Just a quick heads-up for anyone who likes to play around with the Keep and Recycle caches.
In 11g Oracle introduced the option for serial direct path reads for tablescans on tables that was sufficiently large – which meant more than the small_table_threshold – provided the table wasn’t already sufficient well cached. (The rules mean that the choice of mechanism can appear to be a little random in the production environment for tables that are near the threshold size – but if you try testing by doing “alter system flush buffer_cache” you find that you always get direct path reads in testing.)
I’ve just discovered a little oddity about this, though. I have a table of about 50MB which is comfortably over the threshold for direct path reads. But if I create a KEEP cache (db_keep_cache_size) that is a little larger than the table and then assign the table to the KEEP cache (alter table xxx storage(buffer_pool keep)) then 126.96.36.199 stops doing direct path reads, and caches the table.
Now this doesn’t seem unreasonable – if you’ve assigned an object to the KEEP cache you probably want it (or once wanted it) to be kept in cache as much as possible; so using the KEEP cache if it’s defined and specified makes sense. The reason I mention this as an oddity, though, is that it doesn’t reproduce in 188.8.131.52.
I think I saw a bug note about this combination a few months ago- I was looking for something else at the time and, almost inevitably, I can’t find it when I want it – but I don’t remember whether it was the 11.1 or 11.2 behaviour that was deemed to be correct.
See comments 1 and 2 below. I’ve written about this previously, and the caching bechaviour is the correct behaviour. The patch is in 184.108.40.206 and backports are available for 220.127.116.11 and 18.104.22.168. The patch ensures that the table will be read into the cache if it is smaller than the db_keep_cache_size. (Although we might want to check – see Tanel’s notes – whether this is based on the high water mark recorded in the segment header or on the optimizer stats for the table; and I suppose it might be worth checking that the same feature applies to index fast full scans). From the MoS document:
With the bug fix applied, any object in the KEEP buffer pool, whose size is less than DB_KEEP_CACHE_SIZE, is considered as a small or medium sized object. This will cache the read blocks and avoid subsequent direct read for these objects.
Here’s a recent request from the OTN database forum – how do you make this query go faster (tkprof output supplied):
from a, b
where A.MARK IS NULL
and a.cntry_code = b.cntry_code and b.dir_code='XX' and b.numb_type='XXX'
and upper(Trim(replace(replace(replace(replace(replace(replace(replace(a.co_name,'*'),'&'),'-'),'/'),')'),'('),' '))) like
When a “cache read” tablescan (or index fast full scan) takes place we generally expect to see waits on “db file scattered read” as Oracle performs multi-block reads to do the scan. But we all know that Oracle will skip over blocks that are already in the cache and can therefore end up doing multi-block reads of many different sizes, even down to the point where it does single block reads (waiting for “db file sequential read”).
A quirky little question came up on OTN a little while ago: “for a large table we expect multiblock reads to be positioned at the end of the LRU for prompt re-use; but when Oracle does a single block read as part of a tablescan does it go to the end of the LRU (because it’s part of a large tablescan) or does it go to the mid-point of the LRU (because it’s a single block read)?”
The description of how blocks are treated in a tablescan has been simplified, of course, but the question is still valid – so what’s the answer, and how (without going into an extreme level of detail) would you demonstrate it ?
Everybody “knows” that when you do a tablescan of a table that it starts with two buffer gets on the segment header, though older versions (ca. 8i and earlier) of Oracle used to do 4 buffer gets on the segment header. The upshot of this is that many people still say that if you create a table and insert a single row then you’re going to get 3 buffer gets when you tablescan a table: two for the segment header and one for the data block:
So here’s a test, with the second set of autotrace stats which, for reasons I’ll describe shortly, may not be immediately reproducible on your system:
Here’s a point that’s probably fairly well-known, but worth repeating – nvl() and coalesce() aren’t identical in behaviour but you may have some cases where you’re usingnvl() when coalesce() would be a more efficient option.
The reason for this is “short-circuiting”. The expression nvl(expr1, expr2) will return expr2 if expr1 is null, otherwise it will return expr1; the expression coalesce(expr1, expr2, …, exprN) will return the first non-null expression in the list so, in particular, coalesce(expr1, expr2) will give the same result as nvl(expr1, expr2) ; the big difference is that nvl() will evaluate both expressions, while coalesce will evaluate expr2 only if it needs to (i.e. only if expr1 evaluates to null). In many cases the difference in performance will be insignificant – but consider the following type of construct (t1 is a table with a single, numeric, column n1 and a single row):
Here’s an example I saw a few months ago of the confusion caused by NULL. As the owner of the problem put it: the first query, run from SQL*Plus for testing purposes, takes no time to complete; but when “put into a pl/sql cursor” (as shown in the second query) it takes ages to complete.
Here’s one of those little details which I would have said just couldn’t be true – except it’s in the manuals, and the manuals happen to be right.
In an earlier post I’ve described how a distributed query can operate at a remote site if it’s a simple select but has to operate at the local site if it’s a CTAS (create as select) or insert as select. There’s (at least) one special case where this turns out to be untrue … provided you write the query in the correct fashion. I discovered this only as a result of doing a few experiments in response to a question on the OTN database forum.
I have a fairly strong preference for choosing simple solutions over complex solutions, and using Oracle-supplied packages over writing custom code – provided the difference in cost (whether that’s in human effort, run-time resources or licence fees) is acceptable. Sometimes, though, the gap between simplicity and cost is so extreme that a hand-crafted solution is clearly the better choice. Here’s an idea prompted by a recent visit to a site that makes use of materialized views and also happens to be licensed for the partitioning option.
While checking out potential scalability threats recently on a client system, I was directed to a time-critical task that was currently executing the same PL/SQL procedure 16 times (with different parameters) between 6:00 and 7:00 am; as the system went through its next phase of expansion the number of executions of this procedure was likely to grow. An interesting detail, though, was that nothing else was going on while the task was running so the machine (which had 6 cores) was running at 16% CPU.
An obvious strategy for handling the required growth target was to make sure that four (possibly 5) copies of the procedure were allowed to run concurrently. Fortunately the different executions were completely independent of each other and didn’t interfere with each other’s data, so the solution simply required a mechanism to control the parallelism. Conveniently 11gR2 gave us one.
There are a couple of posts on the blog describing problems with updateable join views or, to be more precise, join views which were key-preserved but which the optimizer did not recognize as key-preserved. Both scenarios are addressed in 12c:
Some time ago – actually a few years ago – I wrote a note about the hint /*+ gather_plan_statistics */ making some informal comments about the implementation and relevant hidden parameters. I’ve recently discovered a couple of notes from Alexander Anokhin describing the feature in far more detail and describing some of the misleading side effects of the implementaiton. There are two parts (so far): part 1 and part 2.
Dominic Brooks published a note recently about some very nasty SQL – originally thinking that it was displaying a run-time problem due to the extreme number of copies of the lnnvl() function the optimizer had produced. In fact it turned out to be a parse-time problem rather than a run-time problem, but when I first read Dominic’s note I was sufficiently surprised that I decided to try modelling the query.
Anyone who has used Kevin Closson’s “Silly Little Oracle Benchmark” will want to know about his significantly updated SLOB2.
I was at a client site recently where one of the end-users seemed to have discovered a cunning strategy for optimising a critical SQL statement. His problem was that his query screen times out after 2 minutes, so any query he runs has to complete in less than two minutes or he doesn’t see the results. Unfortunately he had a particular query which took nearly 32 minutes from cold to complete – partly because it’s a seven-table join using ANSI OUTER joins, against tables ranging through the 10s of millions of rows and gigabytes of data – the (necessary) tablescan of the table that had to be first in the join order took 70 seconds alone.
But our intrepid user seems to have made an important discovery and engineered a solution to his performance problem. I think he’s noticed that when you run a query twice in a row the second execution is often faster than the first. I can’t think of any other reason why the same person would run the same query roughly every four minutes between 8:00 and 9:00 am every morning (and then do the same again around 5:00 in the afternoon).
Looking at the SQL Monitoring screen around 10:00 the first day I was on-site I noticed this query with a very pretty graphic effect of gradually shrinking blue bars as 32 minutes of I/O turned into 2 minutes of CPU over the course of 8 consecutive executions which reported run times something like: 32 minutes, 25 minutes, 18 minutes, 12 minutes, 6 minutes, 4 minutes, 2.1 minutes, 2 minutes.
It’s lucky (for that user) that the db_cache_size is 60GB. On the other hand this machine is one of those Solaris boxes that likes to pretend that it’s got 128 CPUs when really it’s only 16 cores with 8 lightweight threads per core – you don’t want anyone running a query that uses 2 solid CPU minute on one of those boxes because it’s taking out 1/16th of your CPU availability, while reporting a load of 1/128 of your CPUs.
Footnote: the query can be optimised (properly) – it accessed roughly 100M rows of data to return roughly 300 rows (with no aggregation), so we just need to do a little bit of work on precise access paths.