page 1 kernel tuning introduction hp-ux kernel tuning a primer bill hassell hp-ux consultant

Kernel Tuning IntroductionKernel Tuning IntroductionPage 1

HP-UX Kernel Tuning HP-UX Kernel Tuning A PrimerA Primer

Bill HassellBill HassellHP-UX ConsultantHP-UX Consultant


Introduction Introduction

What is kernel ‘tuning’?What is kernel ‘tuning’? Parameters and their effectsParameters and their effects What tuning cannot doWhat tuning cannot do


VocabularyVocabulary Kernel fileKernel file

» /stand/system (and /stand/build/system)/stand/system (and /stand/build/system)» Tunable parametersTunable parameters» sysdefsysdef

Measurement toolsMeasurement tools» sar, vmstat, Glancesar, vmstat, Glance


Parameter changingParameter changing SAMSAM

» easiest to useeasiest to use» built-in documentationbuilt-in documentation» some bounds checking and interaction some bounds checking and interaction

teststests Manual methodManual method

» manual editmanual edit» no boundary checksno boundary checks


Filesystem ParametersFilesystem Parameters nfile, nflocksnfile, nflocks ninode, vx_ncsize, ninode, vx_ncsize, ncdnodencdnode

bufpages, nbuf, bufpages, nbuf, dbc_max_pct, dbc_min_pctdbc_max_pct, dbc_min_pct

maxfiles, maxfiles_limmaxfiles, maxfiles_lim fs_async, fs_async, default_disk_irdefault_disk_ir

disksort_secondsdisksort_seconds


nfile - max file opensnfile - max file opens

Maximum number of file opensMaximum number of file opens» open counted for same files or differentopen counted for same files or different» min 3 per processmin 3 per process

stdinstdin stdoutstdout stderrstderr

file: table is fullfile: table is full


nflocks - max file locksnflocks - max file locks

Maximum number of open file locksMaximum number of open file locks Highly application-dependentHighly application-dependent

» one file may have several locksone file may have several locks» Databases may need hundredsDatabases may need hundreds


ninode - kernel inode cacheninode - kernel inode cache In-core cache of In-core cache of uniqueunique current and current and

recent HFS inodes locationsrecent HFS inodes locations Speeds reopens, multi-process file accessSpeeds reopens, multi-process file access Keep small for typical databases, larger Keep small for typical databases, larger

for development and NFS (HFS only)for development and NFS (HFS only) ninode < nfileninode < nfileinode: table is fullinode: table is full

Formula may be way too large (vxfs)Formula may be way too large (vxfs)


ncdnodencdnode

In-core cache of In-core cache of uniqueunique current and current and recent CDROM inodesrecent CDROM inodes

Speeds reopens, multi-process file accessSpeeds reopens, multi-process file access Generally very small (one user for Generally very small (one user for

CDROM)CDROM) Copy CD’s to hard disk for multi-user Copy CD’s to hard disk for multi-user

accessaccess


bufpages, nbuf, dbc_max_pct, bufpages, nbuf, dbc_max_pct, dbc_min_pct - buffer cachedbc_min_pct - buffer cache

Buffercache similar to DOS Buffercache similar to DOS SMARTDRVSMARTDRV

used only with file access, not rawused only with file access, not raw nbuf non-zero not recommendednbuf non-zero not recommended fix with bufpagesfix with bufpages range 200 Mb to 2 Gb (rd/wt ratio)range 200 Mb to 2 Gb (rd/wt ratio) Dynamic Buffer Cache min/max %Dynamic Buffer Cache min/max %


maxfiles, maxfiles_limmaxfiles, maxfiles_lim

Single process file open limitsSingle process file open limits defaults to 60defaults to 60 soft limit, extend with soft limit, extend with setrlimitsetrlimit(2) up (2) up

to to maxfiles_limmaxfiles_lim POSIX shell: ulimit -fPOSIX shell: ulimit -f


fs_async - sync policyfs_async - sync policy

Async filesystem writes (inc. metadata)Async filesystem writes (inc. metadata) 0 = safe for panics, powerfails0 = safe for panics, powerfails 1 = risky (filesystem corruption is likely) 1 = risky (filesystem corruption is likely)

but can improve writes (only) by 30%but can improve writes (only) by 30%


default_disk_ir - immediate default_disk_ir - immediate reportingreporting

Disk Immediate Reporting - no wait for Disk Immediate Reporting - no wait for write to completewrite to complete

Similar to Similar to fs_asyncfs_async except applies to all except applies to all disk writes including rawdisk writes including raw

Kernel param or scsictl by diskKernel param or scsictl by disk


disksort_secondsdisksort_seconds

HP-UX gives priority to serial rd/wtHP-UX gives priority to serial rd/wt Intense serial I/O slows random I/OIntense serial I/O slows random I/O Value to wait before changing priorityValue to wait before changing priority


Process ParametersProcess Parameters

nproc, maxuprcnproc, maxuprc maxdsiz, maxdsiz_64bitmaxdsiz, maxdsiz_64bit maxssiz, maxssiz_64bitmaxssiz, maxssiz_64bit maxtsiz, maxtsiz_64bitmaxtsiz, maxtsiz_64bit


nproc, maxuprc - processesnproc, maxuprc - processes

nproc nproc = max number of processes= max number of processes» proc: table is fullproc: table is full

maxuprc maxuprc = max processes per UID= max processes per UID» collective UID processes, not per logincollective UID processes, not per login


maxdsiz - data segmentmaxdsiz - data segment Maximum data segment sizeMaximum data segment size Both default to 64 megs (typically too Both default to 64 megs (typically too

small)small) Set to apx. 900 megs or 1750 megs Set to apx. 900 megs or 1750 megs no kernel size penalty (just a fence for no kernel size penalty (just a fence for

runaways)runaways) maxdsiz_64bit max is 4 Tbmaxdsiz_64bit max is 4 Tb MAGIC notes (/usr/share/doc)MAGIC notes (/usr/share/doc)


maxssiz - stack sizemaxssiz - stack size

Maximum stack sizeMaximum stack size runaway recursion can exceedrunaway recursion can exceed 79 megs is a 10.20 kernel limit, 200 79 megs is a 10.20 kernel limit, 200

megs for 32 bit 11.0 kernels and 64bit megs for 32 bit 11.0 kernels and 64bit 11.0 allows up to 1 Gb11.0 allows up to 1 Gb

FORTRAN arrays passed as data can FORTRAN arrays passed as data can easily exceed - use COMMONeasily exceed - use COMMON


maxtsiz - text sizemaxtsiz - text size Unchanging executable instructionsUnchanging executable instructions Directly related to file sizeDirectly related to file size Seldom needs changing unless a process has an Seldom needs changing unless a process has an

exceptionally large number of instructionsexceptionally large number of instructions 10.20:10.20:

» 64 Mb default, 4 Gb max64 Mb default, 4 Gb max 11.x:11.x:

» 64 Mb default, 2 Gb max (32 bit), 4 Tb max (64 bit)64 Mb default, 2 Gb max (32 bit), 4 Tb max (64 bit)


Network ParametersNetwork Parameters NetworkingNetworking

» nninni number of network interfacesnumber of network interfaces multiple LAN cardsmultiple LAN cards SLIP/CSLIP/PPP interfacesSLIP/CSLIP/PPP interfaces

» nettune nettune (10.20),(10.20), ndd ndd (11.0)(11.0) Setup in /etc/rc.config.dSetup in /etc/rc.config.d


Virtual Memory (swap)Virtual Memory (swap) maxswapchunksmaxswapchunks swchunkswchunk nswapdev, nswapfsnswapdev, nswapfs swapmem_onswapmem_on


maxswapchunks, swchunk - maxswapchunks, swchunk - Swap SpaceSwap Space

Maximum addressing limits for swapMaximum addressing limits for swap Usable swap is defined as: Usable swap is defined as:

» maxswapchunks * swchunk * dev_bsize where: maxswapchunks * swchunk * dev_bsize where: swchunk=2048 and dev_bsize=1024swchunk=2048 and dev_bsize=1024

Leave swchunk to defaultLeave swchunk to default Formula simplifies to:Formula simplifies to:

» maxswapchunks=DESIRED-SWAP / 2097152 or maxswapchunks=DESIRED-SWAP / 2097152 or » Maxswapchunks=DESIRED-SWAP / 2megsMaxswapchunks=DESIRED-SWAP / 2megs


nswapdev, nswapfsnswapdev, nswapfs

nswapdevnswapdev - max number of swap - max number of swap devicesdevices

nswapfs - max number of filesystems that nswapfs - max number of filesystems that will be used for swapwill be used for swap» filesystem swap performancefilesystem swap performance» one way assignment policy (not returned)one way assignment policy (not returned)


swapmem_onswapmem_on

HP-UX normally needs 1:1 RAM:swapHP-UX normally needs 1:1 RAM:swap swapmem_onswapmem_on creates an overallocation creates an overallocation

policy, typically 75% of RAMpolicy, typically 75% of RAM Can be used in both low RAM and high Can be used in both low RAM and high

RAM systemsRAM systems



SWAPSWAP500 meg500 meg

RAMRAM1 Gb1 Gb

Unusable forUnusable for processesprocesses

In this example, only 500 megs is usable In this example, only 500 megs is usable for processes since Virtual Memory is for processes since Virtual Memory is only 500 megs.only 500 megs.

swapmem_on=0swapmem_on=0



SWAPSWAP1 Gb1 Gb

RAMRAM1 Gb1 Gb

In this example, 1 Gb is usable for In this example, 1 Gb is usable for processes since Virtual Memory is 1 Gb processes since Virtual Memory is 1 Gb too...but no paging neededtoo...but no paging needed




SWAPSWAP500 meg500 meg

RAMRAM1 Gb1 Gb

In this example, 1.25 Gb is usable In this example, 1.25 Gb is usable for processes since Virtual for processes since Virtual Memory is 75% RAM + swap Memory is 75% RAM + swap (750 + 500)(750 + 500)


Total Total Vmemory = Vmemory = 1250 megs1250 megs


Miscellaneous ParametersMiscellaneous Parameters MiscellaneousMiscellaneous

» timezonetimezone, , dstdst (0,1,2,3 policy) (0,1,2,3 policy)» nptynpty (SAM vs. manual) (SAM vs. manual)» timeslicetimeslice

Non-parameters:Non-parameters:» maxusersmaxusers (pseudo/formula parameter) (pseudo/formula parameter)


Some last tipsSome last tips Web help: Web help: docs.hp.com//hpux/onlinedocs/os/KCparams.OverviewAll.htmldocs.hp.com//hpux/onlinedocs/os/KCparams.OverviewAll.html

SysAdmin CoursesSysAdmin Courses SAM help (10.xx+)SAM help (10.xx+) Adjust carefullyAdjust carefully

» Small changesSmall changes» Few changes per sysgenFew changes per sysgen

page 1 kernel tuning introduction hp-ux kernel tuning a primer bill hassell hp-ux consultant

Documents