Professional Documents
Culture Documents
Advanced RAC Troubleshooting
Advanced RAC Troubleshooting
By Riyaj Shamsudeen
Who am I?
18 years using Oracle products/DBA OakTable member Oracle ACE Certified DBA versions 7.0,7.3,8,8i,9i &10g Specializes in RAC, performance tuning, Internals and E-business suite Chief DBA with OraInternals Email: [email protected] Blog : orainternals.wordpress.com URL: www.orainternals.com
Disclaimer
These slides and materials represent the work and opinions of the author and do not constitute official positions of my current or past employer or any other organization. This material has been peer reviewed, but author assume no responsibility whatsoever for the test cases. If you corrupt your databases by running my scripts, you are solely responsible for that. This material should not be reproduced or used without the authors' written permission.
Concepts
Cache coherency
There are multiple buffer caches in an instance and Oracle RAC uses shared everything architecture. Cache coherency is the method by which consistency of the database is maintained. Only one instance can hold a block in exclusive current mode and a block can be modified only if the block is held in exclusive current mode. There can be two pending transactions modifying the same block, but a block can only be held in exclusive mode in an instance.
If the buffer is not in the Local buffer cache, process identifies the master node of that block. Then the process sends a request to a LMS process running in the master node over the interconnect. While sending the request, it is not known whether the block is in any instance buffer cache. Until LMS responds, User process waits for a place-holder wait event such as gc cr read, gc current read etc. Time is accounted to appropriate events after the response is received from the LMS process.
Demo: demo_01a.sql
OraInternals Riyaj Shamsudeen 6
Block is not in any buffer cache. LMS grants a PR mode lock on the resource and asks FG to read from the disk.
FG LMD LMS
1 2
Buffer
FG LMD LMS
FG LMD LMS
GRD
GRD
Buffer
GRD
Buffer
SGA
SGA
SGA
Directory node for the resource FG Foreground Process LMD Lock Manager Daemon GRD Global Resource Directory
Disk files
Trace lines
Following trace lines shows that session waited for a 2-way grant, followed by a disk read.
WAIT #18446741324875049632: nam='gc cr grant 2-way' ela= 499 p1=7 p2=6867 p3=1 obj#=76484 tim=4597940025 WAIT #18446741324875049632: nam='db file sequential read' ela= 758 file#=7 block#=6867 blocks=1 obj#=76484 tim=4597941129
Lock mode of PR (Protected Read) granted to the instance before reading the block from the disk.
KJBLNAME2 KJBLGRANT KJBLROLE KJBLREQUES
KJBLNAME
If the buffer is in the remote instance in a compatible mode, LMS process grants a lock. Remote LMS process transfers the block to the foreground process. Foreground process copies the buffer to the buffer cache. Instances with that block may acquire lock on that block (CR block transfer does not GRD updates). You can see gc events, but no disk events following the gc events.
WAIT #18446741324875056000: nam='gc current block 2-way' ela= 1453 p1=7 p2=6852 p3=1 obj#=76483 tim=6688296584 FETCH #18446741324875056000:
GCS structures
A resource structure created in the directory instance, a lock created in instance 2 X$bh X$le X$kjbl Shadow
[0x1ac4][0x7],[BL]
BH LE Shadow
[0x1ac4][0x7],[BL]
Buffer
X$kjbr
Block is in the buffer cache of instance 3. Instance 2 is the directory instance of the resource. LMS process transfers the blocks from instance 3 over the interconnect. 3 FG FG FG
LMD LMS
LMD LMS
LMD LMS
2
Buffer GRD
GRD
Buffer
GRD
Buffer
SGA
SGA
SGA
Disk files
FG Foreground Process LMD Lock Manager Daemon GRD Global Resource Directory
11
GRD After the transfer, GRD is updated with ownership changes. Both instances are the owner of the block.
KJBLNAME2 KJBLGRANT KJBLROLE KJBLREQUES KJUSERPR 0 KJUSERNL
KJBLNAME
If the block is transferred from one instance to another instance in PR mode, then the block mode is considered current mode transfer. Subsequently, gc current blocks received statistics incremented.
12
Buffer changes
Before modifying a buffer, BL lock on that buffer must be acquired in Exclusive mode (EX). Other instances will downgrade or flush that buffer from their instance, if that buffer is already in their cache. Instance acquired the block in EX mode and other instance(s) flushed the buffer.
KJBLNAME2 KJBLGRANT KJBLROLE KJBLREQUES KJUSEREX 0 KJUSERNL
KJBLNAME
Enter value for block: 6852 STATE MODE_HELD LE_ADDR DBARFIL DBABLK CR_SCN_BAS CR_SCN_WRP CLASS
---------- ---------- ---------------- ---------- ---------- ---------- ---------- ---------1 0 000000006D3E3AB0 7 6852 0 0 1
Busy
Event gc cr block busy, gc current block busy indicates that those blocks were busy. In this case, that block was in EX mode in another instance and LMS process applied undo blocks to reconstruct a consistent mode buffer reconstructing a CR mode buffer. Excessive *busy events would indicate application affinity is not in play. Application affinity will reduce *busy events as the buffers will be modified in the same instance.
14
But, if the instances crash right after the block is transferred to other node, how does RAC maintain consistency? Actually, before sending a current mode block LMS process will request LGWR for a log flush. Until LGWR sends a signal back to LMS process, LMS process will wait on gcs log flush event. CR block transfer might need log flush if the block was considered busy. One of the busy condition is that if the block was constructed by applying undo records.
15
CUR mode
What happens if two instances modify same block, but different rows? Row level locks prevent the same row being updated from two different instances. Before an instance can modify a block, the instance must acquire EX mode lock on the buffer. No two instances can hold the block in EX mode and a compatible buffer state.
CUR mode
What happens if there are two pending transactions from two different instances in the same block? No two instances are allowed to hold XCUR mode buffers with EX mode GCS lock concurrently.
FG LMS FG LMS
EX PI
EX
Buffer cache
Buffer cache
17
18
Types of packets
Block oriented packets Consistent Read blocks Read blocks Message oriented packets Single block grants
Multi Current
block grants
layer packets
19
CR Wait events
Following are the top wait events associated with CR mode transfers: gc cr block 2-way Transfers without congestion or concurrency. gc cr block 3-way gc cr multi block request gc cr block busy gc buffer busy (acquire/release) gc cr grant 2-way gc cr grant congested gc cr block congested
Multi block read Concurrency related Grants Congestion related
20
Gc cr block 2/3-way
Time is accounted for gc cr block 2-way if the block owner and master is an instance. If the owner and master instance are different than 3-way wait events are used. Time is accounted to these wait events if there was no need for additional work such as CR block creation or contention.
nam='gc cr block 2-way' ela= 627 p1=7 p2=6852 p3=1 obj#=76483 tim=37221074057
Analysis
These two events gc cr block 2-way and gc cr block 3-way can be considered as baseline events to calibrate cache fusion performance.
These events are in the top events consuming Considerable time. Differentiate between these two cases.
Histogram of waits indicates that elapsed time per event wait is high.
22
If the time_waited histogram indicated for this event is higher, it could be due to: High CPU usage in the nodes, leading to processes not getting CPU quick enough. Network performance or Network configuration issue.
Since concurrency or congestion related waits are not factored in to these waits, these are good baseline indicators for cache fusion performance.
23
Diagnostics
---------- -------------------- --------------- ---------- ---------1 gc cr block 2-way 1 gc cr block 2-way 1 gc cr block 2-way 1 gc cr block 2-way 1 gc cr block 2-way 1 gc cr block 2-way ... 1 gc cr block 2-way 1 gc cr block 2-way 8192 16384 9 24 0 0 1 2 4 8 16 32 3720856 148016413 140006974 40140870 2491886 43253 1.11 44.25 41.86 12 .74 .01
Demo: event_histogram.sql
24
Recommendations
Keep CPU usage below 80-85%. Above 80% CPU usage, scheduling inefficiency kicks in and multiplies the cache fusion performance issues. Possibly consider jumbo frames. Jumbo frames reduces assembly and disassembly of packets, so will reduce CPU usage slightly. Review network performance using OS tools. Review if cache fusion traffic is using private interconnect. Review if the cache fusion traffic is mixed with other network traffic.
25
If there are numerous waits for this wait event, identify the object and SQL causing these waits. SQL Trace or ASH data can be used to identify the object associated with these wait events. ASH data is a sampled data, so caution should be taken so that big enough samples are used. Object_id from the SQLTrace file can be used to identify the objects too.
26
Diagnostics
---------- -------------------- -------------------------------- -------------------- ---------... 1 APPLSYS 1 INV 1 INV 1 APPLSYS 1 INV 1 INV 1 FND_CONCURRENT_PROCESSES MTL_SERIAL_NUMBERS MTL_TRANSACTIONS_INTERFACE_N1 FND_CONCURRENT_REQUESTS MTL_MATERIAL_TRANSACTIONS MTL_TRANSACTIONS_INTERFACE TABLE TABLE INDEX TABLE TABLE TABLE Undo Header/Undo block? 118 144 176 184 211 216 18483
For undo header blocks/undo blocks, current_obj# is set to 0 and for undo blocks, curent_obj# is set to -1. Demo: ash_gcwait_to_obj.sql
OraInternals Riyaj Shamsudeen 27
Recommendations
Consider application affinity. Huge number of blocks transferred back and forth between the instances are indicating that application affinity might help. SGA size might be smaller for the workload. Try to see if increasing SGA size is an option. Stretch clusters will suffer from longer latencies due to network latency between the end points.
28
These wait events indicate that there were CPU resource starvation issues. For example, sudden spikes in PQ processing can increase CPU load average leading to CPU starvation. Reducing CPU usage by tuning costly SQL statement, scheduling jobs to run different times, or even adding new nodes is generally required. In a really busy and active environments, there will be few of these wait events; These events are concerns only if the AWR or SQLTrace indicates high amount of wait times for these events.
29
Gc cr grants 2-way
Time is accounted to this wait event, if the block is not in any of the buffer cache. Trace file will indicate this wait event followed by a disk read.
nam='gc cr grant 2-way' ela= 659 p1=1 p2=88252 p3=1 obj#=77779 nam='db file sequential read' ela= 938 file#=1 block#=88252 blocks=1 obj#=77779
Typical latency is 1-2ms. Any thing above needs to be reviewed as these are light-wait events. Process sends a request to remote master LMS process and the LMS process simply responds with read from disk. This is another base line wait event to measure interconnect response time, as LMS processing is limited.
OraInternals Riyaj Shamsudeen 30
RAC-Tuning objects
31
Partitioning
Partitioning can be used to improve performance and scalability in RAC instances. Few Guidelines:
Range Partitioning : For physical segment segregation Hash Partitioning Index : To improve insert concurrency Hash Partitioning Index : To reduce GC traffic for Select Hash Partitioning Table with local indexes: To improve insert concurrency Hash Partitioning Table : To reduce GC traffic for Select
From Version 10g onwards, partitioned indexes can be created on non-partitioned tables.
OraInternals Riyaj Shamsudeen 32
Btree indexes store ordered (key, rowid) pair. If the key column values are generated using a sequence value or monotonically increasing values, then those values are stored in the right most leaf block of the index. If many sessions are concurrently inserting into the index, all those sessions will be trying to insert in to right most leaf block of the index. This leads to contention in right most leaf block and known as right hand index growth contention.
33
Non-partitioned indexes
Ses 1 [1000] Ses 2 [1001] Ses 3 [1002] Ses 4[1003] Ses 5[1004] Ses 6 [1005]
34
In RAC
Right hand growth indexes will suffer from buffer busy wais in single instance. In RAC, this problem is magnified with enormous waits on gc buffer busy events and other downstream events.
35
Hash partitioning Hash partitioning is an option to resolve concurrency issues associated with right hand growth indexes. Hashing algorithm uniformly distributes values to various leaf blocks leading to increased concurrency. For example, by converting an unique non-partitioned index to a partitioned index with 2 partition, concurrency can be doubled. Conversion to 32 partition index will lead to a concurrency increase of near 32 fold. This action may be needed for even non-unique indexes if the data is almost unique, such as timestamp column.
OraInternals Riyaj Shamsudeen 36
37
Partition count
Keep partition count to be a binary power of 2 as hash partitioning algorithm uses hashing algorithm. Err on caution: Use bigger number of partitions such as 32, 64, or 128 partitions if the concurrency is higher. Of course, this might induce more logical reads, but the effect of that increase is negligible.
Another option to resolve right-hand-growth index contention is to convert the table to partitioned table and create the indexes as a local indexes. Since the algorithm uses hashing techniques, keep the partition count as binary power of 2. Hash partitioning indexes or table is a proven way to scale the application concurrency in RAC. Index Organized Tables also can suffer from this insert concurrency, if the row is short.
39
ASSM
ASSM avoids the need for manual freelist management. It is out of scope to go deeper in to ASSM, but in ASSM, L1 bitmaps are keeping the list of free blocks for insert. L2 bitmaps points to L1 bitmaps and L3 bitmaps in turn points to L2 bitmaps. L3 bitmaps is not common though. L1 bitmaps are searched to find free blocks.
Each instance assumes ownership of few L1 bitmaps. Processes in that instance search L1 bitmaps owned by that instance. Essentially, ASSM avoids the need for freelist groups, by instance-owning L1 bitmaps and second/third level indirect bitmaps. Dbms_space_admin package can be used to dump segment header information in ASSM tablespace.
c1.tablespace_name, c1.relative_fno, c1.header_block);
dbms_space_admin.segment_dump(
41
L1 bitmaps
When Instance 1 processes were inserting in to the segment, all L1 bitmaps were owned by that instance. When both instances were inserting in to the segment, new L1 bitmaps were owned by the second instance.
L1 Ranges : ------------------------------0x01c02470 Free: 1 Inst: 1 0x01000b08 Free: 1 Inst: 1 0x01000b28 Free: 1 Inst: 1 0x01000b48 Free: 1 Inst: 1 0x01000b60 Free: 1 Inst: 1 0x01000b78 Free: 1 Inst: 1 0x01c02588 Free: 1 Inst: 1 0x01c025b0 Free: 1 Inst: 1 0x01000b80 Free: 1 Inst: 1 0x01000b81 Free: 5 Inst: 1
L1 Ranges : --------------------------0x01c02470 Free: 1 Inst: 1 0x01000b08 Free: 1 Inst: 1 0x01000b28 Free: 1 Inst: 1 0x01000b48 Free: 1 Inst: 1 0x01000b60 Free: 1 Inst: 1 0x01000b78 Free: 1 Inst: 1 0x01c02588 Free: 1 Inst: 1 0x01c025b0 Free: 1 Inst: 1 0x01000b80 Free: 1 Inst: 1 0x01000b81 Free: 1 Inst: 2 0x01c02600 Free: 1 Inst: 2 0x01c02601 Free: 1 Inst: 2 0x01000c80 Free: 5 Inst: 2 0x01000c81 Free: 5 Inst: 2
42
Instance ownership
Dump of First Level Bitmap Block -------------------------------nbits : 4 nranges: 1 parent dba: 0x01c02471 poffset: 9 unformatted: 0 total: 64 first useful block: 0 owning instance : 2 instance ownership changed at 01/28/2011 22:43:06 Last successful Search 01/28/2011 22:43:06 Freeness Status: nf1 0 nf2 0 nf3 0 nf4 0
Excessive deletes doesnt lead to ill effects similar to freelist blocks. Free blocks are correctly accounted to L1 bitmaps and instance ownership maintained. In a nutshell, consider using ASSM in RAC environment.
43
Sequences
Incorrect configuration of sequences can be fatal to performance. For cached sequences, each instance caches a range of sequence values. Problem is that sessions running from different nodes can get non-sequential values. For example, if the cache is 20, then session #1 in instance 1 will get a value of 1 and session #2 in instance 2 will retrieve a value starting at 21. In normal operations, there is no loss of values, just possible gaps.
OraInternals Riyaj Shamsudeen 44
3 Second access caches 1 First access to sequence caches values from 10 to 29 value from 30-49 1. 60 access to sequence results in 3 changes to block. 10-29 30-49 2. These changes might not result in physical reads/writes. Inst 1 2 SEQ$ updated with last_value as 29 7 SEQ$ updated with last_value as 69 Inst 2 3. Gaps in sequence values. 4. Still, log flush needed for cache transfer. 4 SEQ$ updated with last_value as 49
45
Sequences
If ordered values are needed, consider order, cache attributes. Order, cache attribute provides better performance since the GES layer is used to maintain the order between the instances. Still, with order, cache it is possible to lose the values in case of instance crashes. You should consider order, nocache only for lightly used sequences such as control table sequences etc.
47
call
count
cpu
elapsed
disk
query
current
-------- ---------- ---------- ---------- ---------0.00 0.94 0.00 0.01 12.60 0.00 0 0 0 0 910 0 0 16440 0
Event waited on ---------------------------------------DFS lock handle enq: HW - contention buffer busy waits
-------- ---------- ---------- ---------- ---------0.00 7.71 0.00 0.00 282.75 0.00 0 3 0 0 333 0 0 20670 0
Event waited on ---------------------------------------row cache lock Disk file operations I/O db file sequential read gc current block busy gc current block 2-way
Break it down..
Coutesy:npowersoftware.com
50
Send GC Message OS,Network stack User session processing Copy to SGA / PGA OS,Network stack Rx Msg CR / CUR block build Msg to LGWR (if needed) Wake up Send Block OS,Network stack Wakeup Log buffer processing Log file write Signal LMS
User
LMSx
LGWR
Node 1
Node 2
51
GC CR latency
GC CR latency ~= Time spent in sending message to LMS + LMS processing (building blocks etc) + LGWR latency ( if any) + LMS send time + Wire latency
Averages can be misleading. Always review both total time and average to understand the issue.
52
Breakdown latency
Wait time gc cr block build time Gc cr block flush time Gc cr block send time
Node 1 Node 2 Node 3 Node 4 Total 402 3016 375 199 870 188 100 978 87 227 2247 265 1679 7111 1290
53
GC CURRENT latency
GC CUR latency ~= Time spent in sending message to LMS + LMS processing : (Pin and build block) + LGWR latency: Log flush + Wire latency Statistics : gc current block flush time gc current block pin time gc current block send time
54
GV$ views
GV$ and V$ views are implemented with an abstraction layer. GV$ views: Fixed views, accessing x$ tables. For example, gv$database is accessing x$kccdi and x$kccdi2 fixed tables.
select di.inst_id,di.didbi,di.didbn, to_date(di.dicts,'MM/DD/RR HH24:MI:SS', 'NLS_CALENDAR=Gregorian'), to_number(di.dirls) , to_date(di.dirlc,'MM/DD/RR HH24:MI:SS','NLS_CALENDAR=Gregor
... fl2,64), 64, 'YES', 'NO'), decode(di2.di2min_req_capture_scn,0, to_number(null), di2.di2min_req_capture_scn) from x$kccdi di, x$kccdi2 di2
55
V$ views
V$ views access gv$ views and filter data specific to current instance. This is true even in a single instance.
Create or replace v$database as select DBID, NAME, CREATED, RESETLOGS_CHANGE#, RESETLOGS_TIME, PRIOR_RESETLOGS_CHANGE#, PRIOR_RESETLOGS_TIME,LOG_MODE,CHECKPOINT_CHANGE#, ARCHIVE_CHANGE#, CONTROLFILE_TYPE, CONTROLFILE_CREATED, CONTROLFILE_SEQUENCE#, CONTROLFILE_CHANGE#, ... from GV$DATABASE where inst_id = USERENV('Instance') ;
Demo: demo_gvdef.sql
56
GV$ implementation
GV$ views retrieves from all instances and merges them to produce final output. Specialized parallel query slaves are used to retrieve rows from different instances (10g and above).
INST_ID QC/Slave Slave Set SID QC SID Requested DOP Actual DOP 1 QC 2 (Slave) 1 (Slave) 1 1 52 55 62 52 52 52 2 2 2 2
Username
------------ ------- ---------- ---------- ------ ------ ------------- ---------SYS - pz99 - pz99
In 9i, normal PQ slaves were used to retrieve rows from remote instances.
OraInternals Riyaj Shamsudeen 57
Caution
Dont use gv$views to find averages. Use AWR reports or custom scripts. gv$views are aggregated data and persistent from the instance restart. For example this query can be misleading:
select b1.inst_id, b2.value "RECEIVED", b1.value "RECEIVE TIME", ((b1.value / b2.value) * 10) "AVG RECEIVE TIME (ms)" from gv$sysstat b1, gv$sysstat b2 where b1.name = gc cr block receive time' and b2.name = 'gc cr blocks received' and b1.inst_id = b2.inst_id
58
gc_traffic_print.sql
You can use my script to print global cache performance data for the past minute. Download from scripts archive:
http://www.orainternals.com/scripts_rac1.php
---------|--------------|---------|----------------|----------|---------------|---------------|-------------| Inst | CR blocks Rx | CR time | CUR blocks Rx | CUR time | CR blocks Tx | CUR blocks Tx |Tot blocks | ---------|--------------|---------|----------------|----------|---------------|---------------|-------------| 1 | 40999| 13.82| 7827| 4.82| 25070| 17855| 91751| 2 | 12471| 5.85| 8389| 5.28| 31269| 9772| 61901| 3 | 28795| 4.11| 18065| 3.97| 28946| 4248| 80054| 4 | 33105| 4.54| 12136| 4.68| 29517| 13645| 88403| ---------|--------------|---------|----------------|----------|---------------|---------------|-------------|
During the same time frame, output of the script from prior slide:
Very misleading!
INST_ID RECEIVED RECEIVE TIME AVG RECEIVE TIME (ms) ---------- ---------- ------------ --------------------4 165602481 104243160 6.2947825 2 123971820 82993393 6.69453695 3 215681074 103170166 4.7834594 1 134814176 66663093 4.9448133
59
It is important to review performance data from all the nodes. It is easy to create AWR reports from all nodes using my script: Refer awrrpt_all_gen.sql. [ Dont forget that access to AWR report needs license ] Or use my script gc_traffic_processing.sql from my script archive.
Default collection period is 60 seconds.... Please wait for at least 60 seconds... ---------|-----------|---------|-----------|----------|------------|------------|------------|----------| Inst | CR blk Tx | CR bld | CR fls tm | CR snd tm| CUR blk TX | CUR pin tm | CUR fls tm |CUR blk TX| ---------|-----------|---------|-----------|----------|------------|------------|------------|----------| 2 | 67061| .08| .88| .23| 34909| 1.62| .2| .23| 3 | 38207| .17| 2.19| .26| 28303| .61| .08| .26| 4 | 72820| .06| 1.76| .2| 40578| 1.76| .24| .19| 5 | 84355| .09| 2.42| .23| 30717| 2.69| .44| .25| --------------------------------------------------------------------------------------------------------
60
Sessions can be seen waiting for these wait events, but will not show up in AWR / ADDM reports. After sending the global cache block request, foreground process waits on these events. On receipt of the response, time is accounted for correct wait event.
OraInternals Riyaj Shamsudeen 61
Histogram
Averages can be misleading. Use v$event_histogram to understand true performance metrics. It is better to take snapshots of this data and compare the differences.
INST_ID EVENT WAIT_TIME_MILLI WAIT_COUNT THIS_PER TOTAL_PER ---------- ------------------------- --------------- ---------- ---------- ---------1 gc cr block 2-way 1 466345 .92 .92 1 gc cr block 2-way 2 23863264 47.58 48.51 1 gc cr block 2-way 4 20543430 40.96 89.47 1 gc cr block 2-way 8 4921880 9.81 99.29 1 gc cr block 2-way 16 329769 .65 99.95 1 gc cr block 2-way 32 17267 .03 99.98 1 gc cr block 2-way 64 2876 0 99.99 1 gc cr block 2-way 128 1914 0 99.99 1 gc cr block 2-way 256 1483 0 99.99 1 gc cr block 2-way 512 618 0 99.99 1 gc cr block 2-way 1024 83 0 99.99 1 gc cr block 2-way 2048 4 0 99.99 1 gc cr block 2-way 4096 3 0 99.99 1 gc cr block 2-way 8192 5 0 99.99 1 gc cr block 2-way 16384 3 0 100
62
GC event histograms
Default collection period is sleep seconds. Please wait.. Enter value for event: gc cr block 2-way Enter value for sleep: 60 ---------|-----------------------|----------------|----------| Inst id | Event |wait time milli |wait cnt | ---------|-----------------------|----------------|----------| 1 |gc cr block 2-way | 1| 37| 1 |gc cr block 2-way | 2| 4277| 1 |gc cr block 2-way | 4| 5074| 1 |gc cr block 2-way | 8| 1410| 1 |gc cr block 2-way | 16| 89| 1 |gc cr block 2-way | 32| 1| 1 |gc cr block 2-way | 64| 0| 1 |gc cr block 2-way | 128| 0| 1 |gc cr block 2-way | 256| 0|
63
GC buffer busy waits are usually symptoms. In many instances, this event can show up the top most waited event. GC Buffer busy simply means that buffer is pinned by another process and waiting for a different global cache event. Understand why that buffer pin holder is waiting. Resolving that will resolve global cache buffer busy waits. Segment header changes dues to insufficient freelist groups also can lead to longer gc buffer busy waits.
OraInternals Riyaj Shamsudeen 64
Example analysis Client had high Global Cache response time waits.
Global Cache and Enqueue Services - Workload Characteristics ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Avg global enqueue get time (ms): 2.5 Avg global cache cr block receive time (ms): Avg global cache current block receive time (ms): Avg global cache cr block build time (ms): Avg global cache cr block send time (ms): Global cache log flushes for cr blocks served %: Avg global cache cr block flush time (ms): Avg global cache current block pin time (ms): Avg global cache current block send time (ms): Global cache log flushes for current blocks served %: Avg global cache current block flush time (ms): 18.2 14.6 0.3 0.2 25.1 5.2 0.4 0.2 1.7 5.2
65
CR latency
Wait time Avg. CR block receive time Avg CUR block receive time
In RAC, node suffering from chronic issues causes GC performance issues in other nodes. With that logic in mind, node 2 should be suffering from chronic issues.
OraInternals Riyaj Shamsudeen 66
Breakdown of latency
Sum of flush time is higher, but it is comparable across the cluster. But, notice the build time in node 2.
Node 1 Node 2 Node 3 Node 4 Total 11,392 148,666 56,634 9,153 75,751 7,779 5,267 34,406 4,018 6,632 53,031 7,905 171,957 219,822 28,855
Statistics gc cr block build time Gc cr block flush time Gc cr block send time
67
Consistent reads
For CR blocks, time is spent in building blocks, which indicates consistent block generation. Very high value compared to other nodes.
Statistics data blocks consistent Reads undo records applied db block changes
Node 1 2,493,242
Node 2 86,988,512
Node 3 3,090,308
Node 4 7,208,575
6,276,149
43,898,418 20,698,189
14,259,340
68
Time line
We wanted to see when this problem started. Surprisingly, instance 2 had a pattern of increasing flush time.
69
Db block changes
with
Unfortunately, AWR report does not capture segments with high db block changes.
segstats as ( select * from ( select inst_id, owner, object_name, object_type , value , rank() over (partition by inst_id, statistic_name order by value desc ) rnk , statistic_name from gv$segment_statistics where value >0 ) where rnk <11 ) , sumstats as ( select inst_id, statistic_name, sum(value) sum_value from gv $segment_statistics group by statistic_name, inst_id) select a.inst_id, a.statistic_name, a.owner, a.object_name, a.object_type,a.value, (a.value/b.sum_value)*100 perc from segstats a , sumstats b where a.statistic_name = b.statistic_name and a.inst_id=b.inst_id and a.statistic_name ='db block changes' order by a.statistic_name, a.value desc /
INST_ID STATISTIC_NAME ------- -----------------2 db block changes 4 3 3 1 ... OWNER ----AR INV AR AR INV OBJECT_NAME -----------------------------CUSTOM_TABLE MTL_MATERIAL_TRANS_TEMP_N1 RA_INTERFACE_LINES_N2 RA_CUSTOMER_TRX_LINES_N2 MTL_MATERIAL_TRANS_TEMP_N1 TYPE VALUE PERC ----- ------------ -----TABLE 122949282400 81.39 INDEX 1348827648 16.59 INDEX 791733296 9.77 INDEX 715855840 8.83 INDEX 652495808 12.44
70
Solution
Finally, it boiled down to a custom code bug which was updating almost all rows in a table unnecessarily. Unfortunately, number of rows that fall in to that criteria was slowly increasing. So, GC CR response time was slowly creeping up and it wasnt easy to identify the root cause. After the code fix, GC CR time came down to normal range.
71
Agenda
Global cache performance Undo, redo and more RAC background process tuning Interconnect issues, lost packets and network layer Network layer tuning Effective use of parallel query Troubleshooting locking issues Object re-mastering
72
Question
73
CR and undo
1 User process in instance 1 requests master for the block in PR mode. 2 Current owner (2) holds the block in Exclusive mode. 3 Instance 2 applies undo to create a version of the block consistent with SCN requested. Then ships the block to instance 1. 1 2 Inst 1 3
undo
Inst 2
Inst 3
74
But, if one instance is a read only instance, then it might request another instance to generate CR copies applying undo blocks excessively. This is avoided by light-works rule. If an instance is excessively serving a block by applying undo blocks, it will downgrade the block mode. Requesting instance then will read from the disk and apply undo blocks (if needed) reducing load on the prior owner.
75
Undo for CR
A query can not read the block image with an SCN later than the query environment SCN. If a block is ahead of time, then undo blocks are applied to create a consistent version of the block. But, if the block was modified by a different instance, then undo blocks may need to be shipped from another instance. This can cause excessive cache transfers or excessive physical reads for undo blocks. Node affinity will be helpful to resolve this.
76
Commit cleanout
ITL entries in the blocks may not be cleaned out immediately. Session reading the block next time will check if the pending transaction is committed or not. If the transaction table is cycled through, then the session will apply undo to find a transaction table version with that transaction. This can lead to vicious cycle of rolling back the transaction table accessing undo blocks excessively. In RAC, this problem is magnified, since the transaction(s) could be in a different instance. So, transaction table blocks and undo blocks from different instance need to be shipped from other node or read from the disk.
OraInternals Riyaj Shamsudeen 77
LGWR performance is important for global cache response time. Even for CR blocks LGWR must flush if the block is considered busy.
Statistics Gc cr block flush time Node 1 Node 2 Node 3 Node 4 129,970 12,289 11,556 27143
78
Agenda
Global cache performance Undo, redo and more RAC background process tuning Interconnect issues, lost packets and network layer Network layer tuning Effective use of parallel query Troubleshooting locking issues Object re-mastering
79
Global Cache waits increases due to increase in LMS latency in the CPU starved node. Much of these GC waits are blamed on interconnect interface and hardware. In many cases, interconnect is performing fine, it is that GCS server processes are introducing latencies.
80
In 9i, increasing priority of LMS processes to RT helps (more covered later). From Oracle release 10.2.0.3 LMS processes run in Real Time priority by default. Two parameters control this behaviour: _high_priority_processes _os_sched_high_priority
81
Parameters in 10gR2
_high_priority_processes: Default value: LMS*|VKTM* This parameter controls what background processes should get Real time priority. Default is all LMS processes and VKTM process.
_os_sched_high_priority : Default value: 1 This is a switch. If set to 0, no background process will run in high priority.
OraInternals Riyaj Shamsudeen 82
oradism
Of course, bumping priority needs higher privileges such as root in UNIX. Oradism utility is used to increase the priority class of these critical background process in UNIX. Verify that LMS processes are using Real time priority in UNIX and if not, oradism might not have been configured properly. In Windows, oradism service is used to increase the priority.
83
Typical response is to increase number of LMS processes adjusting _lm_lms (9i) or gcs_server_processes(10g). Increase in LMS processes without enough need increases xcalls/ migrates/tlb-misses in massive servers. Further, LMS process runs in RT CPU priority and so, CPU usage will increase.
84
In huge servers, by default, number of LMS processes might be quite high. It is possible to get up to 26 LMS processes by default. Typically, same number of LMS processes as interconnect or remote nodes is a good starting point. If there is enormous amount of interconnect traffic, then configure LMS processes to be twice the interconnect.
85
LGWR and CPU priority LGWR performance is akin to Global cache performance. If LGWR suffers from performance issues, it will reflect on Global cache performance. For example, If LGWR suffers from CPU latency issues, then LMS will have longer waits for gcs log flush sync event This leads to poor GC performance in other nodes.
86
LGWR priority Method to increase priority for LGWR and LMS in 9i (Example for Solaris). If you dont want to increase priority to RT for LGWR, at least, consider FX priority.
priocntl -e -c class -m userlimit -p priority priocntl -e -c RT -p 59 `pgrep -f ora_lgwr_${ORACLE_SID}` priocntl -e -c FX -m 60 -p 60 `pgrep -f ora_lms[0-9]*_${ORACLE_SID}`
alter system set "_high_priority_processes"="LMS*|LGWR*" scope=spfile sid='*'; alter system set "_high_priority_processes"="LMS*|VKTM*|LGWR*" scope=spfile sid='*'; (11g)
Pitfalls of RT mode
Of course, there are few! RT is kernel preemptive. LMS process can continuously consume CPU and can introduce CPU starvation in servers with few CPUs. A bug was opened to make LMS process sleep intermittently, but that causes LMS to be less active and can cause GC latency. Another undocumented parameter _high_priority_process_num_yields_before_sleep was introduced as a tunable. But, hardly a need to alter this parameter. So, LMS might wait for LGWR. If LGWR is not running in RT, then LMS can preempt LGWR leading to not-so-optimal wait graph. But, LGWR can block interrupt which LGWR might need!
OraInternals Riyaj Shamsudeen 88
Binding..
Another option is to bind LGWR/LMS to specific processors or processor sets. Still, interrupts can pre-empt LMS processors and LGWR. So, binding LMS to processor set without interrupts helps (see psradm in solaris).
But, of course, processor binding is useful in servers with higher number of CPUs such as E25K / M9000 platforms.
89
CSSD/CRSD
CSSD is a critical process. Few CSSD processes must be running with RT priority. crsctl set css priority 4 CPU starvation in the server can lead to missed network or disk heart beat. This can lead to node reboots. It is important to have good and consistent I/O performance to ORA_CRS_HOME directories. If CSSD cant access those directories efficiently (i.e. due to NFS or other file system issues), then that can lead to node reboots too.
Summary
In summary, Use optimal # of LMS processes Use RT or FX high priority for LMS and LGWR processes. Configure decent hardware for online redo log files. Tune LGWR writes and Of course, avoid double buffering and double copy using optimal file systems. Of course, tune SQL statement to reduce logical reads and reduce redo size.
OraInternals Riyaj Shamsudeen 91
Agenda
Global cache performance Few important RAC wait events and statistics RAC background process tuning Interconnect issues, lost packets and network layer Effective use of parallel query Troubleshooting locking issues Object re-mastering
92
gc blocks lost Probably, the most direct statistics indicating interconnect issues. Consistent high amount of gc blocks lost is an indication of problem with underlying network infrastructure. (Hardware, firmware,setup etc). Need to understand which specific component is an issue. Usually, this is an inter-disciplinary analysis. Ideal value is near zero. But, only worry about this, if there are consistently higher values.
93
Higher number of block loss can lead to timeouts in GC traffic wait events. Many processes will be waiting for place-holder events. Use total_timeouts column in v$system_event to see if the timeouts are increasing. Percent of total_timeouts should be very small.
94
Network layers
LMSx User Process Socket layer Udp_xmit_hiwat Udp_recv_hiwat Udp_max_buf Net.core.rmem_max Fragmentation and Assembly Socket layer Socket queues
Interface layer
UDP is a send-and-forget type protocol. Sending process does not get any acknowledgement. UDP Tx/Rx buffers are allocated per process. When the process executes CPU, it drains the UDP buffers. If the buffer is full, then incoming packets to that process are dropped. Default values for the UDP buffers are small for the bursty nature of interconnect traffic. Increase UDP buffer space to 128KB or 256KB.
OraInternals Riyaj Shamsudeen 96
Due to CPU latency, process might not be able to acquire CPU quick enough. This can lead to buffer full conditions and lost packets.
It is essential to keep CPU usage under 80% to avoid latencies and lost packets.
97
Agenda
Global cache performance Few important RAC wait events and statistics RAC background process tuning Interconnect issues, lost packets and network layer Effective use of parallel query Troubleshooting locking issues Object re-mastering
98
Parallel Query slaves can be allocated from multiple instances for a query. It is imperative that PQ messages are transmitted between producers and consumers. Insufficient network bandwidth with PQ storm can cause higher GC latency and possible packet loss.
99
PQ Optimization
Inst 1 QC P9 P10 P11 P16
Communication between producers/consumers are Not limited to one node. Gigabytes of data flew Between node 1 and node 2.
Inst 2
Consumers
P1 P2 P3
P8
P1 P2 P3
P8
Producers
100
Optimizations in 10g/11g
PQ algorithms are optimized in Oracle versions 10g and 11g. Only few discussed here. In 11g, interconnect traffic due to PQ is also reported in the AWR reports. Oracle code tries to allocate all PQ slaves in one node, if possible. This minimizes PQ induced interconnect traffic. If it not possible to allocate all slaves from a node, then the least loaded node(s) are chosen for PQ slave allocation.
101
Partition-wise joins
Interconnect traffic is kept minimal as the equivalent partitions are joined by a PQ process and final result is derived by Query Co-ordinator.
QC
P1
P2
P3
P4
P5
P6
P1 P1
P2 P2
P3 P3 Inst 2
P4 P4
P5 P5
P6 P6 Inst 3
Table T1 Table T2
102
PQ-Summary
Inter instance parallelism need to be carefully considered and measured. For partition based processing, when processing for a set of partitions is contained within a node, performance will be better. Excessive inter instance parallelism will increase interconnect traffic leading to performance issues.
http://www.oracle.com/technology/products/bi/db/11g/pdf/ twp_bidw_parallel_execution_11gr1.pdf
Agenda
Global cache performance Few important RAC wait events and statistics RAC background process tuning Interconnect issues, lost packets and network layer Effective use of parallel query Troubleshooting locking issues Object re-mastering
104
Globalization
GES layer locks are externalized through x$kjirft and x$kjilkft tables.
Owners Waiters granted_q
Lock
Lock
Lock
Converting_q
x$ksqeq Lock
Lock
x$kjilkft Lock
Lock
x$ksqrs
x$kjirft
105
---------- ---------- -- ---------- ---------- ---------- ----- ------- ---------1 1 48 TM 48 TX 76483 3866640 0 49 55 55 3 6 0 0 2 2
106
Updating the same row in another instance, globalised the TX resource and created lock structures in the master instance.
------------------Resource details... ------------------Resource name [0x470001][0x159],[TX][ext 0x2, Master 0,Instance 1 Resource name [0x470001][0x159],[TX][ext 0x2, Master 0,Instance 2 ------------------Lock details... ------------------Res name [0x470001][0x159],[TX][ext 0x2, owner 0 ...Transaction_id0 2097153,Level KJUSEREX ,State GRANTED ...blocked 0,1 Res name [0x470001][0x159],[TX][ext 0x2, owner 1 ...Transaction_id0 0,Level KJUSERNL ,State GRANTED ...blocked 0,0 Res name [0x470001][0x159],[TX][ext 0x2, owner 1 ...Transaction_id0 2228226,Level KJUSERNL ,State OPENING , Req. lvl KJUSEREX
For TX resource, instance creating the transaction is the master instance, starts with 0.
Waiting session
Agenda
Global cache performance Few important RAC wait events and statistics RAC background process tuning Interconnect issues, lost packets and network layer Effective use of parallel query Troubleshooting locking issues Object re-mastering
108
Object re-mastering
Before reading the block, an user process must request master node of the block to access that block. Typically, a batch process will access few objects aggressively. If an object is accessed excessively from a node then remastering the object to that node reduces Global cache grants. Local grants (affinity locks) are very efficient compared to remote grants avoiding global cache messaging traffic.
OraInternals Riyaj Shamsudeen 109
Dynamic remastering is file based in 10gR1. If a block need to be remastered, then every block in that data file must be remastered to an instance. In 10gR2, remastering is object based. If a block to be remastered, then all blocks associated with that object is remastered to an instance. Three background processes work together to implement dynamic remastering functionality.
110
LCK0 process maintains object level statistics and determines if remastering must be triggered. If an object is chosen, a request is queued. LMD0 reads the request queue and initiates GES freeze. LMD0 trace file
*** 2010-01-08 19:41:26.726 * kjdrchkdrm: found an RM request in the request queue Dissolve pkey 6984390 *** 2010-01-08 19:41:26.727 Begin DRM(189) - dissolve pkey 6984390 from 2 oscan 1.1 ftd received from node 1 (8/0.30.0) ftd received from node 0 (8/0.30.0) ftd received from node 3 (8/0.30.0) all ftds received
*** 2010-01-08 19:41:26.793 Begin DRM(189) sent syncr inc 8 lvl 5577 to 0 (8,0/31/0) synca inc 8 lvl 5577 rcvd (8.0)
111
_gc_affinity_limit default value is 50. Not documented well, but, it is number of times a node should open an object more than other nodes. _gc_affinity_time default value is 10. Frequency in seconds to check if remastering to be triggered or not. _gc_affinity_minimum determines minimum activity per minute to trigger remastering default to 2400.
OraInternals Riyaj Shamsudeen 112
Defaults
Default for these parameters may be too low in a very busy, highend instances. If your database have higher waits for gc remaster and gcs drm server freeze then dont disable this feature completely. Instead tune it. Some good starting points (for a very busy environment) are: [ YMMV]
_gc_affinity_limit
to 250 to 2500.
_gc_affinity_minimum
113
11g
In 11g, these are few changes. Three new parameters are introduced: _gc_affinity_locking _gc_affinity_locks
_gc_affinity_ratio
114
An example
Top 5 Timed Events Avg %Total ~~~~~~~~~~~~~~~~~~ wait Call Event Waits Time (s) (ms) Time Wait Class ------------------------------ ------------ ----------- ------ ------ ---------gc buffer busy 1,826,073 152,415 83 62.0 Cluster CPU time 30,192 12.3 enq: TX - index contention 34,332 15,535 453 6.3 Concurrenc gcs drm freeze in enter server enq: TX - row lock contention 22,789 46,926 11,279 4,493 495 96 4.6 Other
1.8 Applicatio
Global Cache and Enqueue Services - Workload Characteristics ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Avg global enqueue get time (ms): 16.8 Avg global cache cr block receive time (ms): Avg global cache current block receive time (ms): 17.1 14.9
115
Views
View v$gcspfmaster_info provides remastering details. For example, you can identify the object with high remastering count.
FILE_ID OBJECT_ID CURRENT_MASTER PREVIOUS_MASTER REMASTER_CNT ---------- ---------- -------------- --------------- -----------0 6983606 0 32767 1 0 5384799 2 1 2 0 6561032 3 2 2 0 5734002 0 2 2 0 6944892 2 0 2 0 5734007 2 0 4 0 6944891 2 0 5 0 6795604 2 0 5 0 6944894 2 0 5 0 6795648 2 0 6 0 5734006 2 0 6 0 4023250 2 0 6 0 5734003 0 2 7
116
Views
select * from x$object_affinity_statistics order by opens ADDR INDX INST_ID OBJECT NODE OPENS ---------------- ---------- ---------- ---------- ---------- --------- FFFFFFFF7C04CB40 8 3 4740170 1 113 FFFFFFFF7C04CB40 109 3 1297745 1 127 FFFFFFFF7C04CB40 21 3 1341531 1 128 FFFFFFFF7C04CB40 2 3 2177393 1 135 FFFFFFFF7C04CB40 153 3 6942171 2 174 FFFFFFFF7C04CB40 108 3 1297724 1 237 FFFFFFFF7C04CB40 3 3 2177593 1 239 FFFFFFFF7C04CB40 106 3 1297685 1 337 FFFFFFFF7C04CB40 53 3 6984154 3 1162
117
This enqueues an object remaster request. LMD0 and LMON completes this request *** 2010-01-08 23:25:54.948 * received DRM start msg from 1 (cnt 1, last 1, rmno 191) Rcvd DRM(191) Transfer pkey 6984154 from 0 to 1 oscan 0.0 ftd received from node 1 (8/0.30.0) ftd received from node 0 (8/0.30.0) ftd received from node 3 (8/0.30.0)
118
Oradebug You can manually remaster an object with oradebug command. Current_master starts from 0.
1* select * from v$gcspfmaster_info where object_id=6984154 SQL> / FILE_ID OBJECT_ID CURRENT_MASTER PREVIOUS_MASTER REMASTER_CNT ---------- ---------- -------------- --------------- -----------0 6984154 1 0 2 SQL> oradebug lkdebug -m pkey 6984154 Statement processed. SQL> select * from v$gcspfmaster_info where object_id=6984154 2 / FILE_ID OBJECT_ID CURRENT_MASTER PREVIOUS_MASTER REMASTER_CNT ---------- ---------- -------------- --------------- -----------0 6984154 2 1 3
119
If you like this presentation, you will love our upcoming intensive RAC webinar in August 11. Watch for updates in: www.tanelpoder.com orainternals.wordpress.com Contact info: Email: [email protected] Blog : orainternals.wordpress.com URL : www.orainternals.com
120
References
1. Oracle support site. Metalink.oracle.com. Various documents 2. Internalsguru Steve Adams website www.ixora.com.au 3. Jonathan Lewis website www.jlcomp.daemon.co.uk 4. Julian Dykes website www.julian-dyke.com 5. Oracle8i Internal Services for Waits, Latches, Locks, and Memory by Steve Adams 6. Randolf Geist : http://oracle-randolf.blogspot.com 7. Tom Kytes website Asktom.oracle.com 8. Richard Stevens, Gary R Wright: TCP/IP Illustrated
121