
> 
>>>I'm not at all sure that my messages are going through, and I seem to
>>>only be getting mail from the list intermittantly.  But let's see if
>>>this makes it through.
> 
> 
>>>I've been running 2.4.3 for a while on one of my systems, and it seems
>>>generally okay.  I'm therefore planning on making the release official.
>>>Has anyone else run into any problems?  (I'll respond to Jeff's message
>>>in just a moment.)
> 
> 
>>I observed a significant performance loss with 2.4.3 compared to 2.4.1.
>>Trying 2.4.2 I saw the same problem.  Obviously this is caused by
>>history functions (history sync/write etc), since they take longer and
>>history disk I/O increased.
> 
> 
> Huh, that's surprising, particularly since there are various performance
> fixes in 2.4.3 (although I think they're mostly with CNFS).  I can't think
> of what would have changed in the history code between 2.4.3 and 2.4.1 to
> have caused that....
> 

from the ChangeLog of 2.4.2 (missing in 2.4.3 ChangeLog):
2004-06-02 20:11  hkehoe
    * history/hisv6/hisv6.c: In hisv6_reopen(), if MMAP_NEEDS_MSYNC and
      INND_DBZINCORE are set and nfsreader is false, override INCORE_NO
      to INCORE_MMAP

2004-06-02 20:02  hkehoe
    * lib/dbz.c: Put msync call in putcore for INCORE_MMAP Move
      writethrough test from dbzsync to putcore because it only applies
      to INCORE_MEM

whereas I assume that my problem is due to the changes in dbzsync / putcore.
If I use in 2.4.3 dbzsync as of 2.4.1 I got the same history performance
as with 2.4.1. Is the writethrough test correct in putcore?


Ulrich