Linux on System z disk I/O performance

Mario Held mario.held@de.ibm.com
Martin Kammerer martin.kammerer@de.ibm.com
07/07/2010

Linux on System z disk I/O performance

visit us at http://www.ibm.com/developerworks/linux/linux390/perf/index.html

© 2010 IBM Corporation

Linux on System z Performance Evaluation

Agenda

 Linux
– file system
– special functions
– logical volumes
– monitoring
– High performance FICON
 Storage server
– internal structure
– recommended usage and configuration
 Host
– Linux I/O stack
– sample I/O attachment
– FICON/ECKD and FCP/SCSI I/O setups
– possible bottlenecks
– recommendations

2 Linux on System z disk I/O performance © 2010 IBM Corporation


Linux file system

 Use ext3 instead of reiserfs
 Tune your ext3 file system
– Select the appropriate journaling mode (journal, ordered, writeback)
– Consider to turn off atime
– Consider upcoming ext4
 Temporary files
– Don't place them on journaling file systems
– Consider a ram disk instead of a disk device
– Consider tmpfs for temporary files
 If possible, use direct I/O to avoid double buffering of files in memory
 If possible, use async I/O to continue program execution while data is fetched
– Important for read operations
– Applications usually are not depending on write completions



I/O options

 Direct I/O (DIO)
– Transfer the data directly from the application buffers to the device driver, avoids copying the
data to the page cache
– Advantages:
• Saves page cache memory and avoids caching the same data twice
• Allows larger buffer pools for databases
– Disadvantage:
• Make sure that no utility is working through the file system (page cache) --> danger of
data corruption
• The size of the I/O requests may be smaller
 Asynchronous I/O (AIO)
– The application is not blocked for the time of the I/O operation
– It resumes its processing and gets notified when the I/O is completed.
– Advantage
• the issuer of a read/write operation is no longer waiting until the request finishes.
• reduces the number of I/O processes (saves memory and CPU)
 Recommendation is to use both
– Database benchmark workloads improved throughput by 50%


I/O schedulers (1)

 Four different I/O schedulers are available
– noop scheduler
only request merging
– deadline scheduler
avoids read request starvation
– anticipatory scheduler (as scheduler)
designed for the usage with physical disks, not intended for storage subsystems
– complete fair queuing scheduler (cfq scheduler)
all users of a particular drive would be able to execute about the same number of I/O requests
over a given time.
 The default in current distributions is deadline
– Don't change the setting to as in Linux on System z. The throughput will be impacted
significantly



HowTo

 How to identify which I/O scheduler is used
– Searching (grep) in file /var/log/boot.msg for phrase 'io scheduler' will find a line like:
• <4>Using deadline io scheduler
– which returns the name of the scheduler in use
 How to select the scheduler
– The I/O scheduler is selected with the boot parameter elevator in zipl.conf
• [ipl]
• target = /boot/zipl
• image = /boot/image
• ramdisk = /boot/initrd
• parameters = "maxcpus=8 dasd=5000 root=/dev/dasda1 elevator=deadline”
• where elevator ::= as | deadline | cfq | noop
 For more details see /usr/src/linux/Documentation/kernel-parameters.txt.



Logical volumes

 Linear logical volumes allow an easy extension of the file system
 Striped logical volumes
– Provide the capability to perform simultaneous I/O on different stripes
– Allow load balancing
– Are also extendable
 Don't use logical volumes for “/” or “/usr”
– If the logical volume gets corrupted your system is gone
 Logical volumes require more CPU cycles than physical disks
– Consumption increases with the number of physical disks used for the logical volume



Monitoring disk I/O
Merged requests
I/Os per second
Throughput
Request size
Queue length
Service times
utilization

Output from iostat -dmx 2 /dev/sda

Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util

sda 0.00 10800.50 6.00 1370.00 0.02 46.70 69.54 7.03 5.11 0.41 56.50
sequential write
---------------------------------------------------------------------------------------------------------

sda 0.00 19089.50 21.50 495.00 0.08 76.59 304.02 6.35 12.31 1.85 95.50
sequential write
---------------------------------------------------------------------------------------------------------

sda 0.00 15610.00 259.50 1805.50 1.01 68.14 68.58 26.07 12.23 0.48 100.00
random write
---------------------------------------------------------------------------------------------------------

sda 1239.50 0.00 538.00 0.00 115.58 0.00 439.99 4.96 9.23 1.77 95.00
sequential read
---------------------------------------------------------------------------------------------------------

sda 227.00 0.00 2452.00 0.00 94.26 0.00 78.73 1.76 0.72 0.32 78.50
random read



High Performance FICON

 Advantages:
– Provides a simpler link protocol than FICON
– Is capable of sending multiple channel commands to the control unit in a single entity
 Performance expectation:
– Should improve performance of database OLTP workloads
 Prerequisites:
– z10 GA2 or newer
– FICON Express2 or newer
– DS8000 R4.1 or newer
– DS8000 High Performance FICON feature
– A Linux dasd driver that supports read/write track data and HPF
• Included in SLES11 and future Linux distributions, service packs or updates)



High Performance FICON results (1)

 For sequential write throughput improves with read/write track data up to 1.2x

Throughput for sequential write [MB/s]
1200

1000
old dasd I/O
800 new read/write track
data
new HPF (includes
600 read/write track data)

400

200

0
1 2 4 8 16 32 64
number of processes



 For sequential read throughput improves with read/write track data up to 1.6x

Throughput for sequential read [MB/s]
2500

2000
old dasd I/O
new read/write track
1500 data
new HPF (includes
read/write/track data)

1000

500

0
1 2 4 8 16 32 64
number of processes



 With small record sizes, as e.g. used by databases, HPF throughput improves up to 2.2x
versus old Linux channel programs
Throughput for random reader [MB/s]
1000
900
800
old dasd I/O
700 new read/write track
600 data
new HPF (includes
500 read/write track data)

400
300
200
100
0
1 2 4 8 16 32 64
number of processes


DS8000 Disk setup

 Don't treat a storage server as a black box, understand its structure
 Principles apply to other storage vendor products as well
 Several conveniently selected disks instead of one single disk can speed up the sequential
read/write performance to more than a triple. Use the logical volume manager to set up the
disks.
 Avoid using subsequent disk addresses in a storage server (e.g. the addresses 5100, 5101,
5102,.. in an IBM Storage Server), because
– they use the same rank
– they use the same device adapter.
 If you ask for 16 disks and your system administrator gives you addresses 5100-510F
– From a performance perspective this is close to the worst case



DS8000 Architecture
●
structure is complex
- disks are connected via two internal
FCP switches for higher bandwidth
●
the DS8000 is still divided into two parts
named processor complex or just server
- caches are organized per server
●
one device adapter pair addresses
4 array sites
●
one array site is build from 8 disks
- disks are distributed over the front and rear
storage enclosures
- have the same color in the chart
●
one RAID array is defined using one
array site
●
one rank is built using one RAID array
●
ranks are assigned to an extent pool
●
extent pools are assigned to one of the
servers
- this assigns also the caches
●
the rules are the same as for ESS
- one disk range resides in one extent pool



Rules for selecting disks

 This makes it fast
– Use as many paths as possible (CHPID -> host adapter)
– Spread the host adapters used across all host adapter bays
• For ECKD switching of the paths is done automatically
• FCP needs a fixed relation between disk and path
– Use Linux multipathing for load balancing
– Select disks from as many ranks as possible!
– Switch the rank for each new disk in a logical volume
– Switch the ranks used between servers and device adapters
– Avoid reusing the same resource (path, server, device adapter, and disk) as long as possible
 Goal is to get a balanced load on all paths and physical disks
 In addition striped Linux logical volumes and / or storage pool striping may help to increase
the overall throghput.



Linux kernel components involved in disk I/O
Application program Issuing reads and writes

VFS dispatches requests to different devices and
Virtual File System (VFS) Translates to sector addressing

Logical Volume Manager (LVM) LVM defines the physical to logical device relation
Multipath Multipath sets the multipath policies
Device mapper (dm) dm holds the generic mapping of all block devices and
Block device layer performs 1:n mapping for logical volumes and/or multipath

Page cache Page cache contains all file I/O data, direct I/O bypasses the
page cache
I/O scheduler I/O schedulers merge, order and queue requests, start
device drivers via (un)plug device

Data transfer handling

Direct Acces Storage Device driver Small Computer System Interface driver (SCSI)

Device drivers z Fiber Channel Protocol driver (zFCP)
Queued Direct I/O driver (qdio)


Disk I/O attachment and storage subsystem

 Each disk is a configured volume in an
extent pool (here extent pool = rank) Device ranks
 An extent pool can be configured for Adapter
either ECKD dasds or SCSI LUNs 1 ECKD
 Host bus adapters connect internally to

Server 0
both servers 2 SCSI
Channel Host Bus
path 1 Adapter 1 ECKD
3
Channel Host Bus
4 SCSI
Switch

path 2 Adapter 2

Channel Host Bus 5 ECKD
path 3 Adapter 3
Server 1
6 SCSI
Channel Host Bus
path 4 Adapter 4
7 ECKD
FICON
Express 8 SCSI


I/O processing characteristics

 FICON/ECKD:
– 1:1 mapping host subchannel:dasd
– Serialization of I/Os per subchannel
– I/O request queue in Linux
– Disk blocks are 4KB
– High availability by FICON path groups
– Load balancing by FICON path groups and Parallel Access Volumes
 FCP/SCSI
– Several I/Os can be issued against a LUN immediately
– Queuing in the FICON Express card and/or in the storage server
– Additional I/O request queue in Linux
– Disk blocks are 512 bytes
– High availability by Linux multipathing, type failover or multibus
– Load balancing by Linux multipathing, type multibus



General layout for FICON/ECKD

 The dasd driver starts the I/O on a subchannel
Application program
 Each subchannel connects to all channel paths in the path group
 Each channel connects via a switch to a host bus adapter
VFS
 A host bus adapter connects to both servers
 Each server connects to its ranks
LVM
Multipath DA ranks
dm
Block device layer Channel 1
subsystem

Server 0
Page cache
I/O scheduler a chpid 1 HBA 1
3
b chpid 2 HBA 2
dasd driver Switch
c chpid 3 HBA 3 5

Server 1
d chpid 4 HBA 4
7



FICON/ECKD dasd I/O to a single disk

 Assume that subchannel a corresponds to disk 2 in rank 1
Application program  The full choice of host adapters can be used
 Only one I/O can be issued at a time through subchannel a
VFS  All other I/Os need to be queued in the dasd driver and in the block device layer until
the subchannel is no longer busy with the preceding I/O

DA ranks
subsystem

Server 0
Page cache
3
chpid 2 HBA 2
dasd driver ! Switch
chpid 3 HBA 3 5

Server 1
chpid 4 HBA 4
7



FICON/ECKD dasd I/O to a single disk with PAV (SLES10 / RHEL5)

 VFS sees one device
Application program  The device mapper sees the real device and all alias devices
 Parallel Access Volumes solve the I/O queuing in front of the subchannel
 Each alias device uses its own subchannel
VFS  Additional processor cycles are spent to do the load balancing in the device mapper
 The next slowdown is the fact that only one disk is used in the storage server. This implies the use of
only one rank, one device adapter, one server

Multipath
dm ! DA ! ranks
subsystem

Server 0
Page cache
3
b chpid 2 HBA 2
dasd driver Switch
c chpid 3 HBA 3 5

Server 1
d chpid 4 HBA 4
7



FICON/ECKD dasd I/O to a single disk with HyperPAV (SLES11)

Application program  The dasd driver sees the real device and all alias devices
 Load balancing with HyperPAV and static PAV is done in the dasd driver. The aliases need only to be
added to Linux. The load balancing works better than on the device mapper layer.
VFS  Less additional processor cycles are needed than with Linux multipath.
 The next slowdown is the fact that only one disk is used in the storage server. This implies the use of
only one rank, one device adapter, one server

DA ! ranks
subsystem

Server 0
Page cache
3
b chpid 2 HBA 2
dasd driver Switch
c chpid 3 HBA 3 5

Server 1
d chpid 4 HBA 4
7



FICON/ECKD dasd I/O to a linear or striped logical volume

 VFS sees one device (logical volume)
Application program  The device mapper sees the logical volume and the physical volumes
 Additional processor cycles are spent to map the I/Os to the physical volumes.
 Striped logical volumes require more additional processor cycles than linear logical volumes
VFS  With a striped logical volume the I/Os can be well balanced over the entire storage server and overcome limitations
from a single rank, a single device adapter or a single server
 To ensure that I/O to one physical disk is not limited by one subchannel, PAV or HyperPAV should be used in
combination with logical volumes
LVM
! DA ranks
dm
subsystem

Server 0
Page cache
3
b chpid 2 HBA 2
dasd driver Switch
c chpid 3 HBA 3 5

Server 1
d chpid 4 HBA 4
7



FICON/ECKD dasd I/O to a storage pool striped volume with
HyperPAV (SLES11)

 A storage pool striped volume makes only sense in combination with PAV or HyperPAV
Application program  VFS sees one device
The dasd driver sees the real device and all alias devices

 The storage pool striped volume spans over several ranks of one server and overcomes the limitations of a single rank and / or a
single device adapter
VFS  Storage pool striped volumes can also be used as physical disks for a logical volume to use both server sides.
 To ensure that I/O to one dasd is not limited by one subchannel, PAV or HyperPAV should be used

! DA ranks
subsystem

Server 0
Page cache
extent pool 3
b chpid 2 HBA 2
dasd driver Switch
c chpid 3 HBA 3 5

Server 1
d chpid 4 HBA 4
7



General layout for FCP/SCSI

 The SCSI driver finalizes the I/O requests
Application program
 The zFCP driver adds the FCP protocol to the requests
 The qdio driver transfers the I/O to the channel
VFS
 A host bus adapter connects to both servers
 Each server connects to its ranks
LVM
Multipath DA ranks
dm
Block device layer

Server 0
Page cache
2
I/O scheduler chpid 1 HBA 1

chpid 2 HBA 2 4
SCSI driver Switch
zFCP driver
chpid 3 HBA 3
qdio driver

Server 1
6
chpid 4 HBA 4

8


FCP/SCSI LUN I/O to a single disk

 Assume that disk 3 in rank 8 is reachable via channel 2 and host bus adapter 2
Application program  Up to 32 (default value) I/O requests can be sent out to disk 3 before the first completion is required
 The throughput will be limited by the rank and / or the device adapter
 There is no high availability provided for the connection between the host and the storage server
VFS

DA ranks
Block device layer

Server 0
Page cache
2

chpid 2 HBA 2 4
SCSI driver Switch
zFCP driver
chpid 3 HBA 3
qdio driver
!

Server 1
6
chpid 4 HBA 4

8
26 Linux on System z disk I/O performance
! © 2010 IBM Corporation


FCP/SCSI LUN I/O to a single disk with multipathing

Application program  The device mapper sees the multibus or failover alternatives to the same disk
 Administrational effort is required to define all paths to one disk
VFS  Additional processor cycles are spent to do the mapping to the desired path for the disk
in the device mapper
 Multibus requires more additonal processor cycles than failover

Multipath DA ranks
dm !
Block device layer

Server 0
Page cache
2

chpid 2 HBA 2 4
SCSI driver Switch
zFCP driver
chpid 3 HBA 3
qdio driver

Server 1
6
chpid 4 HBA 4

8


FCP/SCSI LUN I/O to a to a linear or striped logical volume

 VFS sees one device (logical volume)
Application program  The device mapper sees the logical volume and the physical volumes
 Additional processor cycles are spent to do map the I/Os to the physical volumes.
 Striped logical volumes require more additional processor cycles than linear logical volumes
VFS  With a striped logical volume the I/Os can be well balanced over the entire storage server and
overcome limitations from a single rank, a single device adapter or a single server
 To ensure high availability the logical volume should be used in combination with multipathing
LVM
! DA ranks
dm
Block device layer

Server 0
Page cache
2

chpid 2 HBA 2 4
SCSI driver Switch
zFCP driver
chpid 3 HBA 3
qdio driver

Server 1
6
chpid 4 HBA 4

8


FCP/SCSI LUN I/O to a storage pool striped volume with
multipathing

 Storage pool striped volumes make no sense without high availability
Application program  VFS sees one device
 The device mapper sees the multibus or failover alternatives to the same disk
The storage pool striped volume spans over several ranks of one server and overcomes the limitations
VFS 
of a single rank and / or a single device adapter
 Storage pool striped volumes can also be used as physical disks for a logical volume to make use of
both server sides.

Multipath DA ranks
dm !
Block device layer

Server 0
Page cache
2

chpid 2 HBA 2 4
SCSI driver Switch
zFCP driver
chpid 3 HBA 3
qdio driver

Server 1
6
chpid 4 HBA 4

extent pool 8


Summary, recommendations and outlook

 FICON/ECKD
– Storage pool striped disks (no disk placement)
– HyperPAV (SLES11)
– Large volume (RHEL5, SLES11)
– High Performance FICON (SLES11)
 FCP/SCSI
– Linux LV with striping (disk placement)
– Multipath with failover



Trademarks

IBM, the IBM logo, and ibm.com are trademarks or registered trademarks of International Business
Machines Corp., registered in many jurisdictions worldwide. Other product and service names might be
trademarks of IBM or other companies. A current list of IBM trademarks is available on the Web at
“Copyright and trademark information” at www.ibm.com/legal/copytrade.shtml.

Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both.

Java and all Java-based trademarks and logos are trademarks of Sun Microsystems, Inc. in the United
States, other countries, or both.

Other product and service names might be trademarks of IBM or other companies.

Source: If applicable, describe source origin


Linux on System z disk I/O performance

More Related Content

What's hot (15)

Viewers also liked (20)

Similar to Linux on System z disk I/O performance (20)

More from IBM India Smarter Computing (20)

Linux on System z disk I/O performance