txn performance penalty

[Matthias Andree]

Greg Louis <glouis at dynamicro.on.ca> writes:

>>0.93.1 took 28 s with default settings
>>   CVS took 22 s with default settings
>>   CVS took 15 s with --disable-transactions
>>
>>(arithmetic average of three runs with warm caches)
>
> (1) my application, and the runs I quoted, were with the cache stone
> cold :)

That didn't escape me. The overhead from synchronous I/O is the same
though, unless the hardware is cheating (fast write cache or something
like that).

> (2) having changed notebooks since then, I can no longer reproduce
> these ridiculous times; I'm now getting, for 1457 spam, 19 seconds of
> wall-clock time with 0.93.1 / qdbm, and 35 seconds with 0.93.1 / db
> with transactions enabled.  Again, starting stone cold, ie right after
> logging in after a reboot.
>
> No idea what produced the difference.  Maybe I wrote the wordlist
> directory for db_txn on a hunk of my 3-year-old HD that happened to
> have a lot of soft errors?  Not very likely; nothing else seemed unduly
> slow.

Slow seek on a full disk, perhaps with scattered free space and
scattered files, is a real performance killer. Fresh installs on disks
with plenty of space are faster.

-- 
Matthias Andree