This discussion is archived
1 2 3 Previous Next 38 Replies Latest reply: Jul 18, 2007 1:56 AM by 16279 Go to original post RSS
  • 15. Re: log_buffer
    Jonathan Lewis Oracle ACE Director
    Currently Being Moderated
    Your comments in italic:

    > OK, educate me. If they are not multiple LGW(n) processes, then how do the “I/O slaves for LGWR” appear in a "ps -ef" command?

    You still haven't grasped the difference between the LGWR process and I/O slaves. Please re-read my previous post. In a unix system, I think they typically appear as with "I2nn" (where nn is a two digit number) in their names - but it's been quite some time since I've been to a site that enabled multiple I/O slaves. (And the dbwr I/O slaves are I1nn).


    Can you elaborate on this, please?
    Just reminding people in general that you can't get multiple db writers and multiple db writer I/O slaves going at the same time. From memory, if you try to set both parameters it's the I/O slaves that get started with just one db writer.

    Touché! . . .
    Humour (or humor) is sometimes a very good thing. I salute you.

    Thanks, but will you use the nofollow tag? Cmon, we both know why you did not hyperlink to my quotes. An apology is for when you make a mistake, and this was no mistake! You chose not to link to my quotes because you don’t want the search engines to increase my search rank.
    It wouldn't have been a mistake, and I do know about the nofollow tag. So far I can think of two places where I didn't point the finger. You will, no doubt, find it hard to believe, but at the time I was trying to be kind by not pointing out that it was another Burleson article that I was criticising. But please let me know about the rest of the "loads" so that I can update them.

    Stop putting words in my mouth. I said no such thing. It’s “apparent” only that I shared my real-world observations from databases where increasing the log_buffer improved performance, nothing more. ... Not at all. I’m an empiricist. When I see a clear-cut case of a too-small log_buffer, I try to publish the symptoms.

    So you've published the symptoms of (two) 'clear cut case of a too-small log_buffer' - neither showing significant log buffer space waits, one showing very poor I/O wait times, and one showing a remarkable percentage of CPU time. It would have been nice to have an explanation in the original text about why the cases were so clear cut.

    Why doodle with test scripts when you can talk to the folks who wrote the source code?
    Personally, I only talk to the senior technicians when I have a high degree of confidence that I'm looking at an Oracle bug - and then I build a test case to demonstrate the problem. It's quite hard to get to the people who wrote the code by logging SRs.

    Ah, but isn’t that want you attempted in your test case?
    Not really - there is a difference (perhaps subtle) between "providing a counter-example" and "arguing from the particular to the general". However, I will accept that there is something to your argument - and then compare it with Bertrand Russell's teapot.

    I can't prove that there isn't a china teapot in orbit somewhere between the Earth and Mars - but I would be a little silly to assume that there was (until Voyager XXXII goes up with a team of astronauts from the UK perhaps). How many CPUs do I need to have, what setting of lgwrio_slaves do I need to have, and how many more Metalink articles do we have to see to convince you that there is only ever (at present) one LGWR process ?

    As I’ve watched your knowledge of tuning mature,
    Tsk, tsk - now who's playing with words.

    the behaviour [sic] of
    No need to put in the [sic] - I'm sure the readers of the forum are aware that you're only copying my English spelling of the word, and haven't made a spelling mistake.


    No, let’s discuss it! Prove to me that Metalink is wrong
    I don't need to, that bit of Metalink says you get multiple I/O slaves - not multiple LGWR processes. That bit of Metalink happens to be right.

    Which, in turn, is facilitated by multiple processors. You could make the exact same argument about OPQ
    "facilitated" not "dependent upon" - but irrelevant to the discussion about multiple LGWR processes.

    But it’s a whole lot easier to learn how Oracle works from the people who wrote it, than speculating with test cases, IMHO.
    Logical fallacy - it would certainly be better (and possibly quicker) to learn how specific bits or Oacle work by talking to the people that wrote the specification and wrote the code (and the code does not always agree with the original spec). But you've switched from "Metalink" to "the people who wrote the code". That's not really likely to be the same people.

    Do you feel comfortable generalizing that your “demonstration” applies to all Oracle databases?
    Logical fallacy. You claim that the log buffer has to be a multiple of 2KB so, as a starting point I only have to show once that you are wrong in order to determine the need for correction or further investigation . The fact that my demonstration is consistent with the manuals (and metalink but I wouldn't take that as a guarantee) is interesting, and it was jolly nice of me to give you a let-out with the bit about "following the O/S block size" - I think the 2KB suggests one of the Sequent models, by the way, if you had a machine running with the default operating system install.


    Yes, but the default values are often set according to rules of optimal performance
    So you're prepared to say that any default value is a recommendation for the maximum value - or are some of them recommendations about the minimum value - or perhaps the recommendation for the only sensible value ? There's a big difference between offering a default and giving a recommendation.

    This strikes me as a tad nitpicky. You admit that LGWR has factotums, and whether or not these slave processes appear as multiple LGWR processes was never my point.
    I don't have to "admit" that LGWR has 'factotum process' - that seems to be just a circumlocution for "you get one LGWR process and and may get multiple I/O slaves" - and you do get multiple I/O slaves if you have adjusted dbwr_io_slaves (or lgwrio_slaves). If you really want to believe that an I/O slave is the same thing as a LGWR process that's your problem. (Try killing I202 (say), then try killing the LGWR process and see nit-picky Oracle is about the difference).


    You mean transactions_per_rollback_segment, right?
    Let's think, maybe that a mistype for "log_buffer" ? No, on second thoughts "transaction_per_rollback_segment" looks almost identical to "transactions_per_rollback_segment". So now that we've established the difference between mistyping a word and typing the wrong word completely, is there any explanation for that parameter appearing in an article about log buffer sizing ?


    I must strongly disagree. Failure to follow the instructions on Metalink can leave a whole shop unsupported
    Logical fallacy again - I say it may be wrong, and you strongly disagree on the basis that failure to follow instructation can ... Now, if I had said "Metalink is always wrong" you would, indeed, only have to point out one exception. (Bet you there are people reading this thread who have run into problems because they followed Metalink instructions - which could, of course, have left them supported but not very happy the quality of the support).


    Metalink is the final word, directly from the software engineers
    You think Oracle has software engineers busy writing Metalink notes ? They're probably busy producing the software. I don't like using the 'argument from authority' but I do know some of the people who write some of the Metalink notes - and I know that they even they don't get easy access to the software engineers. I can't make any definitive statement about the percentage of Metalink that has come directly from the software engineers - but I'd be prepared to bet that not much of it did.


    For example, to “prove” that multiple LGWR processes don’t exist ...
    Bertrand Russell's teapot again. I can't "prove" that multiple LGWR processes won't come into existence, but there is sufficient evidence to show that the statement has a very high probability of being true.


    Let me add one more assertion:
    * After the Oracle software and disk I/O sub-system are optimized, there are very limited options for improving throughout:
    - Use faster disks (SSD)
    - Turn-off ARCHIVELOG mode (using triple mirrored disks and hardware-level incremental backups)

    That seems like a fairly limited set of options - but I'd have to point out to anyone who tried the second one that if DBWR managed to corrupt (say) the system tablespace, they might have a very expensive recovery problem.

    Regards
    Jonathan Lewis
    http://jonathanlewis.wordpress.com
    http://www.jlcomp.demon.co.uk

    Message was edited by: Jonathan Lewis to italicise one paragraph with a missing tag.
  • 16. Re: log_buffer
    Jonathan Lewis Oracle ACE Director
    Currently Being Moderated
    Your comments in italic:
    Regarding the multiple log writer processes, I got this, from Steve Karam (OCM):
    It's called the "argument from authority", but you don't have to be an OCM to have a Metalink account and do some cut and paste. Moreover, the cut and paste simply states (as it did the last time the note was referenced) that you get I/O slaves.

    You could try asking Steve Karam (OCM and employee of Burleson Consulting) what the difference is between the LGWR process and a set of I/O slaves.

    (Note: It is NOT necessary for the same process to use the same adaptor number always. What this implies is that the adaptor numbers are not 'reserved' for processes.)
    That's an interesting detail. I haven't seen a system using I/O slaves for quite some time, but my memory is that the dbwr I/O slaves always seemed to be I1nn and the lgwr slaves (often only one or two of the default 4) were I2nn

    Thus a typical Unix ps -ef listing after instance startup may look like :
    There you go - it's the answer to the first question you asked above - you don't see LGW(n) - you see IXnn.

    Regards
    Jonathan Lewis
    http://jonathanlewis.wordpress.com
    http://www.jlcomp.demon.co.uk
  • 17. Re: log_buffer
    108476 Journeyer
    Currently Being Moderated
    <p>Hi Jonathan,</p>
    <p><i>>> but at the time I was trying to be kind by not pointing out that it was
    another Burleson article that I was criticising.</i></p>
    <p>You know, I get comments from people asking me why you resort to
    really-obvious sophistry.  It only fools beginners, you know . . </p>
    <p><i>>> You still haven't grasped the difference between <b>the</b> LGWR
    process and I/O slaves.</i></p>
    <p>Oh, I've got it now.  You were just playing a "word game":</p>
    <p>- Are there multiple log writer processes?  Yes?</p>
    <p>- Are there multiple LGW(n) processes?  No, and I never said that there
    were any.</p>
    <p><b>All this nonsense, just because I abbreviate "log writer" as LGWR?</b></p>
    <p>It reminds me of that hilarious "proof", where you went-on for pages when you
    could have simply said: "Don did not mention an obscure exception, and if you
    use the MTS, the results will be different".</p>
    <p>This reminds me of the Gary Larson "Far Side" HELF cartoon:</p>
    <p>http://thethaifamily.com/son/jokes/helf.jpg</p>
    <p><i>>> It would have been nice to have an explanation in the original text
    about why the cases were so clear cut.</i></p>
    <p>There!  You did it again!  I never said that!  Tuning Oracle
    is RARELY clear cut, you know that!</p>
    <p><i>>> It's quite hard to get to the people who wrote the code by logging SRs.</i></p>
    <p>If you say so. . . </p>
    <p><i>>> there is a difference (perhaps subtle) between "providing a
    counter-example" and "arguing from the particular to the general". </i></p>
    <p>That's a problem with all counter-examples, not just yours. </p>
    <p>It's like trying to "prove" that the Loch Ness monster doesn't want tree
    fitty. </p>
    <p>>> Bertrand Russell's teapot</p>
    <p>Ah Bertrand Russell, my favorite [favourite] nihilist.  I like this
    Russell quote, appropriate to this discussion: </p>
    <p>"The most savage controversies are about matters as to which there is no good
    evidence either way."</p>
    <p>BTW, Russell used teapots frequently in analogy.  When folks hear
    "Russell's teapot", it's normally associated with "Russell's paradox", but I
    believe that you are referring to his "celestial teapot" argument, not the same
    thing.</p>
    <p><i>>> No need to put in the [sic].  I'm sure the readers of the forum
    are aware that you're only copying my English spelling of the word, and haven't
    made a spelling mistake.</i></p>
    <p>I thought it might serve to remind you that there are many "correct" ways to
    communicate.  Think about it.  </p>
    <p><i>>> You claim that the log buffer <b>has to be</b> a multiple of 2KB</i></p>
    <p>Jeesh!  PLEASE, stop putting words in my mouth.  Metalink said
    that, not me.  I think it's inconsequential and trivial.</p>
    <p><i>>> No, on second thoughts "transaction_per_rollback_segment" looks almost
    identical to "transactions_per_rollback_segment".</i></p>
    <p>Sort of like making a huge distinction between my use of the terms "log
    writer" and LGWR, right?  </p>
    <p>I'm surprised you took my bait! I really didn't think that you would fall for
    such an obvious set-up.  </p>
    <p><i>>> You think Oracle has software engineers busy writing Metalink notes?</i></p>
    <p>Yes, I know it to be true in many cases.  Not always, but more often
    than not.</p>
    <p>>> I'd have to point out to anyone who tried the second one that if DBWR
    managed to corrupt (say) the system tablespace, they might have a very expensive
    recovery problem.</p>
    <p>Oh yeah, I tried . . . .  But, you know, they can recover in under 30
    minutes with hardware-level B&R.</p>
    <p>It makes me wonder if people will stop taking backups a the Oracle-level, and
    drive it down to the disk hardware?</p>
    <p>BTW, did you get your paper laptop yet?</p>
    <p>http://www.dba-oracle.com/oracle_news/news_cardboard_laptops.htm</p>
  • 18. Re: log_buffer
    Jonathan Lewis Oracle ACE Director
    Currently Being Moderated
    >> I'd have to point out to anyone who tried the second one that if DBWR
    managed to corrupt (say) the system tablespace, they might have a very expensive
    recovery problem.

    Oh yeah, I tried . . . . But, you know, they can recover in under 30
    minutes with hardware-level B&R


    The thing I'd be most concerned about is a block that got corrupted by the hardware or dbwr with a few days (or even hours) time-lag before the corruption was noticed.

    If you adopted this "hardware-level B&R" option on the platforms I am familiar with, the worst case scenario is that you would have to restore the last (hardware) backup of the database, roll in the incremental (hardware) backups up to the one before the corruption occurred (how do you tell which one that is ?), and then start the database with instance recovery. Then you have to bring the data back up to date.

    Alternative accidents could be a lot less painful, of course. A corrupt block in a user index could be resolved by dropping and rebuilding the index. A corrupt block in a user table could be resolved by extracting data whilst skipping the corrupt block - but then you might want to recreate the missing data.


    Running in noarchivelog mode is, to me, a safe option only if you can seriously argue the case that you don't mind running the risk of losing some data.

    Regards
    Jonathan Lewis
    http://jonathanlewis.wordpress.com
    http://www.jlcomp.demon.co.uk
  • 19. Re: log_buffer
    The Human Fly Oracle ACE Director
    Currently Being Moderated
    >>
    You claim that the log buffer has to be a multiple of 2KB
    Jeesh! PLEASE, stop putting words in my mouth. Metalink said
    that, not me. I think it's inconsequential and trivial.
    >>

    Do you think we need to agree and rely whatever metalink note says?

    If something is written in the metalink, I am sure, its not guarantee all time to be right. For an instance, just couple of days I go, I was trying to add a new instance to an existing windows 2 node RAC and I found a metalink note which has couple of missing points, those points are covered from the notes which I got from the internet.

    In my opinion, experts would suggest based on their experience, practical knowledge and their testing result, not just simply referring from the metalink notes.

    >>
    Oh yeah, I tried . . . . But, you know, they can recover in under 30
    minutes with hardware-level B&R.
    >>

    Does this guarantee compelte recovery of the database and 30 minutes downtime for an financial institue would cost too much $$$ lost..

    Finally, we believe in experts comments which comes based on their extensive knolwedge.

    Jaffar
  • 20. Re: log_buffer
    108476 Journeyer
    Currently Being Moderated
    Hi Jaffar,
    Do you think we need to agree and rely whatever metalink note says?
    For patch requirements and notes that say "if you do this, we won't support you", yes, absolutely! For everything else, no!
    Does this guarantee compelte recovery of the database and 30 minutes downtime for an financial institue would cost too much $$$ lost
    Good point! This was a non-critical app, and it was not the end of the world if it crashed sporadically. Howeverm last I heard, they had been running for over 5 years without interruption.
    we believe in experts comments which comes based on their extensive knolwedge.
    Just knowledge alone, no experience necessary? I wonder how the landscape will change after Oracle starts certifying people as experts. I can see it now: "Don't worry, I'm an Oracle certified expert":

    http://www.dba-oracle.com/oracle_news/news_oracle_expert_certification.htm
  • 21. Re: log_buffer
    The Human Fly Oracle ACE Director
    Currently Being Moderated
    Burleson,

    >>
    Just knowledge alone, no experience necessary? I wonder how the landscape will change after Oracle starts certifying people as experts. I can see it now: "Don't worry, I'm an Oracle certified expert":
    >>

    It seems you haven't read my earlier point, where I said the following:
    "
    In my opinion, experts would suggest based on their experience, practical knowledge and their testing result, not just simply referring from the metalink notes.
    "

    When I said practical knowledge and their testing result, it is pretty clear that it will be real experience, just not the books or documents knowledge.

    I am 100% agree with you about the point which you have mentioned regarding an Oracle Certified Experts certifications and in fact on one of the forum, I too asked the same quetion as yo mentioned:

    "I firmly believe that becoming an expert in any technology requires many years of experience."

    Anybody take the Beta Version of Oracle RAC Expert Test?

    Oracle has added another an Oracle certified Expert certification (beta) for SQL recently.

    Well, for the sake of certification, I have registered myself for this beta exam.


    Jaffar
  • 22. Re: log_buffer
    108476 Journeyer
    Currently Being Moderated
    Hi Jaffar,
    Oracle has added another Oracle certified Expert for SQL recently.
    Yeah, I saw that. Laurent Schneider has signed-up for it:

    http://laurentschneider.com/wordpress/2007/07/oracle-certified-sql-expert.html

    Of course, Laurent has "real" experience, so it will be interesting to read his take on it . . .

    Also, remember that guy who advertized the services of a fictional Oracle expert, using a stock photo?

    http://www.dba-oracle.com/oracle_news/2004_11_22a_webb.htm
    experts would suggest based on their experience, practical knowledge and their testing result, not just simply referring from the metalink notes.
    Yes, but for me, it's the "collective wisdom" (CW) that matters more than any single observation. If one experienced person says something, that's good, but if I see 20 experienced people saying something, I give it more credibility . . .

    Message was edited by:
    burleson
  • 23. Re: log_buffer
    Jonathan Lewis Oracle ACE Director
    Currently Being Moderated
    Alan,

    You're quite right.

    Sometimes it's a little hard, though, to decide what's important and what isn't - and why. I can't find a way to delete the post, so I've overwritten it with this message.

    Regards
    Jonathan Lewis
    http://jonathanlewis.wordpress.com
    http://www.jlcomp.demon.co.uk
  • 24. Re: log_buffer
    546494 Journeyer
    Currently Being Moderated
    don't take this the wrong way but can we please keep the petty arguements between people out of the forum.

    It makes the forum something less than it should be.

    Alan
  • 25. Re: log_buffer
    The Human Fly Oracle ACE Director
    Currently Being Moderated
    I personally think that if the argument is in professional way to prove the the point right, then, its really worthy to have arguments.

    Jaffar
  • 26. Re: log_buffer
    587954 Newbie
    Currently Being Moderated
    Don & Jonathan, please stop now; it's all a bit too Tom & Jerry.

    This (seemingly ancient) Jonathan vs. Don/Don vs. Jonathan thing has gone way beyond being prettified with the often trotted-out, "But I'm just a professional correcting/rebuffing vigorously."; it's now nothing more than two talented adults who seem to adore kicking the other in the b*lls, while, IMHO, many of the Oracle community must be watching in increasing disbelief and growing boredom.

    Come on gents, give it a rest. Just try it for a week (how about forever?) ... please. We'll all still remember that you exist, I promise.
  • 27. Re: log_buffer
    108476 Journeyer
    Currently Being Moderated
    Hi,
    it's all a bit too Tom & Jerry
    I've heard of Tom, but who is Jerry? Jerry Springer?

    BTW, I find the "hijacking" of this thread as naughty as you do, but I could not just sit-by while (whilst) somebody proffered a misleading proof:

    http://www.dba-oracle.com/t_biased_test_cases.htm

    If the OP is still here, I apologize (apologise). . . .

    I was kicked in the b*lls, and I should not have taken the bait . . .


    Sincerely,

    Don Burleson
    Nursing my wounded b*lls . . .

    Message was edited by:
    burleson
  • 28. Re: log_buffer
    16279 Newbie
    Currently Being Moderated
    Dear Doanld,
    "somebody proffered a misleading proof"

    you said on http://dba.ipbhost.com post (01/04/2005)
    "I was interviewing a DBA yesterday (30 years DBA experience, Certified Oracle Master, ex-Oracle University instructor), ...."

    considering that the first RDBMS was not available until 1978, Multics Relational Data Store if i remember well..

    mmhh, interesting, so if mathematics does not represent an opinion for you

    2005-30 = 1975

    before 1978!!!! Gasp therefore you would have interviewed Adamo of the DBAs !!!!!


    credibility is very important thing and it is established with the intelligence and the reasonableness of that everyone asserts

    Regards
    Alessandro
  • 29. Re: log_buffer
    108476 Journeyer
    Currently Being Moderated
    Hi Alessandro,

    You really think that DBA's did not exist in 1975? Really?

    There were large numbers of DBA's in the 1950 (back when shops wrote their own persistent storage managers using BAL), and a giant flood of DBA's after IMS came out in the late 1960's . . .

    Here is an article I wrote about pre-relational databases, back in 1995, where I personally interviewed Ted Codd:

    http://www.dba-oracle.com/art_sm_legacy_relational_migration.htm

    Ted noted that in 1995, many shops had not yet moved to relational databases: "It is a serious problem to convert between databases such as IMS and relational because a lot of the activities with IMS were developed by application programmers instead of DBAs"
    30 years DBA experience
    Yeah, that's 100% true.

    I have 25 years of DBA experience myself, are you calling me a liar, Alessandro?
    Gasp therefore you would have interviewed Adam of the DBAs !!!!!
    Yeah, your can call older DBA's "Adam's", although I prefer the term "wild-eyed old coot DBA's".
    credibility is very important thing
    Agreed, but what about folks who falsely attack the credibility of others?

    ******************
    Edited to add:
    ******************
    Certified Oracle Master
    FYI, way-back before Oracle introduced the Oracle Certified Master program (OCM), Oracle issued Certified Oracle Master (COM) certificates . . .

    Message was edited by:
    burleson