Transcript 100% of IO operations take less than Z

October 1st 2014, Oracle OpenWorld 2014
Eric Grancher, head of database services group, CERN-IT
Video screen capture at
https://indico.cern.ch/event/344531/
CERN DB blog: http://cern.ch/db-blog/
Storage, why does it matter?
•
•
•
•
•
Database Concepts 12c Release 1 (12.1) "An
essential task of a relational database is data
storage."
Performance (even if DB would be in memory,
commit are synchronous operations)
Availability, many (most?) recovery due to failing IO
subsystems
Most of the architecture can fail, storage not!
Fantastic evolution in recent years, more to come!
3
Outlook
•
•
•
•
•
Oracle IOs
Technologies
Planning
Lessons from experience / failure
Capturing information
4
Understand DB IO patterns (1/2)
•
You do not know what you cannot measure
(24x7!)
• Oracle DB does different types of IO
• Many OS tools enable to understand them
• Oracle DB provides timing and statistics
information about the IO operations (as seen
from the DB)
5
Understand DB IO patterns (2/2)
•
•
•
snapper (Tanel Poder)
strace (Linux) / truss (Solaris)
perf / gdb (see Frits Hoogland’s blog)
6
7
8
Overload at CPU level (1/2)
•
•
•
Observed many times: “the storage is slow”
(and storage administrators/specialists say
“storage is fine / not loaded”)
Typically happens that observed (from
Oracle rdbms point of view) IO wait times
are long if CPU load is high
Instrumentation / on-off CPU
Overload at CPU level (2/2)
t1
t2
t1
t2
t1
t2
Oracle
OS
Acceptable
load
IO
time
t1
t2
t1
t2
Oracle
OS
Off cpu
IO
High load
Technology
•
•
•
•
Rotating disk
Flash based devices
Local and shared storage
Functionality
11
Rotating disk (1/2)
•
Highest capacity / cost
credit: (WLCG) computing Model update
credit: Wikipedia
12
Rotating disk (2/2)
•
Well known technology… but complex
credit: Computer Architecture: A Quantitative Approach
13
Flash memory / NAND (1/2)
•
•
Block erasure (128kB)
Memory wear
credit: Intel (2x)
credit: Wikipedia
14
Flash memory / NAND (2/2)
•
•
•
•
Multi-level (MLC) and single-level (SLC), even TLC
Low latency 4kB: read in 0.02ms, write in 0.2ms
Complex algorithms to improve write performance, use
over-provisioning
Enormous
differences
15
Local versus shared (1/2)
•
•
•
Locally attached storage
Remotely attached storage
(NFS, FC)
Real Application Cluster
credit: Broadberry
credit: Oracle
16
Local versus shared (2/2)
•
Serial ATA (SATA):
•
•
•
•
PCI Express: 1.5 GB/s PCI-Express 2.0 x4
Serial Attached SCSI (SAS)
•
•
•
•
Current: SATA 3.0, 6 Gb/s, half duplex, 550 MB/s
SATA express: 2 PCIe 3.0 lanes -> 1969 MB/s
6/12 Gb/s, multiple initiators, full duplex
NFS: 10GbE
FC: 16 Gb/s: 1.6GB/s (+ duplex)
Infiniband: QDR 4x 40Gb/s
17
Bandwidth (1/2)
•
How many spinning disks to fill a 16GFC
channel?
•
•
•
Each spinning disk (15k RPM) at 170 MB/s.
1.6GB/s / 170 MB/s ~= 9.4
Each flash disk SATA 3.0 at 550 MB/s. 1.6GB/s /
550 MB/s ~= 2.9
Link aggregation!
references
credit: Tomshardware, wikipedia
18
Bandwidth (2/2)
•
Only few disk saturate storage networking, so
•
•
•
either many independent sources and ports on the server
or (FC) multipathing
or Direct NFS link aggregation
•
PCI Express (local)
•
Balanced systems in term of storage networking
require good planning
19
Queuing / higher load
20
Writeback / memory caching
•
•
Major gain (x ms to <1ms), but it requires solid execution
9 August 2010
•
•
•
•
18 August 2010
•
•
•
•
Power stop
No Battery Backup Unit and write-back enabled
Database corruption, 2 minutes data loss
Disk failure
Double controller failures (write-back)
Database corruption
1 September 2010
•
•
•
Planned intervention, clean database stop
Power stop (write-back data has not been flushed)
Database corruption including backup on disk
21
Functionality
•
Thin provisioning (ex: db1 5TB in a 15TB
volume)
• Snapshot at the storage level, restore at the
storage level (ex: 10TB in 15s)
• Cloning (*) at the storage level
Integration with Oracle Multitenant
22
ASM, (cluster/network) file system
•
•
•
•
[754305.1] 11.2 DBCA does not support raw
devices anymore
[12.1 doc] Raw devices have been desupported and deprecated in 12.1
Local filesystem, cluster and NFS for RAC
ASM
24
Filesystem
•
•
•
•
•
•
FILESYSTEMIO_OPTIONS=SETALL enables
direct IO and asynchronous IO
NFS with Direct NFS
Cluster filesystem (example ACFS)
Advanced features like snapshots
Important to do regular de-fragmentation if
there is some sort of copy-on-write
Simplicity
25
ASM
•
•
•
•
ASMLib /dev/oracleasm controversial: (ASMLib in RHEL6)
ASM filter driver -> AFD:* /dev/oracleafd/
ACFS interesting, many use cases, logs in databases for
example
Setup has evolved at CERN:
•
•
•
•
Powerpath / EMC
RHEL3: Qlogic driver
RHEL4: dm + chmod
RHEL5/6: udev (permission) and dm (multipathing)
26
Planning
•
•
•
•
•
If random IOs are crucial: “IOPS for sale,
capacity for free” (Jeffrey Steiner)
Read: for IOPS, for bandwidth
Write: log/redo and DB writer
Identify capacity, latency, bandwidth,
functionality needs
Validate with reference workloads (SLOB, fio)
and your workload (Real Application Testing)
27
Real Application Testing
Capture
Original
Replay
Target
28
Disks (1/5)
•
•
•
Disks are not reliable (any vendor, enterprise or not...), “there are
only two types of disk drives in the industry. Drives that have failed,
and drives that are about to fail “ (Jeff Bonwick)
RAID 4/5 gives a false sense reliability (see next slide)
Experience: datafile not reachable anymore, disk array double
error, (smart) move to local disk, 3rd disk failure, major recovery,
bandwidth issue, 4th disk failure... Post-mortem
29
29
Disks (2/5)
30
Disks (3/5), lessons
1* – 3 – 5*
2 – 4* – 6
•
•
Monitoring storage is crucial
Regular media check is very important (will it be
possible to read when needed what is not read on a
regular basis)
•
•
1 – 3* – 5
2* – 4 – 6*
•
Parity drive / parity blocks
ASM partner extent Note:416046.1 “A corruption in the
secondary extent will normally only be seen if the block in the
primary extent is also corrupt. ASM fixes corrupt blocks in
secondary extent automatically at next write of the block.” (if
it is over-written before it is needed ), >=11.1 preferred
read failure group before RMAN full backup, >=12.1 ASM
diskgroup scrubbing (ALTER DISKGROUP data SCRUB)
Double parity / triple mirroring
* Primary extent
31
Disks (4/5)
•
Disks are larger and larger
•
•
•
speed stay ~constant -> issue with speed
bit error rate stay constant (10-14 to 10-16), increasing
issue with availability
With x as the size and α the “bit error rate”
32
Disks, redundancy comparison
(5/5)
5
14
28
Bit error rate
10^-14
1 TB SATA desktop
RAID 1
7.68E-02
5
14
28
Bit error rate
10^-15
1TB SATA enterprise
RAID 1
7.96E-03
RAID 5 (n+1)
3.29E-01
6.73E-01
8.93E-01
RAID 5 (n+1)
3.92E-02
1.06E-01
2.01E-01
~RAID 6 (n+2)
1.60E-14
1.46E-13
6.05E-13
~RAID 6 (n+2)
1.60E-16
1.46E-15
6.05E-15
~triple mirror
8.00E-16
8.00E-16
8.00E-16
~triple mirror
8.00E-18
8.00E-18
8.00E-18
Bit error rate
10^-16
10TB SATA enterprise
450GB FC
RAID 1
Bit error rate
10^-15
RAID 1
4.00E-04
7.68E-02
RAID 5 (n+1)
2.00E-03
5.58E-03
1.11E-02
RAID 5 (n+1)
3.29E-01
6.73E-01
8.93E-01
~RAID 6 (n+2)
7.20E-19
6.55E-18
2.72E-17
~RAID 6 (n+2)
1.60E-15
1.46E-14
6.05E-14
~triple mirror
3.60E-20
3.60E-20
3.60E-20
~triple mirror
8E-17
8E-17
8E-17
33
33
Upgrades (1/3)
•
•
•
“If it is working, you should not change it!”…
But often changes come through the little
door (disk replacement, IPv6 enablement on
the network, “just a new volume”, etc.)
Example:
34
Upgrades (2/3)
•
SunCluster, new storage (3510) introduced,
1 LUN ok
• Second LUN introduced (April), all fine
• Standard maintenance operation (July),
cluster does not restart
• Multipathing: DMP / MPxIO, cluster SCSI
reservations on the shared disks
35
Upgrades (3/3)
•
•
•
•
23 hours downtime in a critical path to LHC magnets testing
Some production with reduced HW availability
Important stress...
Up to date multi-pathing layer would have avoided it (or no change)
36
Measure
•
•
•
“You can't manage what you don't measure”
The Oracle DB has wait information
(v$session, v$event_histogram) AWR /
statspack, set retention
It has additional information, ready to be
used…
37
38
AWR history
39
Slow IO is different than IO outlier
•
•
•
•
•
•
•
•
•
IO tests and planning: SLOB, fio, etc. help to size
Latency variation is the user experience (C. Millsap)
1, 19, 0, 20, 10: average 10
9, 11, 10, 9.5, 10.5 : average 10
10.01, 9.99, 10, 10.1, 9.9: average 10
10.01, 9.99, 10.01, 9.99, 500 , 10.01, 9.99: average ~10
Ex: web page: 5 SQL statements, 10 IOs per request
50 IOs at 0.2 ms =
5 ms
50 Ios at 300 ms = 1500 ms = 1.5s
40
Pathological cases, latency
•
Spinning disk latency O(10ms), Flash O(0.1ms)
• Is it always in this order of magnitude? If not,
you should know, with detailed and timing(*)
information.
• Reasons include bugs, disk failures, temporary
or global overload, etc.
(*) correlation with other sources (OS logs, storage
sub-system logs, ASM logs, etc.)
41
Complementing AWR for IO
•
•
•
AWR captures histogram information, not
single IO
AWR does not capture information on ADG
In addition, desirable to extract information
•
•
Longer term (across migrations, upgrades,
capacity planning)
With timing, slow IO operations from ASH
42
Capture Active Session History long IO
direct path read
ON CPU
db file sequential read
session 5
session 4
session 3
session 2
session 1
log file parallel write
lgwr
time
db file sequential read
direct path read
ON CPU
log file parallel write
Relative wait times indicative…
43
ASH long IO repository
•
Stores information about long(*) IO
operations
• Query it to identify major issues
• Correlate with histograms and total IO
operations
(*) longer than expected, >1s, >100ms, >10ms
44
45
46
47
48
“Low latency computing”
•
from Kevin Closson
Needed: specify storage with latency targets:
•
•
•
•
With a specified workload
99% of IO operations take less than X
99.99% of IO operations take less than Y
100% of IO operations take less than Z
49
Conclusion
•
•
Critical: availability, performance and functionality
Performance:
•
•
•
•
•
“High-availability
•
•
•
•
•
•
Spinning disk: IOPS matter, capacity “for free”
Flash: “low latency computing’, bandwidth, especially in storage network
Btw, absorbing lot of IO require CPU
Planning: SLOB, fio, Real Application Testing
HA is low technology” (Carel-Jan Engel) = complexity is the enemy of high-availability
Mirror / triple-mirror / triple parity
Some sort of scrubbing is essential
(N)FS or ASM are both solid and capable, depends on what is simpler and best known in your
organisation
RAC adds requirement for shared storage, be careful that the storage interconnect is not a
bottleneck
Latency is complex, measure and keep long term statistics (AWR retention, extract data, data
visualization), key differentiator between Flash solutions
50
I am a proud member of an
EMEA Oracle User Group
Are you a member yet?
Meet us at
south upper lobby of Moscone South
www.iouc.org
References
•
•
•
•
•
•
•
•
•
Kyle Hailey R scripts https://github.com/khailey/fio_scripts
Kevin Closson SLOB http://kevinclosson.net/slob/
Frits Hoogland gdb/strace
http://fritshoogland.wordpress.com/tag/oracle-io-performance-gdb-debuginternal-internals/
fio http://freecode.com/projects/fio
Luca Canali OraLatencyMap
http://db-blog.web.cern.ch/blog/luca-canali/2014-06-recent-updatesoralatencymap-and-pylatencymap
My Oracle Support, White Paper: ASMLIB Installation & Configuration On
MultiPath Mapper Devices (Step by Step Demo) On RAC Or Standalone
Configurations. (Doc ID 1594584.1)
Computer Architecture: A Quantitative Approach
CERN DB blog: http://cern.ch/db-blog/
53