Technical Education

October 2011
IBM Systems Group

Disk IO Tuning in AIX 6.1

Author: Dan Braden dbraden@us.ibm.com Presenter: Steve Nasypany IBM Advanced Technical Skills http://w3.ibm.com/support/americas/pseries

2010 IBM Corporation

2010 IBM Corporation

Agenda
The importance of IO tuning Disk basics and performance overview AIX IO stack Data layout Characterizing application IO Disk performance monitoring tools Testing disk subsystem performance Tuning

2
2011 IBM Corporation

Why is disk IO tuning important?

Moore's law Processors double in price performance every 18 months Disk growth Disk densities are doubling every 12 months Customers are doubling storage capacities every 12-18 months Actuator and rotational speed increasing relatively slowly Network bandwidth - doubling every 6 months Approximate CPU cycle time Approximate memory access time Approximate disk access time 0.0000000005 seconds 0.000000270 seconds 0.010000000 seconds

Memory access takes 540 CPU cycles Disk access takes 20 million CPU cycles, or 37,037 memory accesses System bottlenecks are being pushed to the disk Disk subsystems are using cache to improve IO service times Customers now spend more on storage than on servers
3
2011 IBM Corporation

Why is disk IO tuning important?

Seagate 15k RPM/3.5" Drive Specifications

+35% 450
Capacity (GB) Max Sustained DR (MB/s) Read Seek (ms)

+15%
73 75 3.6

180

-1%

3.4

2002

2010

Disk IO service time not improving compared to processors

4
2011 IBM Corporation

Performance metrics
Disk metrics MB/s IOPS With a reasonable service time Application metrics Response time Batch job run time System metrics CPU, memory and IO Size for your peak workloads Size based on maximum sustainable thruputs Bandwidth and thruput sometimes mean the same thing, sometimes not For tuning - it's good to have a short running job that's representative of your workload
5
2011 IBM Corporation

Performance metrics
Use a relevant metric for testing Should be tied to business costs, benefits or requirements Batch job run time Maximum or sufficient application transactions/second Query run time Metrics that typically are not so relevant Application transaction time if < a few seconds Metrics indicating bottlenecks CPU, memory, network, disk Important if the application metric goal isnt met Be aware of IO from other systems affecting disk performance to shared disk If benchmarking two systems, be sure the disk performance is apples to apples and youre not really comparing disk subsystem performance

6
2011 IBM Corporation

Disk performance
ZBR Geometry

Interface type
ATA SATA SCSI FC SAS
7
2011 IBM Corporation

IO service times are predominately seek + rotational latency + queueing time

Disk performance
When do you have a disk bottleneck? Random workloads Reads average > 15 ms With write cache, writes average > 2.5 ms Sequential workloads Two sequential IO streams on one disk You need more thruput
IOPS vs IO service time - 15,000 RPM disk IO service time (ms)
500 400 300 200 100 0 25 50 75 100 125 150 175 200 225 250 275 300 325

IOPS
8
2011 IBM Corporation

How to improve disk performance

Reduce the number of IOs Bigger caches Application, file system, disk subsystem Use caches more efficiently No file system logging No access time updates Improve average IO service times Better data layout Reduce locking for IOs Buffer/queue tuning Use SSDs or RAM disk Faster disks/interfaces, more disks Short stroke the disks and use the outer edge Smooth the IOs out over time Reduce the overhead to handle IOs
9
2011 IBM Corporation

What is %iowait?
A misleading indicator of disk performance A type of CPU idle Percent of time the CPU is idle and waiting on an IO so it can do some more work High %iowait does not necessarily indicate a disk bottleneck Your application could be IO intensive, e.g. a backup You can make %iowait go to 0 by adding CPU intensive jobs Low %iowait does not necessarily mean you don't have a disk bottleneck The CPUs can be busy while IOs are taking unreasonably long times If disk IO service times are good, you arent getting the performance you need, and you have significant %iowait consider using SSDs or RAM disk Improve performance by potentially reducing %iowait to 0

10
2011 IBM Corporation

Solid State Disk (SSD)

High performance electronic disk From 14,000 27,000 IOPS possible for a single SSD SSD IO bandwidth varies across Power and disk subsystems Typically small (69-177 GB) and expensive compared to HDDs Read or write IOs typically < 1 ms About the same IO service time as compared to writes to disk subsystem cache About 5-15X faster than reads from disk Positioned for high access density (IOPS/GB) random read data Implementation involves finding the best data to place on the SSDs SSDs can save disk costs by reducing the number of spindles needed When high access density data exists A mix of SSDs and HDDs is often best

11
2011 IBM Corporation

SSD vs. HDD performance

SSD offers up to 33x 125x more IOPS
125X

HDD IO service time typically 5X to 40X slower*

40X

33X 5X 1X HDD SSD

2011 IBM Corporation

1X HDD SSD
12

Access time is drive-to-drive, ignoring any caching by SAS controller

RAM disk
Use system RAM to create a virtual disk Data is lost in the event of a reboot or system crash IOs complete with RAM latencies For file systems, it takes away from file system cache Taking from one pocket and putting it into another A raw disk or file system only no LVM support
# mkramdisk 16M /dev/rramdisk0 # mkfs -V jfs2 /dev/ramdisk0 mkfs: destroy /dev/ramdisk0 (yes)? y File system created successfully. 16176 kilobytes total disk space. Device /dev/ramdisk0: Standard empty filesystem Size: 32352 512-byte (DEVBLKSIZE) blocks # mkdir /ramdiskfs # mount -V jfs2 -o log=NULL /dev/ramdisk0 /ramdiskfs # df -m /ramdiskfs Filesystem MB blocks Free %Used Iused %Iused Mounted on /dev/ramdisk0 16.00 15.67 3% 4 1% /ramdiskfs

13
2011 IBM Corporation

The AIX IO stack

Application Logical file system Raw disks Raw LVs
NFS caches file attributes NFS has a cached filesystem for NFS clients JFS and JFS2 cache use extra system RAM JFS uses persistent pages for cache JFS2 uses client pages for cache Application memory area caches data to avoid IO

JFS JFS2

NFS

Other

VMM LVM (LVM device drivers)

Multi-path IO driver (optional) Disk Device Drivers Queues exist for both adapters and disks Adapter Device Drivers Adapter device drivers use DMA for IO Disk subsystem (optional) Disk subsystems have read and write cache Disks have memory to store commands/data Disk Read cache or memory area used for IO Write cache
IOs can be coalesced (good) or split up (bad) as they go thru the IO stack IOs adjacent in a file/LV/disk can be coalesced IOs greater than the maximum IO size supported will be split up
2011 IBM Corporation

14

Synchronous vs Asynchronous IOs

Definition depends on the frame of reference Programmers/application When an application issues a synchronous IO, it waits until the IO is complete Asynchronous IOs are handed off to the kernel, and the application continues, and uses the AIO facilities in AIX When a group of asynchronous IOs complete, a signal is sent to the application Allows IO and processing to run simultaneously Filesystem IO Synchronous write IOs to a file system must get to disk Asynchronous IOs only need to get to file system cache GLVM or disk subsystem mirroring Synchronous mirroring requires that writes to both mirrors complete before returing an acknowledgement to the application Asynchronous mirroring returns an acknowledgement when the write completes at the local storage Writes to remote storage are done in the same order as locally
15
2011 IBM Corporation

Data layout
Data layout affects IO performance more than any tunable IO parameter Good data layout avoids dealing with disk hot spots An ongoing management issue and cost Data layout must be planned in advance Changes are often painful iostat and filemon can show unbalanced IO Best practice: evenly balance IOs across all physical disks Random IO best practice: Spread IOs evenly across all physical disks For disk subsystems Create RAID arrays of equal size and RAID level Create VGs with one LUN from every array Spread all LVs across all PVs in the VG The SVC can, and XIV does do this automatically
16
2011 IBM Corporation

Random IO data layout

Disk subsystem

1 2 3 4 5
RAID array LUN or logical disk PV
17
2011 IBM Corporation

datavg
# mklv lv1 e x hdisk1 hdisk2 hdisk5 # mklv lv2 e x hdisk3 hdisk1 . hdisk4 .. Use a random order for the hdisks for each LV

Data layout for sequential IO

Many factors affect sequential thruput RAID setup, number of threads, IO size, reads vs. writes Create RAID arrays with data stripes a power of 2 RAID 5 arrays of 5 or 9 disks RAID 10 arrays of 2, 4, 8, or 16 disks Do application IOs equal to, or a multiple of, a full stripe on the RAID array Or use multiple threads to submit many IOs N disk RAID 5 arrays can handle no more than N-1 sequential IO streams before the IO becomes randomized N disk RAID 10 arrays can do N sequential read IO streams and N/2 sequential write IO streams before the IO becomes randomized Sometimes smaller strip sizes (around 64 KB) perform better Test your setup if the bandwidth needed is high

18
2011 IBM Corporation

Data layout
Best practice for VGs and LVs Use Big or Scalable VGs Both support no LVCB header on LVs (only important for raw LVs) These can lead to issues with IOs split across physical disks Big VGs require using mklv T O option to eliminate LVCB Scalable VGs have no LVCB Only Scalable VGs support mirror pools (AIX 6100-02) For JFS2, use inline logs For JFS, one log per file system provides the best performance If using LVM mirroring, use active MWC Passive MWC creates more IOs than active MWC Use RAID in preference to LVM mirroring Reduces IOs as theres no additional writes for MWC Use PP striping in preference to LV striping
2011 IBM Corporation

19

LVM limits
Standard VG Max PVs/VG Max LVs/VG Max PPs/VG Max LPs/LV 32 256 32,512 32,512 Big VG 128 512 130,048 130,048 Scalable VG 1024 4096 2,097,152 2,097,152

Max PPs per VG and max LPs per LV restrict your PP size Use a PP size that allows for growth of the VG Use a PP size that allows your LVs to be spread across all PVs Unless your disk subsystem ensures your LVs are spread across all physical disks Valid LV strip sizes range from 4 KB to 128 MB in powers of 2 for striped LVs
20
2011 IBM Corporation

LUN size and how many?

Does LUN size matter? It depends. Fewer larger LUNs are easier to manage We can do many IOs in parallel to a LUN depends on its queue_depth Typically limited by: Backend physical disks Other disk subsystem bottlenecks Theoretical bandwidth is: Queue_depth/average IO service time IOPS # physical disks x physical disk IOPS taking into account use of RAID Assumes no other disk subsystem bottlenecks More LUNs mean more hdisk driver threads Very high IOPS rates will require more LUNs

21
2011 IBM Corporation

Application IO characteristics
Random IO Typically small (4-32 KB) Measure and size with IOPS Usually disk actuator limited Sequential IO Typically large (32KB and up) Measure and size with MB/s Usually limited on the interconnect to the disk actuators To determine application IO characteristics Use filemon
# filemon o /tmp/filemon.out O lf,lv,pv,detailed T 500000; sleep 90; trcstop

Check for trace buffer wraparounds which may invalidate the data, run filemon with a larger T value or shorter sleep

22
2011 IBM Corporation

filemon summary reports

Summary reports at PV and LV layers
Most Active Logical Volumes -----------------------------------------------------------------------util #rblk #wblk KB/s volume description -----------------------------------------------------------------------1.00 10551264 5600 17600.8 /dev/rms09_lv /RMS/bormspr0/oradata07 1.00 6226928 7584 10394.4 /dev/rms06_lv /RMS/bormspr0/oradata04 1.00 128544 3315168 5741.5 /dev/rms04_lv /RMS/bormspr0/oracletemp 1.00 13684704 38208 22879.4 /dev/rms02_lv /RMS/bormspr0/oradata01 0.99 11798800 16480 19698.9 /dev/rms03_lv /RMS/bormspr0/oradata02 0.99 600736 7760 1014.5 /dev/rms13_lv /RMS/bormspr0/oradata11 0.98 6237648 128 10399.9 /dev/oraloblv01 /RMS/bormspr0/oralob01 0.96 0 3120 5.2 /dev/hd8 jfslog 0.55 38056 104448 237.6 /dev/rms041_lv /RMS/bormspr0/oraredo 0.48 2344656 3328 3914.6 /dev/rms11_lv /RMS/bormspr0/oradata09 Most Active Physical Volumes -----------------------------------------------------------------------util #rblk #wblk KB/s volume description -----------------------------------------------------------------------1.00 3313059 4520 5531.2 /dev/hdisk66 SAN Volume Controller 1.00 7563668 22312 12647.6 /dev/hdisk59 SAN Volume Controller 1.00 53691 1868096 3204.1 /dev/hdisk61 SAN Volume Controller 1.00 11669 6478 30.3 /dev/hdisk0 N/A 1.00 6247484 4256 10423.1 /dev/hdisk77 SAN Volume Controller 1.00 6401393 10016 10689.3 /dev/hdisk60 SAN Volume Controller 1.00 5438693 3128 9072.8 /dev/hdisk69 SAN Volume Controller
2011 IBM Corporation

Device Device Device Device Device Device

23

filemon detailed reports

Detailed reports at PV and LV layers (only for one LV shown) Similar reports for each PV
VOLUME: /dev/rms09_lv reads: read sizes (blks): read times (msec): read sequences: read seq. lengths: writes: write sizes (blks): write times (msec): write sequences: write seq. lengths: seeks: seek dist (blks): description: /RMS/bormspr0/oradata07 23999 (0 errs) avg 439.7 min 16 max 2048 sdev 814.8 avg 85.609 min 0.139 max 1113.574 sdev 140.417 19478 avg 541.7 min 16 max 12288 sdev 1111.6 350 (0 errs) avg 16.0 min 16 max 16 sdev 0.0 avg 42.959 min 0.340 max 289.907 sdev 60.348 348 avg 16.1 min 16 max 32 sdev 1.2 19826 (81.4%) init 18262432, avg 24974715.3 min 16 max 157270944 sdev 44289553.4 time to next req(msec): avg 12.316 min 0.000 max 537.792 sdev 31.794 throughput: 17600.8 KB/sec utilization: 1.00

24
2011 IBM Corporation

Using filemon
Look at PV summary report Look for balanced IO across the disks Lack of balance may be a data layout problem Depends upon PV to physical disk mapping LVM mirroring scheduling policy also affects balance for reads IO service times in the detailed report is more definitive on data layout issues Dissimilar IO service times across PVs indicates IOs are not balanced across physical disks Look at most active LVs report Look for busy file system logs Look for file system logs serving more than one file system At 6.1, filemon also has reports showing the processes/threads doing IO to files
25
2011 IBM Corporation

Using iostat
Use a meaningful interval, 30 seconds to 15 minutes The first report is since system boot (if sys0s attribute iostat=true) Examine IO balance among hdisks Look for bursty IO (based on syncd interval) Useful flags: -T Puts a time stamp on the data -a Adapter report (IOs for an adapter) for both physical and virtual -m Disk path report (IOs down each disk path) -s System report (overall IO) -A or P For standard AIO or POSIX AIO -D for hdisk queues and IO service times -R to reset min and max values for each interval -l puts data on one line (better for scripts) -p for tape statistics -f/-F for file system statistics (AIX 6.1 TL1)
2011 IBM Corporation

26

Using iostat
# iostat <interval> <count> For individual disk and system statistics tty: tin tout avg-cpu: % user % sys % idle % iowait 24.7 71.3 8.3 2.4 85.6 3.6 Disks: % tm_act Kbps tps Kb_read Kb_wrtn hdisk0 2.2 19.4 2.6 268 894 hdisk1 5.0 231.8 28.1 1944 11964 hdisk2 5.4 227.8 26.9 2144 11524 hdisk3 4.0 215.9 24.8 2040 10916 ... # iostat ts <interval> <count> For total system statistics System configuration: lcpu=4 drives=2 ent=1.50 paths=2 vdisks=2 tty: tin tout avg-cpu: % user % sys % idle % iowait physc % entc 0.0 8062.0 0.0 0.4 99.6 0.0 0.0 0.7 Kbps tps Kb_read Kb_wrtn 82.7 20.7 248 0 0.0 13086.5 0.0 0.4 99.5 0.0 0.0 0.7 Kbps tps Kb_read Kb_wrtn 80.7 20.2 244 0 0.0 16526.0 0.0 0.5 99.5 0.0 0.0 0.8

27
2011 IBM Corporation

Using iostat
# iostat -f <interval> <count> FS Name: % tm_act / /usr /var /tmp /home /admin /proc /opt /var/adm/ras/livedum /oracle /staging /ggs -

Kbps 85.7 961.1 0.0 0.0 0.0 0.0 7.6 0.0 0.0 2.2 0.0 0.0

tps 113.3 274.1 0.0 0.0 0.0 0.0 17.3 0.0 0.0 22.9 0.0 0.0

Kb_read 257 2892 0 0 0 0 22 0 0 0 0 0

Kb_wrtn 0 0 0 0 0 0 0 0 0 6 0 0

28
2011 IBM Corporation

Using iostat
# iostat -DRTl <interval> <count>
Disks: xfers read write queue time -------------- -------------------------------- ------------------------------------ ------------------------------------ -------------------------------------- --------%tm bps tps bread bwrtn rps avg min max time fail wps avg min max time fail avg min max avg avg serv act serv serv serv outs serv serv serv outs time time time wqsz sqsz qfull hdisk41 4.6 89.8K 5.7 24.8K 65.0K 3.0 8.5 0.2 28.9 0 0 2.6 9.4 0.4 233.2 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk44 21.6 450.2K 52.0 421.5K 28.7K 51.5 4.3 0.2 39.0 0 0 0.6 5.9 0.5 30.9 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk42 6.6 57.3K 6.8 42.3K 15.0K 5.2 10.9 0.2 32.7 0 0 1.6 7.0 0.3 22.4 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk43 37.2 845.5K 101.4 818.2K 27.3K 99.9 4.0 0.2 47.6 0 0 1.5 17.2 0.4 230.2 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk37 94.4 700.0K 2.2 0.0 700.0K 0.0 0.0 0.0 0.0 0 0 2.2 1.1S 117.9 4.1S 0 0 0.0 0.0 0.1 0.0 0.1 0.0 04:52:25 hdisk53 23.5 296.2K 35.5 269.5K 26.8K 32.9 7.7 0.2 47.0 0 0 2.6 2.5 0.4 27.7 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk51 32.5 471.2K 55.6 445.5K 25.7K 54.4 6.7 0.2 58.8 0 0 1.2 3.1 0.4 13.0 0 0 0.0 0.0 0.1 0.0 0.0 0.0 04:52:25 hdisk56 19.5 178.0K 20.7 122.3K 55.7K 14.9 9.8 0.2 55.0 0 0 5.7 55.8 0.4 318.9 0 0 2.8 0.0 194.4 0.0 0.0 0.6 04:52:25 hdisk48 18.0 149.6K 18.0 101.0K 48.6K 12.3 10.6 0.2 38.5 0 0 5.7 19.0 0.4 250.2 0 0 0.0 0.0 3.7 0.0 0.0 0.3 04:52:25 hdisk46 12.9 167.4K 19.8 156.7K 10.6K 19.1 6.8 0.2 37.5 0 0 0.7 4.4 0.4 17.0 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk57 55.2 608.8K 71.1 574.4K 34.4K 69.5 8.9 0.2 118.3 0 0 1.6 10.1 0.4 216.3 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk55 13.4 244.9K 29.8 234.0K 10.9K 28.6 4.8 0.2 36.9 0 0 1.3 2.6 0.4 22.3 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk50 48.6 616.7K 73.3 575.5K 41.2K 70.3 7.9 0.2 84.5 0 0 3.1 5.7 0.4 40.1 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25 hdisk52 14.5 174.2K 20.6 116.0K 58.1K 14.2 7.7 0.2 36.9 0 0 6.5 10.7 0.4 270.1 0 0 0.0 0.0 0.0 0.0 0.0 0.0 04:52:25

Shows average IO service times for reads and writes, IO rates, IOPS (tps) and time in the hdisk driver queue One can calculate R/W ratio and average IO size Time spent in the queue indicates increasing queue_depth may improve performance sqfull = number of times the hdisk drivers service queue was full avgserv = average IO service time avgsqsz = average service queue size This can't exceed queue_depth for the disk avgwqsz = average wait queue size Waiting to be sent to the disk If avgwqsz is often > 0, then increase queue_depth If sqfull in the first report is high, then increase queue_depth
29
2011 IBM Corporation

Using sar
sar -d formerly reported zeros for avwait and avserv avque definition changes in AIX 5.3 # sar -d 1 2 AIX sq1test1 3 5 00CDDEDC4C00 06/22/04 System configuration: lcpu=2 drives=1 ent=0.30 10:01:37 10:01:38 10:01:39 Average device %busy avque r+w/s Kbs/s avwait avserv hdisk0 100 36.1 363 46153 51.1 8.3 hdisk0 99 38.1 350 44105 58.0 8.5 hdisk0 99 37.1 356 45129 54.6 8.4

avque - average IOs in the wait queue Waiting to get sent to the disk (the disk's queue is full) Values > 0 indicate increasing queue_depth may help performance Used to mean number of IOs in the disk queue avgwait - time (ms) waiting in wait queue avgserv - IO service time (ms) when sent to the disk
30
2011 IBM Corporation

Using lvmstat
Provides IO statistics for LVs, VGs and PPs Useful for SSD data placement You must enable data collection first for a VG: # lvmstat e v <vgname> Useful to find busy LVs and PPs
root/ # lvmstat -sv rootvg <interval length> <number of intervals> Logical Volume iocnt Kb_read Kb_wrtn Kbps hd8 212 0 848 24.00 hd4 11 0 44 0.23 hd2 3 12 0 0.01 hd9var 2 0 8 0.01 .. hd8 3 0 12 8.00 . hd8 12 0 48 32.00 hd4 1 0 4 2.67 # lvmstat -l lv00 1 Log_part mirror# iocnt Kb_read Kb_wrtn Kbps 1 1 65536 32768 0 0.02 2 1 53718 26859 0 0.01 Log_part mirror# iocnt Kb_read Kb_wrtn Kbps 2 1 5420 2710 0 14263.16 Log_part mirror# iocnt Kb_read Kb_wrtn Kbps 3 1 4449 2224 0 13903.12 2 1 979 489 0 3059.38
2011 IBM Corporation

31

Using NMON
# nmon - then press a for all adapters or ^ for FC adapters

Easy way to monitor adapter thruput NMON can also be used to create Excel graphs showing IO over time Plus CPU, memory, and network IO data
32
2011 IBM Corporation

Testing thruput
Sequential IO Test sequential read thruput from a device: # timex dd if=<device> of=/dev/null bs=1m count=100 # timex dd if=/dev/rhdisk20 of=/dev/null bs=1m count=1024 1024+0 records in. 1024+0 records out. real 3.44 user 0.00 sys 0.17 1024 MB/3.44 s = 297.7 MB/s Test sequential write thruput to a device: # timex dd if=/dev/zero of=<device> bs=1m count=100 Note that /dev/zero writes the null character, so writing this character to files in a file system will result in sparse files For file systems, either create a file, or use the lptest command to generate a file, e.g., # lptest 127 32 > 4kfile
33
2011 IBM Corporation

Testing thruput with ndisk

Use ndisk which is part of the nstress package, or ndisk64 for structures > 2GB http://www.ibm.com/collaboration/wiki/display/WikiPtype/nstress Do IO to a file or raw LV or hdisk Do IO to multiple devices or files Specify the number of threads doing IO You need a lot of threads doing IO to stress a disk subsystem Synchronous or asynchronous writes to file system files Specify the IO size or a set of IO sizes Specify the R/W ratio

34
2011 IBM Corporation

Testing thruput with ndisk

# ndisk64 -R -f /dev/rhdisk20 -r 100 -t 30 M 20 s 716800 Command: ndisk -R -f /dev/rhdisk20 -r 100 -t 60 Synchronous Disk test (regular read/write) No. of processes = 20 I/O type = Random Block size = 4096 Read-Write = Read only Sync type: none = just close the file Number of files = 1 File size = 716800 MB Run time = 30 seconds Snooze % = 0 percent ----> Running test with block Size=4096 (4KB) . Proc - <-----Disk IO----> | <-----Throughput------> RunTime Num TOTAL IO/sec | MB/sec KB/sec Seconds 1 20 TOTALS 12577 12577 251540 419.2 | 1.64 1676.94 30.00

419.2 | 1.64 1676.98 30.00 8384.8 | 32.75 Rand procs= 20 read=100% bs=4KB

35
2011 IBM Corporation

Dealing with cache effects

Prime the cache (recommended) Run the test twice or more and ignore the first run It's recommended to prime the cache, as most applications will be using it and you've paid for it, so you should use it or Flush the cache Unmount and remount file systems For disk subsystems, use #cat <unused file(s)> > /dev/null The unused file(s) must be larger than the disk subsystem read cache Write cache If we fill up the write cache, IO service times will be at disk speed, not cache speed Use a long running job Reads from the disk subsystem will also inhibit unloading of the write cache
36
2011 IBM Corporation

AIX 6.1 Restricted Tunables

Some ioo/vmo/schedo/raso/nfso/no tuning parameters are now restricted Generally should only be changed per AIX Support Display all the tunables using: # <ioo|vmo|schedo|raso|nfso|no> -FL Display non-restricted tunables without the F flag smit access via # smitty tuningDev Dynamic change will show a warning message Permanent changes require a confirmation Permanent changes will result in a warning message at boot in the error log Some restricted tunables relating to disk IO tuning include:
Most aio tunables j2_nBufferPerPagerDevice minpgahead numclust pv_min_pbuf sync_release_ilock lru_file_repage lru_poll_interval maxperm minperm strict_maxclient strict_maxperm page_steal_method
37
2011 IBM Corporation

Tuning IO buffers and queues

General rule increase buffers or queue depths until either: You arent running out of buffers or filling the queue IO service times indicate a bottleneck at the disk subsystem or SAN IOs are delayed due to lack of a buffer or a queue slot Disks and disk subsystem have limits to the maximum number of in-flight IOs they can handle More than this will result in lost IOs, time outs and resubmission of the IO which severely affects IO performance

38
2011 IBM Corporation

The AIX IO stack

Application Logical file system Raw disks Raw LVs

JFS JFS2

NFS

Other

File system buffers at this layer

VMM LVM (LVM device drivers) Multi-path IO driver (optional) Disk Device Drivers Adapter Device Drivers Disk subsystem (optional) Disk Write cache
Disk buffers (pbufs) at this layer hdisk queue_depth adapter num_cmd_elems

Read cache or memory area used for IO

39
2011 IBM Corporation

Tuning IO buffers
# vmstat -v | tail -5 <-- only last 5 lines needed 0 pending disk I/Os blocked with no pbuf 0 paging space I/Os blocked with no psbuf 8755 filesystem I/Os blocked with no fsbuf 0 client filesystem I/Os blocked with no fsbuf 2365 external pager filesystem I/Os blocked with no fsbuf

First field is a count of IOs delayed since boot due to lack of the specified buffer For pbufs, use lvmo to increase pv_pbuf_count (see the next slide) For psbufs, stop paging or add paging spaces For filesystem fsbufs, increase numfsbufs with ioo For external pager fsbufs, increase j2_dynamicBufferPreallocation with ioo For client filesystem fsbufs, increase nfso's nfs_v3_pdts and nfs_v3_vm_bufs (or the NFS4 equivalents) Run # ioo FL to see defaults, current settings and whats required to make the changes go into effect

40
2011 IBM Corporation

Disk pbuf tuning

Disk buffers are configurable for disks in VGs
lvmo -v VgName -o Tunable [ =NewValue ] lvmo [-v VgName] -a # lvmo -a vgname = rootvg pv_pbuf_count = 512 - Number of pbufs added when one PV is added to the VG total_vg_pbufs = 512 - Current pbufs available for the VG max_vg_pbuf_count = 16384 - max pbufs available for the VG, requires remount pervg_blocked_io_count = 1 - delayed IO count since last varyon for the VG pv_min_pbuf = 512 - Minimum number of pbufs added when PV is added to any VG global_blocked_io_count = 1 - System wide delayed IO count

To increase a VGs pbufs dynamically: # lvmo v <vgname> -o pv_pbuf_count=<new value> pv_min_pbuf is tuned via ioo or lvmo Changes to pv_pbuf_count via lvmo are dynamic Increase value, collect statistics and change again if necessary See # lvmo L

41
2011 IBM Corporation

Queue depth tuning

If IO service times are reasonably good, but queues are getting filled up, then Increase queue depths until: You arent filling the queues or IO service times start degrading Youve moved the bottleneck to the disk For hdisks, queue_depth controls the maximum number of in-flight IOs For FC adapters, num_cmd_elems controls maximum number of in-flight IOs Drivers for hdisks and adapters have service and wait queues When the queue is full and an IO completes, then another is issued

42
2011 IBM Corporation

Queue depth tuning

IO flow: Multipath IO code submits IO to hdisk driver SDD queues IOs and wont submit more than queue_depth IOs to a hdisk Disable this with # datapath qdepth disable for heavy IO SDDPCM does not queue IOs Hdisk driver has in process and wait queues in process queue contains up to queue_depth IOs Hdisk driver submits IOs to adapter driver Adapter driver has in process and wait queues FC adapter in process queue contains up to num_cmd_elems IOs Adapter driver uses DMA to do IOs Adapter driver submits IOs to disk subsystem

43
2011 IBM Corporation

Queue depth tuning

Useful things to know about attributes for all AIX devices: List device attributes with # lsattr EHl <device>
# lsattr -EHl hdisk10 attribute value PCM algorithm hcheck_cmd hcheck_interval hcheck_mode max_transfer pvid queue_depth reserve_policy PCM/friend/vscsi fail_over test_unit_rdy 0 nonactive 0x40000 00cee79eaed0872d0000000000000000 3 no_reserve description Path Control Module Algorithm Health Check Command Health Check Interval Health Check Mode Maximum TRANSFER Size Physical volume identifier Queue DEPTH Reserve Policy user_settable False True True True True True False True True

Attributes with user_settable=True can be changed List allowable values for an attribute with # lsattr Rl <device> -a <attribute>
# lsattr -Rl hdisk10 -a queue_depth 1...256 (+1)

To change an attribute use # chdev l <device> -a <attribute>=<new value> -P Then reboot, or if the device is not in use, eliminate the -P so the change is immediately effective

44
2011 IBM Corporation

Queue depth tuning

Fibre channel adapter attributes:
# lsattr -El fcs0 bus_intr_lvl 8355 bus_io_addr 0xffc00 bus_mem_addr 0xf8040000 init_link al intr_priority 3 lg_term_dma 0x1000000 max_xfer_size 0x100000 num_cmd_elems 200 pref_alpa 0x1 sw_fc_class 2 Bus interrupt level False Bus I/O address False Bus memory address False INIT Link flags True Interrupt priority False Long term DMA True Maximum Transfer Size True Maximum number of COMMANDS to queue to the adapter True Preferred AL_PA True FC Class for Fabric True

The max_xfer_size attribute also controls a DMA memory area used to hold data for transfer, and at the default is 16 MB Changing to other allowable values increases it to 128 MB and increases the adapters bandwidth Often changed to 0x200000 This can result in a problem if there isnt enough memory on PHB chips in the IO drawer with too many adapters/devices on the PHB Make the change and reboot check for Defined devices or errors in the error log, and change back if necessary For NPIV and virtual FC adapters the DMA memory area is 128 MB at 6.1 TL2 or later
45
2011 IBM Corporation

Queue depth tuning

How to determine if queues are being filled With SDDPCM: # pcmpath query devstats <interval> <count> # pcmpath query adaptstats <interval> <count> With SDD: # datapath query devstats <interval> <count> # datapath query adaptstats <interval> <count> With iostat: # iostat D for data since system boot and # iostat D <interval> <count> for interval data With fcstat: # fcstat fcs0

46
2011 IBM Corporation

Queue depth tuning

Sample output:
# pcmpath query devstats DEV#: 4 DEVICE NAME: hdisk4 =============================== Total Read Total Write I/O: 446958 1510783 SECTOR: 13006902 25715160 Transfer Size: <= 512 1032 <= 4k 1825589 Active Read 0 0 <= 16K 40591 Active Write 0 0 <= 64K 54946 Maximum 128 24536 > 64K 35583

Indicates that we queued 128 IOs to the hdisk driver If queue_depth is 128, we filled the queue
# iostat -D hdisk10 System configuration: lcpu=4 drives=35 paths=35 vdisks=2 hdisk10 xfer: read: write: queue: %tm_act 0.1 rps 0.0 wps 0.2 avgtime 5.4 bps 1.4K avgserv 4.6 avgserv 8.2 mintime 0.0 tps 0.2 minserv 0.3 minserv 0.5 maxtime 579.5 bread bwrtn 442.6 940.9 maxserv timeouts 67.9 0 maxserv timeouts 106.8 0 avgwqsz avgsqsz 0.0 0.0

fails 0 fails 0 sqfull 0.4

The sqfull value represents the number of times we filled the queue per second Non-zero values indicate we filled the queue
47
2011 IBM Corporation

Adapter queue depth tuning

# pcmpath query adaptstats Adapter #: 0 ============= Total Read Total Write Active Read Active Write Maximum 200 I/O: 1105909 78 3 0 SECTOR: 8845752 0 24 0 88

Maximum of 200 with num_cmd_elems=200 means we filled the queue

# fcstat fcs0 FC SCSI Adapter Driver Information No DMA Resource Count: 4490 No Adapter Elements Count: 105688 No Command Resource Count: 133

<- Increase max_xfer_size <- Increase num_cmd_elems

48
2011 IBM Corporation

VIO
The VIO Server (VIOS) uses multi-path IO code for the attached disk subsystems The VIO client (VIOC) always uses SCSI MPIO if accessing storage thru two VIOSs In this case only entire LUNs are served to the VIOC Set the queue_depth at the VIOC equal to queue_depth at the VIOS for the LUN If you increase vFC adapter num_cmd_elems, also do it on the real FC adapter Preferably treat the real FC adapter num_cmd_elems as a shared resource The VSCSI adapter has a queue also To avoid queuing on the VSCSI adapter: Max LUNs/VSCSI adapter =INT(510/(Q+3)) Q is the queue depth of the LUN assuming all are the same One can monitor adapters with NMON in the oem_setup_env shell

49
2011 IBM Corporation

Read ahead
Read ahead detects that we're reading sequentially and gets the data before the application requests it Reduces %iowait Too much read ahead means you do IO that you don't need Operates at the file system layer - sequentially reading files Set maxpgahead for JFS and j2_maxPgReadAhead for JFS2 Values of 1024 for max page read ahead are not unreasonable Disk subsystems read ahead too - when sequentially reading disks Improves IO service time and thruput Tunable on DS4000, fixed on ESS, DS6000, DS8000 and SVC

50
2011 IBM Corporation

Write behind
Initiates writes from file system cache before syncd does it Write behind tuning for sequential writes to a file Tune numclust for JFS Tune j2_nPagesPerWriteBehindCluster for JFS2 These represent 16 KB clusters Larger values allow IO to be coalesced When the specified number of sequential 16 KB clusters are updated, start the IO to disk rather than wait for syncd Write behind tuning for random writes to a file Tune maxrandwrt for JFS Tune j2_maxRandomWrite and j2_nRandomCluster for JFS2 Max number of random writes allowed to accumulate to a file before additional IOs are flushed, default is 0 or off j2_nRandomCluster specifies the number of clusters apart two consecutive writes must be in order to be considered random If you have bursty IO, consider using write behind to smooth out the IO rate
51
2011 IBM Corporation

More on write behind and syncd

syncd - system wide file system cache flushing Historical Unix feature to improve performance Applies to asynchronous IOs (not necessarily aio) inode is locked when each file is synced There is a tradeoff: Longer intervals allow more IO to be coalesced Longer intervals can: Create bursty IO Bursty IO can slow down other IO As IOPS increase, IO service times also increase

52
2011 IBM Corporation

JFS2 Sync Tunables

ioo JFS2 Sync Tunables
The file system sync operation can be problematic in situations where there is very heavy random I/O activity to a large file. When a sync occurs all reads and writes from user programs to the file are blocked. With a large number of dirty pages in the file the time required to complete the writes to disk can be large. New JFS2 tunables are provided to relieve that situation. j2_syncPageCount
Limits the number of modified pages that are scheduled to be written by sync in one pass for a file. When this tunable is set, the file system will write the specified number of pages without blocking IO to the rest of the file. The sync call will iterate on the write operation until all modified pages have been written. Default: 0 (off), Range: 0-65536, Type: Dynamic, Unit: 4KB pages

j2_syncPageLimit
Overrides j2_syncPageCount when a threshold is reached. This is to guarantee that sync will eventually complete for a given file. Not applied if j2_syncPageCount is off. Default: 16, Range: 1-65536, Type: Dynamic, Unit: Numeric

If application response times are impacted by syncd, try j2_syncPageCount settings from 256 to 1024. Smaller values improve short term response times, but still result in larger syncs that impact response times over larger intervals. These will likely require a lot of experimentation, and detailed analysis of IO behavior. Does not apply to mmap() memory mapped files. May not apply to shmat() files (TBD)

53
2011 IBM Corporation

Mount options
Release behind: rbr, rbw and rbrw Says to throw the data out of file system cache rbr is release behind on read rbw is release behind on write rbrw is both Applies to sequential IO only DIO: Direct IO Bypasses file system cache No file system read ahead No lrud or syncd overhead No double buffering of data Half the kernel calls to do the IO Half the memory transfers to get the data to the application Requires the application be written to use DIO CIO: Concurrent IO The same as DIO but with no i-node locking
54
2011 IBM Corporation

The AIX IO stack

Application Logical file system Raw disks Raw LVs
i-node locking: when 2 or more threads access the same file, and one is a write, the write will block all other read/write threads at this level

JFS JFS2

NFS

Other

VMM LVM (LVM device drivers) Multi-path IO driver Disk Device Drivers Adapter Device Drivers Disk subsystem (optional) Disk Write cache Read cache or memory area used for IO

55
2011 IBM Corporation

Mount options
Direct IO IOs must be aligned on file system block boundaries IOs that dont adhere to this will dramatically reduce performance Avoid large file enabled JFS file systems - block size is 128 KB after 4 MB

# mount -o dio # chfs -a options=rw,dio <file system>

DIO process
1 Application issues IO request to kernel 2 Kernel initiates disk IO 3 Data transferred to/from application buffer to disk

Normal file system IO process File system reads

1 Application issues read request 2 Kernel checks FS cache 1 Data found - kernel copies it to the application buffer 2 Data not found - kernel does disk IO 3 Data transferred FS cache 4 Kernel copies data to app buffer

File system writes

1 Application issues write request 2 Kernel writes data to FS cache and returns acknowledgment to app 1 Application continues 3 Syncd periodically flushes dirty cache to disk

56
2011 IBM Corporation

Mount options
Concurrent IO for JFS2 (not JFS): # mount -o cio # chfs -a options=rw,cio <file system> Assumes that the application ensures data integrity for multiple simultaneous IOs to a file Changes to meta-data are still serialized I-node locking: When two threads (one of which is a write) to do IO to the same file are at the file system layer of the IO stack, reads will be blocked while a write proceeds Provides raw LV performance with file system benefits Requires an application designed to use CIO For file system maintenance (e.g. restore a backup) one usually mounts without cio during the maintenance Some applications now make CIO/DIO calls directly without requiring cio/dio mounts, in which case dont use the mount options Important for times when alignment requirements arent met, or when file system read ahead helps (like during backups)
57
2011 IBM Corporation

No time for atime

Ingo Molnar (Linux kernel developer) said: "It's also perhaps the most stupid Unix design idea of all times. Unix is really nice and well done, but think about this a bit: 'For every file that is read from the disk, lets do a ... write to the disk! And, for every file that is already cached and which we read from the cache ... do a write to the disk!'" If you have a lot of file activity, you have to update a lot of timestamps File timestamps File creation (ctime) File last modified time (mtime) File last access time (atime) The noatime mount option disables last access time updates for JFS2 File systems with heavy inode access activity due to file opens can have significant performance improvements First customer benchmark efix reported 35% improvement with DIO noatime mount (20K+ files) Most customers should expect much less for production environments # mount a rw,noatime /myfilesystem

58
2011 IBM Corporation

Disabling file system journaling

You may lose data in the event of a system crash! Improves performance for meta-data changes to file systems When you frequently add, delete or change the size of files Eliminates IOs to file system log JFS # mount o nointegrity <file system> JFS2 # mount -o log=NULL <file system>

59
2011 IBM Corporation

IO Pacing
IO pacing - causes the CPU to do something else with a specified amount of IOs to a file in process Turning it off improves backup times and thruput Turning it on ensures that no process hogs CPU for IO, and ensures good keyboard response on systems with heavy IO workloads Default is on with minpout=4096 and maxpout=8193 Originally used to avoid HACMP's dead man switch Old HACMP recommended values of 33 and 24 significantly inhibit thruput but are reasonable for uniprocessors with noncached disk Changed via:
# chgsys l sys0 a maxpout=<new value> -a minpout=<new value>

IO pacing per file system via the mount command # mount -o minpout=256 o maxpout=513 /myfs

60
2011 IBM Corporation

Asynchronous IO
Asynchronous IO automatically turned on at AIX 6.1 AIO kernel threads automatically exit after aio_server_inactivity seconds AIO kernel threads not used for AIO to raw LVs or CIO mounted file systems Only aio_maxservers and aio_maxreqs need to be changed Defaults are 21 and 8K respectively per logical CPU Set via ioo Some may want to adjust minservers for heavy AIO use maxservers is the maximum number of AIOs that can be processed at any one time maxreqs is the maximum number of AIO requests that can be handled at one time and is a total for the system (they are queued to the AIO kernel threads) Typical values:
Default 3 10 4096 OLTP 200 800 16384 SAP 400 1200 16384

minservers maxservers maxreqs

61
2011 IBM Corporation

Asynchronous IO tuning
Use iostat A to monitor AIO (or -P for POSIX AIO)
#

iostat -A <interval> <number of intervals> System configuration: lcpu=4 drives=1 ent=0.50 aio: avgc avfc maxg maxf maxr avg-cpu: %user %sys %idle %iow physc %entc 25 6 29 10 4096 30.7 36.3 15.1 17.9 0.0 81.9 Disks: % tm_act Kbps tps Kb_read Kb_wrtn hdisk0 100.0 61572.0 484.0 8192 53380

avgc - Average global non-fastpath AIO request count per second for the specified interval avfc - Average AIO fastpath request count per second for the specified interval for IOs to raw LVs (doesnt include CIO fast path IOs) maxg - Maximum non-fastpath AIO request count since the last time this value was fetched maxf - Maximum fastpath request count since the last time this value was fetched maxr - Maximum AIO requests allowed - the AIO device maxreqs attribute If maxg or maxf gets close to maxr or maxservers then increase maxreqs or maxservers
62
2011 IBM Corporation

SAN attached FC adapters

Set fscsi dynamic tracking to yes Allows dynamic SAN changes Set FC fabric event error recovery fast_fail to yes if the switches support it Switch fails IOs immediately without timeout if a path goes away Switches without support result in errors in the error log
# lsattr -El attach dyntrk fc_err_recov scsi_id sw_fc_class fscsi0 switch no delayed_fail 0x1c0d00 3

How this adapter is CONNECTED False Dynamic Tracking of FC Devices True FC Fabric Event Error RECOVERY Policy True Adapter SCSI ID False FC Class for Fabric True

# chdev l fscsi0 a dyntrk=yes a fc_err_recov=fast_fail P # shutdown Fr

63
2011 IBM Corporation

VMM the Virtual Memory Manager

All IO requires memory All input goes to memory, all output comes from memory File system cache consumes memory File system cache takes CPU cycles to manage Initial tuning recommendations: minfree = max [ 960, lcpus x 120 / mempools ] maxfree = minfree + MaxIOSizeIn4KBunits x lcpus / mempools lcpus = number of logical CPUs mempools = number of memory pools (from vmstat v) MaxIOSizeIn4KBunits is j2_maxPageReadAhead for file system IOs, or what you expect your applications biggest IO request to be The idea is that when we call lrud to free up memory for a large IO, we only call it once and it frees up sufficient memory, but not too much
64
2011 IBM Corporation

QUESTIONS? COMMENTS?

65
2011 IBM Corporation