Z>> Bernhard Heidegger <bheide@hyperwave.com> writes:
>> >>>>> ">" == Eric W Biederman <ebiederm@inetnebr.com> writes:
>>
>> >>>> No. Major performance problem.
>>
BH> Why?
>>
BH> Imagine an application which has most of the (index) file pages in memory
BH> and many of the pages are dirty. bdflush will flush the pages regularly,
BH> but the pages will get dirty immediately again.
BH> If you can be sure, that the power cannot fail the performance should be
BH> much better without bdflush, because kflushd has to write pages only if
BH> the system is running low on memory...
>>
>> >> The performance improvement comes when looking for free memory. In
>> >> most cases bdflush's slow but steady writing of pages keeps buffers
>> >> clean. When the application wants more memory with bdflush in the
>> >> background unsually the pages it needs will be clean (because the I/O
>> >> started before the application needed it), so they can just be dropped
>> >> out of memory. Relying on kflushd means nothing is written until an
>> >> application needs the memory and then it must wait until something is
>> >> written to disk, which is much slower.
>>
>> >> Further
>> >> a) garanteeing no power failure is hard.
>>
>> Use and UPS and regularly flush/sync the primary data to disk from
>> the application
Z>> Update/bdflush costs you nothing. UPS costs you lots of money. Big
Z>> difference.
Yes, but on a real big server where performance does matter (greetings
from Godzilla ;-) you will have an UPS anyway...
Z>> Also, flushing/syncing data to disk doesn't always mean data really
Z>> got to media. Check your favorite sync(2) manpage. :)
Correct, but you doesn't win anything with update/bdflush in this case.
Z>> Using completely synchronous API in applications would consideraly cut
Z>> performances down. Why would your application wait for disk to commit
Z>> buffers, when your CPU can do other useful things in the meantime.
Z>> Also, don't forget that disk latency times are measured in
Z>> milliseconds, where modern CPU's run in units of (almost) nanoseconds.
>>
>> >> b) generally there is so much data on the disk you must write it
>> >> sometime, because you can't hold it all in memory.
>>
>> only a question of how much RAM you can put in your PC
Z>> Still requires money. :)
RAM isn't that expensive anymore
>>
>> >> c) I have trouble imagining a case where a small file would be rewritten
>> >> continually.
>>
>> Not really small, but a database application may use btree based indexes,
>> where many blocks will get dirty when inserting/deleting data. If you flush
>> the dirty buffers and the next insertion dirty the same buffer(s) you have
>> lost performance (Note: the btree based indexes are secondary data; you
>> can rebuild it from scratch if the system fails)
>>
Z>> Right, we agree. But performance doesn't go down if you write buffers
Z>> every few tens of seconds. That is a LOT of time, if you ask your
Z>> application. Some of them never get so old. :)
Hey, I speak of (database) applications which (should ;-) run until the
earth go down :-)
I did some measurements with our application and there were peaks which
were 10 times higher than the average time. I will make some further
tests in order to see if the overall performance drops. Anyway, this
application is also used interactively and if you try to get some data
during such a peak you'll have to wait ;-)
Z>> And (big) databases mostly like to have their own memory management,
Z>> because "they know better".
That's another point I agree with you, but this is a really big task...
Bernhard
get my pgp key from a public keyserver (keyID=0x62446355)
-----------------------------------------------------------------------------
Bernhard Heidegger bheide@hyperwave.com
Hyperwave Software Research & Development
Schloegelgasse 9/1, A-8010 Graz
Voice: ++43/316/820918-25 Fax: ++43/316/820918-99
-----------------------------------------------------------------------------
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.altern.org/andrebalsa/doc/lkml-faq.html