home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!ukma!netsys!decwrl!gossip.pyramid.com!pyramid!infmx!davek
- From: davek@informix.com (David Kosenko)
- Newsgroups: comp.databases.informix
- Subject: Re: SE vs OL - Lets get to the crux.
- Message-ID: <1992Nov9.170244.20742@informix.com>
- Date: 9 Nov 92 17:02:44 GMT
- References: <1992Nov6.194129.4846@elsouth.uucp>
- Sender: news@informix.com (Usenet News)
- Organization: Informix Software, Inc.
- Lines: 83
-
- John White writes:
- > On the other hand, I do think think this is the right forum for
- >one of the non-application-specific issues that was brought up.
- >It has been repeatedly pointed out that OnLine doesn't show its
- >performance increase at lower user numbers. I've been promised that
- >OnLine is buttloads better when the user count goes up. I don't
- >think so.
- >
- > TEST:
- > o Put 50 users on each of two machines
- > o One machine running OnLine
- > o One running Standard Engine
- > o Have all of the users pull up perform screens.
- > o Have the users do nothing (easy).
- > o Run database performance tests on each machine.
- >
- > I think you'll get very similar numbers to when there is a single
- >user. I submit that the important factor in the equation is not the
- >number of users, it is more likely the number or transactions per
- >second, or the number of transactions per user per second.
-
- Well of course users doing nothing do not affect the performance. I'm sure
- most of those (myself included) who suggested more users meant more users
- doing work! TPS is of course, key, but with a single user it is almost
- impossible to stress OnLine sufficiently to get anything approaching peak
- performance from it. For that you need a lot of users doing a lot of work,
- i.e. a lot of transactions.
-
- > Having said that and feeling the incoming corrections in my
- >peripheral conscience, I ask four things:
- >
- > 2) If I'm wrong, what is the correct answer? What is that elusive
- > most-important-factor (X), that as X increases, the performance
- > of OnLine does not degrade as rapidly as Standard Engine.
-
- You asking for X is , in my opinion, misplaced. Buffering of data in
- shared memory is pretty much what buys OnLine performance over SE (at least
- in the simplest analysis). Raw disk access helps too, but only if
- you are doing a lot of i/o relative to processing.
-
- Again, keep in mind that OnLine is doing a bit more work to achieve the
- same end as SE (for example, physical logging, an OnLine consistency/recovery
- mechanism, is not present at all in SE). This added work cost gets amortized
- over higher user counts, to the point where it becomes a benefit.
-
- Also consider shared memory buffering. If two users want the same data, one
- of them causes the data to be read from disk into shared memory. The second
- user can just read it directly from shared memory. So the i/o cost for
- such an operation for one user is X, that same operation for two users is X/2,
- and so on. As the user count increases, the cost of that i/o for any single
- user gets smaller. With SE, each user pays cost X, regardless of how many
- users there are (and when updates are added into the equation, the costs
- increase considerably for SE, but very little for OnLine).
-
- > 3) How does a person measure this factor on his active system? How
- > can I say with confidence that at my shop X = 4.63?
-
- Have your benchmark time operations against the database for each user.
- For example, perform inserts into the table and time a number of inserts
- for each user. This gives you inserts/second. Run 1 user, 2 users, etc.
- For multi-user runs, sum the resulting inserts/second. Perform the tests
- on OnLine and SE. You can use selects (count records fetched), updates,
- deletes, etc. Compare the results. Unless you have your OnLine system
- tuned extremely poorly, you will no doubt find that as the user count
- increases, SE performs worse while OnLine performs better.
-
- > 4) What is the break point? At what point do the performance gains
- > that OnLine brings make up for its overhead costs? i.e. When X > 2.
-
- Depends on the application, tuning, CPU, etc. etc. For an example, on
- an RS6000 running a more-or-less tpc-B benchmark, an untuned OnLine
- system peaked (for me) at about 25 users. This means that after 25, the
- performance started to drop. Tuning the machine and OnLine could have
- increased that substantially (esp. since I had one disk for everything),
- but the purpose of that exercise was to see what "out of the box" performance
- looked like. You number will depend on many different variables.
-
- Dave
- --
- Disclaimer: These opinions are not those of Informix Software, Inc.
- **************************************************************************
- "I look back with some satisfaction on what an idiot I was when I was 25,
- but when I do that, I'm assuming I'm no longer an idiot." - Andy Rooney
-