Re: performance problem - 10.000 databases

Quoth sfrost(at)snowman(dot)net (Stephen Frost):
> * Christopher Browne (cbbrowne(at)libertyrms(dot)info) wrote:
>> On one of our test servers, I set "fsync=false", and a test load's
>> load time dropped from about 90 minutes to 3 minutes. (It was
>> REALLY update heavy, with huge numbers of tiny transactions.)
>>
>> Which is, yes, quite spectacularly faster. But also quite
>> spectacularly unsafe.
>>
>> I'm willing to live with the risk on a test box whose purpose is
>> _testing_; it's certainly not a good thing to do in production.
>
> Would it be possible to have the effectively done for a specific
> transaction? If this was done as a single large transaction could
> there be an option to say "don't fsync this until it's all done and
> then do it all" or something? Just looking for a way to get the
> 'best of both worlds'...

Oh, for sure, the whole thing could be invoked as one giant
transaction, which would reduce the cost dramatically.

But it diminishes the value of the benchmark for my purposes. It's
useful to measure how costly those individual transactions are.
--
wm(X,Y):-write(X),write('@'),write(Y). wm('cbbrowne','cbbrowne.com').
http://www.ntlug.org/~cbbrowne/oses.html
"As far as Saddam Hussein being a great military strategist, he is
neither a strategist, nor is he schooled in the operational arts, nor
is he a tactician, nor is he a general, nor is he as a soldier. Other
than that, he's a great military man, I want you to know that."
-- General Norman Schwarzkopf, 2/27/91