_I have written two posts in one, about a performance issue with writes in system tablespace and introduction of a little DTrace stack sampling script.
_
Have you noticed that DML on tables residing in SYSTEM tablespace is slower than tables in other tablespaces?
Here’s an example, I’ll create two similar tables, one in USERS tablespace, other in SYSTEM, and inset into the first one (Oracle 10.2.0.3 on Solaris x64):
SQL> create table t1(a int) tablespace USERS; Table created. SQL> create table t2(a int) tablespace SYSTEM; Table created. SQL> exec for i in 1..100000 loop insert into t1 values(i); end loop; PL/SQL procedure successfully completed. Elapsed: 00:00:03.09
Insert into table in USERS tablespace took 3 seconds.
Ok, let’s commit and flush dirty buffers that they wouldn’t get on the way of next insert.
SQL> commit; Commit complete. Elapsed: 00:00:00.02 SQL> alter system checkpoint; -- checkpointing to flush dirty buffers from previous inserts System altered. Elapsed: 00:00:01.34 SQL>
And now to the insert into the SYSTEM tablespace table:
SQL> exec for i in 1..100000 loop insert into t2 values(i); end loop; PL/SQL procedure successfully completed. Elapsed: 00:00:08.98 SQL>
What?! The same insert took 3 times longer, almost 9 seconds?
Fine! Let’s troubleshoot it!
First (as usual) I start from my Snapper for getting a performance snapshot of the session. I ran both inserts again and used Snapper from another session to monitor the insert activity (for brevity I didn’t monitor all stat counters, but used the sinclude filter to leave only various “gets” in from V$SESSTAT).
This is the insert to USERS tablespace table:
SQL> @snapper out,gather=tsw,sinclude=gets 10 1 148 -- Session Snapper v1.07 by Tanel Poder ( ) --------------------------------------------------------------------------------------------------------------------------------------------- HEAD, SID, SNAPSHOT START , SECONDS, TYPE, STATISTIC , DELTA, DELTA/SEC, HDELTA, HDELTA/SEC --------------------------------------------------------------------------------------------------------------------------------------------- DATA, 148, 20080901 17:20:53, 10, STAT, db block gets , 103448, 10345, 103.45k, 10.34k DATA, 148, 20080901 17:20:53, 10, STAT, db block gets from cache , 103448, 10345, 103.45k, 10.34k DATA, 148, 20080901 17:20:53, 10, STAT, consistent gets , 495, 49, 495, 49.5 DATA, 148, 20080901 17:20:53, 10, STAT, consistent gets from cache , 495, 49, 495, 49.5 DATA, 148, 20080901 17:20:53, 10, STAT, consistent gets - examination , 219, 22, 219, 21.9 DATA, 148, 20080901 17:20:53, 10, STAT, no work - consistent read gets , 89, 9, 89, 8.9 DATA, 148, 20080901 17:20:53, 10, STAT, rollbacks only - consistent read gets , 18, 2, 18, 1.8 DATA, 148, 20080901 17:20:53, 10, STAT, cluster key scan block gets , 65, 6, 65, 6.5 DATA, 148, 20080901 17:20:53, 10, TIME, hard parse elapsed time , 4090, 409, 4.09ms, 409us DATA, 148, 20080901 17:20:53, 10, TIME, PL/SQL compilation elapsed time , 3139, 314, 3.14ms, 313.9us DATA, 148, 20080901 17:20:53, 10, TIME, parse time elapsed , 5131, 513, 5.13ms, 513.1us DATA, 148, 20080901 17:20:53, 10, TIME, PL/SQL execution elapsed time , 174288, 17429, 174.29ms, 17.43ms DATA, 148, 20080901 17:20:53, 10, TIME, DB CPU , 3108026, 310803, 3.11s, 310.8ms DATA, 148, 20080901 17:20:53, 10, TIME, sql execute elapsed time , 3104607, 310461, 3.1s, 310.46ms DATA, 148, 20080901 17:20:53, 10, TIME, hard parse (sharing criteria) elapsed ti, 4110, 411, 4.11ms, 411us DATA, 148, 20080901 17:20:53, 10, TIME, DB time , 3108026, 310803, 3.11s, 310.8ms DATA, 148, 20080901 17:20:53, 10, WAIT, log file switch completion , 71104, 7110, 71.1ms, 7.11ms DATA, 148, 20080901 17:20:53, 10, WAIT, SQL*Net message to client , 3, 0, 3us, .3us DATA, 148, 20080901 17:20:53, 10, WAIT, SQL*Net message from client , 6684030, 668403, 6.68s, 668.4ms -- End of snap 1 PL/SQL procedure successfully completed.
You see, the DB CPU equals to DB time (in other words all the time spent servicing my request in database was spent on CPU, thus there were no waits, blocks involved). And the insert still took roughly 3 seconds.
This is the insert to SYSTEM tablespace table:
SQL> @snapper out,gather=tsw,sinclude=gets 10 1 148 -- Session Snapper v1.07 by Tanel Poder ( ) --------------------------------------------------------------------------------------------------------------------------------------------- HEAD, SID, SNAPSHOT START , SECONDS, TYPE, STATISTIC , DELTA, DELTA/SEC, HDELTA, HDELTA/SEC --------------------------------------------------------------------------------------------------------------------------------------------- DATA, 148, 20080901 17:21:23, 11, STAT, db block gets , 102616, 9329, 102.62k, 9.33k DATA, 148, 20080901 17:21:23, 11, STAT, db block gets from cache , 102616, 9329, 102.62k, 9.33k DATA, 148, 20080901 17:21:23, 11, STAT, consistent gets , 384, 35, 384, 34.91 DATA, 148, 20080901 17:21:23, 11, STAT, consistent gets from cache , 384, 35, 384, 34.91 DATA, 148, 20080901 17:21:23, 11, STAT, consistent gets - examination , 218, 20, 218, 19.82 DATA, 148, 20080901 17:21:23, 11, STAT, no work - consistent read gets , 94, 9, 94, 8.55 DATA, 148, 20080901 17:21:23, 11, STAT, rollbacks only - consistent read gets , 17, 2, 17, 1.55 DATA, 148, 20080901 17:21:23, 11, STAT, cluster key scan block gets , 65, 6, 65, 5.91 DATA, 148, 20080901 17:21:23, 11, TIME, hard parse elapsed time , 2350, 214, 2.35ms, 213.64us DATA, 148, 20080901 17:21:23, 11, TIME, PL/SQL compilation elapsed time , 1751, 159, 1.75ms, 159.18us DATA, 148, 20080901 17:21:23, 11, TIME, parse time elapsed , 3359, 305, 3.36ms, 305.36us DATA, 148, 20080901 17:21:23, 11, TIME, PL/SQL execution elapsed time , 172153, 15650, 172.15ms, 15.65ms DATA, 148, 20080901 17:21:23, 11, TIME, DB CPU , 8751163, 795560, 8.75s, 795.56ms DATA, 148, 20080901 17:21:23, 11, TIME, sql execute elapsed time , 8749139, 795376, 8.75s, 795.38ms DATA, 148, 20080901 17:21:23, 11, TIME, hard parse (sharing criteria) elapsed ti, 2367, 215, 2.37ms, 215.18us DATA, 148, 20080901 17:21:23, 11, TIME, DB time , 8751163, 795560, 8.75s, 795.56ms DATA, 148, 20080901 17:21:23, 11, WAIT, SQL*Net message to client , 3, 0, 3us, .27us DATA, 148, 20080901 17:21:23, 11, WAIT, SQL*Net message from client , -1442849, -131168, -1.44s, -131.17ms -- End of snap 1 PL/SQL procedure successfully completed.
You see the DB Time is still equal to DB CPU, thus all response was spent on CPU. Moreover, the time model stats tell me that all the elapsed time was spent executing SQL statements (which makes sense as I’m running INSERT SQL in a tight loop).
So, for some reason the INSERT SQL statements take almost 3 times longer when inserting to SYSTEM tablespace.
In cases where there is no waiting going on (things like Oracle Wait Interface / WaitProf would not help here), but only difference in CPU time, I normally look into V$SESSTAT counter differences first. But as seen from above, there is no major difference in counters such logical IOs (db block gets, consistent gets) to justify 3x increase in CPU usage (also there was no major difference in other counters which I have excluded for brevity).
So, the obvious next step is to sample the process stack during execution of both statements and see where the difference is. In my previous posts I’ve used pstack for getting a few stack samples out of a process and a pstack loop could be used here as well.
But for demo purposes (and convenience) I will use my DTrace stack sampling script today.
It’s dead easy to sample user process and kernel stacks using DTrace in Solaris 10 (and MacOS X and FreeBSD too). However the DTrace stack aggregator also takes account of the current program counter offset inside functions when distinguishing between stacks (this means that CPU program counter location main() + 0x10 will be different from main() + 0x12). But I don’t want such granularity, I want to see where the execution was at function level, not instruction level.
So I wrote my script to strip off the PC function offsets from DTrace output and order the stacks the way that most frequently sampled stack “branch” is displayed last, in bottom of the output. So, whatever stack and function you see outputted last, is probably the “troublemaker”. I still recommend scrolling up and reviewing few other stacks too.
So, I ran the insert to USERS tablespace table again and ran my dstackprof script on the server process with SPID 859.
This is from the “well behaving” process, so normally you should just compare this stack profile to the “bad behaving” process stack profile and see if you can spot any major differences in these. I will explain few of the functions on the go.
# ./dstackprof.sh 859 DStackProf v1.02 by Tanel Poder ( ) Sampling pid 859 for 5 seconds with stack depth of 100 frames... 147 samples with stack below __________________ kcbchg1_main kcbchg1 <- buffer change function in buffer cache layer ktuchg <- undo layer generate change ktbchg2 <- block (header) change kdtchg <- data layer generate change kdtwrp kdtInsRow <- data layer row insert function insrowFastPath insdrvFastPath inscovexe insExecStmtExecIniEngine insexe <- OPI insert execution entry point opiexe <- OPI execute (executing my recursive INSERT INTO statement) opipls opiodr rpidrus <- all the RPI functions belong to recursive program interface (which allows making recursive calls in db) skgmstack rpidru <- these RPI functions (and few OPI ones above) are what PL/SQL context switches physically are rpiswu2 rpidrv psddr0 <- most of the functions starting with "p" are PL/SQL execution ones psdnal pevm_EXECC pfrinstr_EXECC pfrrun_no_tool pfrrun plsql_run peicnt <- PL/SQL execution engine entry point kkxexe <- cursor to PL/SQL execution interface opiexe <- OPI call execution dispatcher kpoal8 opiodr ttcpip opitsk opiino opiodr opidrv sou2o opimai_real main 0xe54ffc 193 samples with stack below __________________ libc.so.1`times <-- during these 193 samples the execution was in opiexe (note that opiexe -- it's a recursive OPI call execution (as there's the RPI layer below in the opipls -- stack and there's another opiexe below. The lower opiexe is for executing opiodr -- my PL/SQL anonymous block and upper is for the recursive INSERT statement rpidrus skgmstack rpidru rpiswu2 rpidrv psddr0 psdnal pevm_EXECC pfrinstr_EXECC pfrrun_no_tool pfrrun plsql_run peicnt kkxexe opiexe kpoal8 opiodr ttcpip opitsk opiino opiodr opidrv sou2o opimai_real main 0xe54ffc 3147 Total samples captured
Anyway, the two stacks which I have copied here don’t show anything obvious. And also, see that there were total 3147 samples taken when the process was on CPU. But the execution point was in the “top” stack during only 193 samples, which is around 6% of all samples.
As the above stacks come from a well-behaving process, let’s leave their deeper analysis and just compare them to the bad-behaving process. So, I ran the SYSTEM tablespace insert again (you still remember, we were troubleshooting why inserts into system tablespace were slower, right ;)
The top stack in profile was following:
# ./dstackprof.sh 859 DStackProf v1.02 by Tanel Poder ( ) Sampling pid 859 for 5 seconds with stack depth of 100 frames... [...some output snipped...] 917 samples with stack below __________________ libc.so.1`qsort kd4_entries_overlap kdb4chk kd4chk kcbchk kco_blkchk <-- what the heck is this? the name suggests that it's block checking of some kind kcoapl kcbapl kcrfw_redo_gen <-- also note this function, looks like the block change function (kcbchg) has called kcbchg1_main <-- a function to generate redo on its behalf kcbchg1 ktuchg ktbchg2 kdtchg kdtwrp kdtInsRow insrowFastPath insdrvFastPath inscovexe insExecStmtExecIniEngine insexe opiexe opipls opiodr rpidrus skgmstack rpidru rpiswu2 rpidrv psddr0 psdnal pevm_EXECC pfrinstr_EXECC pfrrun_no_tool pfrrun plsql_run peicnt kkxexe opiexe kpoal8 opiodr ttcpip opitsk opiino opiodr opidrv sou2o opimai_real main 0xe54ffc 6202 Total samples captured
As most of the top stacks from the bad behaving process were spending their time somewhere in kco_blkchk() function, it rang a bell – database block checking!
However it was weird as the db_block_checking parameter was set false in my instance:
SQL> show parameter db_block_checking NAME TYPE VALUE ------------------------------------ ----------- ----- db_block_checking string FALSE SQL>
After a little digging, I found this parameter:
NAME VALUE DESCRIPTION ----------------------------- --------- ------------------------------------------------------------- _db_always_check_system_ts TRUE Always perform block check and checksum for System tablespace
This parameter makes Oracle (since version 8.1.6) do block structure checking in SYSTEM tablespace even if db_block_checking itself is set to FALSE. Considering the criticality of data dictionary objects, it’s a good idea.
So, now you might say: so what that there’s a performance issue, you shouldn’t keep your application tables in SYSTEM tablespace anyway!
Well, I have three words for you: AUD$, FGA_LOG$, DEF$_AQCALL :)
If you rely on auditing functionality heavily and especially if you use advanced replication, having the crucial tables (and associated LOB segments) in SYSTEM tablespace can mean quite a hit to your CPUs and system throughput. In which cases you probably would want to explore the option of moving these tables out of system tablespace.
Warning: before moving anything, consult Metalink or Oracle Support as only few strictly defined ways for moving these object are supported. See notes 731908.1 for Audit and 1037317.6 for Advanced Replication.
So, the key point of this article is that I could have started guessing or setting various events or undocumented parameters for probing Oracle here and there, for figuring out what’s going on. But I used a simple systematic approach instead. When Oracle Wait Interface and V$SESSTAT counters didn’t help, looked into process stack next. And this gave me the vital clue (kdo_blkchk) that there’s block checking going on. And if I hadn’t figured out the cause myself, I would have had some solid evidence for support.
Note that you don’t have to use DTrace for this kind of stack profiling (even though it is very convenient). On non-Solaris 10 OS’es I would use whatever safe stack dump mechanism is available (pstack for example) and post-process the results. I probably wouldn’t take thousands of samples though, but probably 100 or even less.
Also, if you want to get a complete overview of the process execution profile, you would need to do much more complex aggregations, as you’d need to aggregate “child” branches of stack samples with their “parents” (currently the samples of a backtrace “a->b->c” are not accounted in “a->b” for example). But for my case (just checking if there’s any major difference between execution profiles) this approach was good enough.
By the way (if you haven’t given up yet :), check the differences in LatchProfX output. The top is from well-behaving process and bottom one is the bad one:
SQL> @latchprofx sid,name,ksllwnam,ksllwlbl 128 % 1000000
-- LatchProfX 1.06 by Tanel Poder ( )
SID NAME KSLLWNAM KSLLWLBL Held Gets Held % Held ms Avg hold ms
---------- ----------------------------------- ---------------------------------------- -------------------- ---------- ---------- ------- ----------- -----------
128 redo copy kcrfw_redo_gen: nowait latch 49399 48937 4.94 674.790 .014
128 cache buffers chains kcbchg1: kslbegin: bufs not pinned 4574 2320 .46 62.481 .027
128 cache buffers chains kcbchg1: kslbegin: call CR func 2101 1445 .21 28.700 .020
128 cache buffers chains kcbgcur_2 1323 1060 .13 18.072 .017
128 cache buffers chains kcbnew: new latch again 269 243 .03 3.675 .015
128 simulator lru latch kcbs_simulate: simulate set setid 170 170 .02 2.322 .014
128 redo allocation kcrfw_redo_gen: redo allocation 1 strand # 68 13 .01 .929 .071
128 cache buffers chains kcbrls_2 29 29 .00 .396 .014
128 cache buffers chains kcbgtcr: fast path 25 25 .00 .342 .014
128 cache buffers chains kcbgcur: fast path (shr) 8 8 .00 .109 .014
128 cache buffers lru chain kcbzgws_1 7 7 .00 .096 .014
SQL> @latchprofx sid,name,ksllwnam,ksllwlbl 128 % 1000000
-- LatchProfX 1.06 by Tanel Poder ( )
SID NAME KSLLWNAM KSLLWLBL Held Gets Held % Held ms Avg hold ms
---------- ----------------------------------- ---------------------------------------- -------------------- ---------- ---------- ------- ----------- -----------
128 redo copy kcrfw_redo_gen: nowait latch 356465 79839 35.65 6088.422 .076
128 cache buffers chains kcbchg1: kslbegin: bufs not pinned 4704 2233 .47 80.344 .036
128 cache buffers chains kcbchg1: kslbegin: call CR func 3542 1997 .35 60.497 .030
128 cache buffers chains kcbgcur_2 1929 1448 .19 32.947 .023
128 redo allocation kcrfw_redo_gen: redo allocation 1 strand # 1018 34 .10 17.387 .511
128 shared pool simulator kglsim_unpin_simhp: fast path 713 713 .07 12.178 .017
128 simulator lru latch kcbs_simulate: simulate set setid 266 266 .03 4.543 .017
128 cache buffers chains kcbnew: new latch again 182 171 .02 3.109 .018
128 cache buffers chains kcbgtcr: fast path (cr pin) 29 29 .00 .495 .017
128 cache buffers chains kcbgtcr: fast path 15 15 .00 .256 .017
See, how the redo copy latches were held for much longer (both total time and average time of individual hold) when the block checking was enabled. This indicates that in addition to keeping the buffer pinned, also a redo copy latch is held during block checking (if not using an IMU transaction). If you use high-throughput advanced replication, think that CPU usage is too high for work done and can’t get rid of the redo copy latch contention, this may be a reason.
