This is the latest version of the slides based on my book "Solaris Performance and Tuning" that has been extended to include Linux and many other more recent topics. It has been presented innumerable times, most recently at the CMG conference, Usenix 08 and LISA 08, and this version will be presented at Usenix 09, San Diego on June 16th, along with the Free Tools slides.
Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...
Solaris Linux Performance, Tools and Tuning
1. Solaris/Linux Performance
Measurement, Tools and Tuning
Adrian Cockcroft, acockcroft@netflix.com
May 1, 2009
2009 5/1/09 Page 1
2. Abstract
• This course focuses on the measurement sources and tuning
parameters available in Unix and Linux, including TCP/IP
measurement and tuning, workload analysis, complex storage
subsystems, and with a deep dive on advanced Solaris metrics
such as microstates and extended system accounting.
• The meaning and behavior of metrics is covered in detail.
Common fallacies, misleading indicators, sources of
measurement error and other traps for the unwary will be
exposed.
• Free tools for Capacity Planning are covered in detail in a
different slide deck, interleaved for this event.
• Updated slide decks live at http://www.slideshare.net/adrianco
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 2
3. Sources
• Adrian Cockcroft
– Sun Microsystems 1988-2004, Distinguished Engineer
– eBay Research Labs 2004-2007, Distinguished Engineer
– Netflix 2007, Director - Web Engineering – Personalization Systems
– CMG 2007 Michelson Award Winner for lifetime contribution to
computer measurement
– Note: I am a Netflix employee, but this material does not refer to and
is not endorsed by Netflix. It is based on the author's work over the
last 20+ years.
• Books by the author
– Sun Performance and Tuning, Prentice Hall, 1994, 1998 (2nd Ed)
– Resource Management, Prentice Hall, 2000
– Capacity Planning for Internet Services, Prentice Hall, 2001
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 3
4. Contents
• Capacity Planning and Performance Definitions
• Workload Characteristics and Analysis
• Implications of Virtualization and Cloud Computing
• Metric collection interfaces
• Free Tools for capacity planning (separate slide deck)
• CPU - measurement issues and virtualization
• Network - Internet Servers and TCP/IP essentials
• Memory – The memory-go-round, Swap space instrumentation
• Disks - virtualization, SSDs, filesystems, simple disks and RAID
• Quick tips and Recipes
• References
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 4
6. Capacity Planning Definitions
• Capacity
– Resource utilization and headroom
• Planning
– Predicting future needs by analyzing historical data and
modeling future scenarios
• Performance Monitoring
– Collecting and reporting on performance data
• Unix/Linux (apologies to users of OSX, HP-UX, AIX etc.)
– Emphasis on Solaris and Linux
– Much of the discussion is independent of the OS
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 6
7. Measurement Terms and Definitions
• Bandwidth - gross work per unit time [unattainable]
• Throughput - net work per unit time
• Peak throughput - at maximum acceptable response time
• Utilization - busy time relative to elapsed time [can be misleading]
• Queue length - number of requests waiting
• Service time - time to process a unit of work after waiting
• Response time - time to complete a unit of work including waiting
• Key Performance Indicator (KPI) – a measurement you have
decided to watch because it has some business value
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 7
8. Service Level Agreements (SLA)
• Behavioral goals for the system in terms of KPIs
• Response time target
– Rule of thumb: Estimate 95th percentile response time as
three times mean response time
– e.g. if SLA says 1 second response, measured average
should be less than 333ms
• Utilization Target (a proxy for Response Time)
– Specified as a minimum and maximum
– Minimum utilization target to keep costs down
– Maximum utilization target for good response times and
capacity headroom for future workload fluctuations
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 8
9. Capacity Planning Requirements
• We care about CPU, Memory, Network and Disk resources, and
Application response times
• We need to know how much of each resource we are using
now, and will use in the future
• We need to know how much headroom we have to handle
higher loads
• We want to understand how headroom varies, and how it relates
to application response times and throughput
• The application workload must be characterized so we can
understand and manage system behaviours
• We want to be able to find the bottleneck in an under-performing
system
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 9
11. Workload Characteristics: One by One
Constant Workloads
• e.g. Numerical computation, compute intensive batch
• Trivial to model, utilization and duration define the work
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 11
12. Simple Random Arrivals
• Random arrival of transactions with fixed mean service time
– Little’s Law: QueueLength = Throughput * Response
– Utilization Law: Utilization = Throughput * ServiceTime
• Complex models are often reduced to this model
– By averaging over longer time periods since the formulas only
work if you have stable averages
– By wishful thinking (i.e. how to fool yourself)
• e.g. Unix Load Average is actually CPU Queue Length
– Throughput up a little, load average up a lot = slow system
– So load average is a proxy metric for response time
– High load average per CPU implies slow response times
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 12
13. Mixed random arrivals of transactions
with stable mean service times
• Think of the grocery store checkout analogy
– Trolleys full of shopping vs. baskets full of shopping
– Baskets are quick to service, but get stuck behind trolleys
– Relative mixture of transaction types starts to matter
• Many transactional systems handle a mixture
– Databases, web services
• Consider separating fast and slow transactions
– So that we have a “10 items or less” line just for baskets
– Separate pools of servers for different services
– Don’t mix OLTP with DSS queries in databases
• Performance is often thread-limited
– Thread limit and slow transactions constrains maximum throughput
– Throughput = Queue / ResponseTime
• Model using analytical solvers like PDQ
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 13
14. Load dependent servers – non-stable
mean service times
• Mean service time increases at high throughput
– Due to non-scalable algorithms, lock contention
– System runs out of memory and starts paging or frequent GC
• Systems have “tipping points”
– Hysteresis means they don’t come back when load drops
– This is why you have to kill catatonic systems
• Model using simulation tools like Hyperformix, Opnet
– Behaviour is non-linear and hard to model
– Practical option is to avoid tipping points
– Best designs shed load to be stable at the limit
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 14
15. Self-similar / fractal workloads – bursty
rather than random
• Self-similar
– Looks “random” at close up, stays “random” as you zoom out
– Work arrives in bursts, transactions aren’t independent
– Bursts cluster together in super-bursts, etc.
• Network packet streams tend to be fractal
• Common in practice, too hard to model
– Probably the most common reason why your model is wrong!
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 15
16. State Dependent Services
• Personalized services that store user history
– Transactions for new users are quick
– Transactions for users with lots of state/history are slower
– As user base builds state and ages you get into lots of
trouble…
• Social Networks, Recommendation Services
– Facebook, Flickr, Netflix, Pandora, Twitter etc.
• “Abandon hope all ye who enter here”
– Not tractable to model, repeatable tests are tricky
– Long fat tail response time distribution and timeouts
– Excessively long service times for some users
– Solutions: careful algorithm design, lots of caching
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 16
17. Workload Modelling Survivalism
• Simplify the workload algorithms
– move from hard or impossible to simpler models
– use caching and pre-compute to get constant service times
• Stand further away
– averaging is your friend – gets rid of complex fluctuations
• Minimalist Models
– most models are far too complex – the classic beginners error…
– the art of modelling is to only model what really matters
• Don’t model details you don’t use
– model peak hour of the week, not day to day fluctuations
– e.g. “Will the web site survive next Sunday night?”
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 17
19. Measurement Data Interfaces
• Several generic raw access methods
– Read the kernel directly
– Structured system data
– Process data
– Network data
– Accounting data
– Application data
• Command based data interfaces
– Scrape data from vmstat, iostat, netstat, sar, ps
– Higher overhead, lower resolution, missing metrics
• Data available is always platform and release specific…
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 19
20. Reading kernel memory - kvm
• The only way to get data in very old Unix variants
• Use kernel namelist symbol table and open /dev/kmem
• Solaris wraps up interface in kvm library
• Advantages
– Still the only way to get at some kinds of data
– Low overhead, fast bulk data capture
• Disadvantages
– Too much intimate implementation detail exposed
– No locking protection to ensure consistent data
– Highly non-portable, unstable over releases and patches
– Tools break when kernel moves between 32 and 64bit address
support
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 20
21. Structured Kernel Statistics - kstat
• Solaris 2 introduced kstat and extended usage in each release
• Used by Solaris 2 vmstat, iostat, sar, network interface stats, etc.
• Advantages
– The recommended and supported Solaris metric access API
– Does not require setuid root commands to access for reads
– Individual named metrics stable over releases
– Consistent data using locking, but low overhead
– Unchanged when kernel moves to 64bit address support
– Extensible to add metrics without breaking existing code
• Disadvantages
– Somewhat complex hierarchical kstat_chain structure
– State changes (device online/offline) cause kstat_chain rebuild
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 21
22. Kernel Trace - TNF, Dtrace, ktrace
• Solaris, Linux, Windows and other Unixes have similar features
– Solaris has TNF probes and prex command to control them
– User level probe library for hires tracepoints allows
instrumentation of multithreaded applications
– Kernel level probes allow disk I/O and scheduler tracing
• Advantages
– Low overhead, microsecond resolution
– I/O trace capability is extremely useful
• Disadvantages
– Too much data to process with simple tracing capabilities
– Trace buffer can overflow or cause locking issues
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 22
23. Dtrace – Dynamic Tracing
• One of the most exiting new features in Solaris 10, rave reviews
• Book: quot;Solaris Performance and Toolsquot; by Richard McDougall
and Brendan Gregg
• Advantages
– No overhead when it is not in use
– Low overhead probes can be put anywhere/everywhere
– Trace data is correlated and filtered at source, get exactly the
data you want, very sophisticated data providers included
– Bundled, supported, designed to be safe for production
systems
• Disadvantages
– Solaris specific, but being ported to BSD/Linux
– No high level tools support yet
– Yet another (awk-like) scripting language to learn
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 23
24. Hardware counters
• Solaris cpustat for X86 and UltraSPARC pipeline and cache counters
• Solaris busstat for server backplanes and I/O buses, corestat for
multi-core systems
• Intel Trace Collector, Vampir for Linux
• Most modern CPUs and systems have counters
• Advantages
– See what is really happening, more accurate than kernel stats
– Cache usage useful for tuning code algorithms
– Pipeline usage useful for HPC tuning for megaflops
– Backplane and memory bank usage useful for database servers
• Disadvantages
– Raw data is confusing, lots of architectural background info
needed
– Most tools focus on developer code tuning
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 24
25. Configuration information
• System configuration data comes from too many sources!
– Solaris device tree displayed by prtconf and prtdiag
– Solaris 8 adds dynamic configuration notification device picld
– SCSI device info using iostat -E in Solaris
– Logical volume info from product specific vxprint and metastat
– Hardware RAID info from product specific tools
– Critical storage config info must be accessed over ethernet…
– Linux device tree in /proc is a bit easier to navigate
• It is very hard to combine all this data!
• DMTF CIM objects try to address this, but no-one seems to use them…
• Free tool - Config Engine: http://www.cfengine.org
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 25
26. Application instrumentation Examples
• Oracle V$ Tables – detailed metrics used by many tools
• Apache logging for web services
• ARM standard instrumentation
• Custom do-it-yourself and log file scraping
• Advantages
– Focussed application specific information
– Business metrics are needed to do real capacity planning
• Disadvantages
– No common access methods
– ARM is a collection interface only, vendor specific tools, data
– Very few applications are instrumented, even fewer have support
from performance tools vendors
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 26
27. Kernel values, tunables and defaults
• There is often far too much emphasis on kernel tweaks
– There really are few “magic bullet” tunables
– It rarely makes a significant difference
• Fix the system configuration or tune the application instead!
• Very few adjustable components
– “No user serviceable parts inside”
– But Unix has so much history people think it is like a 70’s car
– Solaris really is dynamic, adaptive and self-tuning
– Most other “traditional Unix” tunables are just advisory limits
– Tweaks may be workarounds for bugs/problems
– Patch or OS release removes the problem - remove the tweak
Solaris Tunable Parameters Reference Manual (if you must…)
– http://docs.sun.com/app/docs/doc/817-0404
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 27
28. Process based data - /proc
• Used by ps, proctool and debuggers, pea.se, proc(1) tools on Solaris
• Solaris and Linux both have /proc/pid/metric hierarchy
• Linux also includes system information in /proc rather than kstat
• Advantages
– The recommended and supported process access API
– Metric data structures reasonably stable over releases
– Consistent data using locking
– Solaris microstate data provides accurate process state timers
• Disadvantages
– High overhead for open/read/close for every process
– Linux reports data as ascii text, Solaris as binary structures
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 28
29. Network protocol data
• Based on a streams module interface in Solaris
• Solaris 2 ndd interface used to configure protocols and interfaces
• Solaris 2 mib interface used by netstat -s and snmpd to get TCP stats etc.
• Advantages
– Individual named metrics reasonably stable over releases
– Consistent data using locking
– Extensible to add metrics without breaking existing code
– Solaris ndd can retune TCP online without reboot
– System data is often also made available via SNMP prototcol
• Disadvantages
– Underlying API is not supported, SNMP access is preferred
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 29
30. Tracing and profiling
• Tracing Tools
– truss - shows system calls made by a process
– sotruss / apitrace - shows shared library calls
– prex - controls TNF tracing for user and kernel code
– snoop/tcpdump – network traces for analysis with wireshark
• Profiling Tools
– Compiler profile feedback using -xprofile=collect and use
– Sampled profile relink using -p and prof/gprof
– Function call tree profile recompile using -pg and gprof
– Shared library call profiling setenv LD_PROFILE and gprof
• Accurate CPU timing for process using /usr/proc/bin/ptime
• Microstate process information using pea.se and pw.se
10:40:16 name lwmx pid ppid uid usr% sys% wait% chld% size rss pf
nis_cachemgr 5 176 1 0 1.40 0.19 0.00 0.00 16320 11584 0.0
jre 1 17255 3184 5743 11.80 0.19 0.00 0.00 178112 110336 0.0
sendmail 1 16751 1 0 1.01 0.43 0.00 0.43 18624 16384 0.0
se.sparc.5.6 1 16741 1186 9506 5.90 0.47 0.00 0.00 16320 14976 0.0
imapd 1 16366 198 5710 6.88 1.09 1.02 0.00 34048 29888 0.1
dtmail 10 16364 9070 5710 0.75 1.12 0.00 0.00 102144 94400 0.0
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 30
31. Free Tools
(See Separate Slide Deck)
http://www.slideshare.net/adrianco
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 31
33. What would you say if you were asked:
How busy is that system?
A: I have no idea…
A: 10%
A: Why do you want to know?
A: I’m sorry, you don’t understand your question….
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 33
34. Headroom Estimation
• CPU Capacity
– Relatively easy to figure out
• Network Usage
– Use bytes not packets/s
• Memory Capacity
– Tricky - easier in Solaris 8
• Disk Capacity
– Can be very complex
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 34
35. Headroom
• Headroom is available usable resources
– Total Capacity minus Peak Utilization and Margin
– Applies to usr+sysRAM, Net, Disk and OS
CPU, CPU for Peak Period
100
Margin
90
80
Headroom
70
60
CPU %
50
40 Utilization
30
20
10
0
Time
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 35
36. Utilization
• Utilization is the proportion of busy time
• Always defined over a time interval
OnCPU Scheduling for Each CPU
Mean CPU Util
OnCPU and
0.56
usr+sys CPU for Peak Period
100
0
90
80 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41
70
Microseconds
60
CPU %
50
40
Utilization
30
20
10
0
Time
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 36
37. Response Time
• Response Time = Queue time + Service time
• The Usual Assumptions…
– Steady state averages
– Random arrivals
– Constant service time
– M servers processing the same queue
• Approximations
– Queue length = Throughput * Response Time (Little's Law)
– Utilzation = Throughput * Service Time (utilization law)
– Response Time = Service Time / (1 - UtilizationM)
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 37
38. Response Time Curves
The traditional view of Utilization as a proxy for response time
Systems with many CPUs can run at higher utilization levels, but degrade more
rapidly when they run out of capacity
Headroom margin should be set according to a response time target.
Response Time Curves R = S / (1 - (U%)m)
10.00
Response Time Increase Factor
9.00
8.00
One CPU
7.00
Two CPUs
6.00
Four CPUs
5.00 Eight CPUs
Headroom 16 CPUs
4.00
margin 32 CPUs
3.00 64 CPUs
2.00
1.00
0.00
0 10 20 30 40 50 60 70 80 90 100
Total System Utilization %
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 38
39. So what's the problem with Utilization?
• Unsafe assumptions! Complex adaptive systems are not simple!
• Random arrivals?
– Bursty traffic with long tail arrival rate distribution
• Constant service time?
– Variable clock rate CPUs, inverse load dependent service time
– Complex transactions, request and response dependent
• M servers processing the same queue?
– Virtual servers with varying non-integral concurrency
– Non-identical servers or CPUs, Hyperthreading, Multicore, NUMA
• Measurement Errors?
– Mechanisms with built in bias, e.g. sampling from the scheduler clock
– Platform and release specific systemic changes in accounting of interrupt time
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 39
40. Variable Clock Rate CPUs
• Laptop and other low power devices do this all the time
– Watch CPU usage of a video application and toggle mains/battery power….
• Server CPU Power Optimization - AMD PowerNow!™
– AMD Opteron server CPU detects overall utilization and reduces clock rate
– Actual speeds vary, but for example could reduce from 2.6GHz to 1.2GHz
– Changes are not understood or reported by operating system metrics
– Speed changes can occur every few milliseconds (thermal shock issues)
– Dual core speed varies per socket, Quad core varies per core
– Quad core can dynamically stop entire cores to save power
• Possible scenario:
– You estimate 20% utilization at 2.6GHz
– You see 45% reported in practice (at 1.2GHz)
– Load doubles, reported utilization drops to 40% (at 2.6GHz)
– Actual mapping of utilization to clock rate is unknown at this point
• Note: Older and quot;low powerquot; Opterons used in blades fix clock rate
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 40
41. Virtual Machine Monitors
• VMware, Xen, IBM LPARs etc.
– Non-integral and non-constant fractions of a machine
– Naiive operating systems and applications that don't expect this
behavior
– However, lots of recent tools development from vendors
• Average CPU count must be reported for each measurement
interval
• VMM overhead varies, application scaling characteristics may
be affected
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 41
42. Threaded CPU Pipelines
• CPU microarchitecture optimizations
– Extra register sets working with one execution pipeline
– When the CPU stalls on a memory read, it switches registers/threads
– Operating system sees multiple schedulable entities (CPUs)
• Intel Hyperthreading
– Each CPU core has an extra thread to use spare cycles
– Typical benefit is 20%, so total capacity is 1.2 CPUs
– I.e. Second thread much slower when first thread is busy
– Hyperthreading aware optimizations in recent operating systems
• Sun “CoolThreads”
– quot;Niagaraquot; SPARC CPU has eight cores, one shared floating point unit
– Each CPU core has four threads, but each core is a very simple design
– Behaves like 32 slow CPUs for integer, snail like uniprocessor for FP
– Overall throughput is very high, performance per watt is exceptional
– Niagara 2 has dedicated FPU and 8 threads per core (total 64 threads)
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 42
43. Measurement Errors
• Mechanisms with built in bias
– e.g. sampling from the scheduler clock underestimates CPU usage
– Solaris 9 and before, Linux, AIX, HP-UX “sampled CPU time”
– Solaris 10 and HP-UX “measured CPU time” far more accurate
– Solaris microstate process accounting always accurate but in Solaris 10
microstates are also used to generate system-wide CPU
• Accounting of interrupt time
– Platform and release specific systemic changes
– Solaris 8 - sampled interrupt time spread over usr/sys/idle
– Solaris 9 - sampled interrupt time accumulated into sys only
– Solaris 10 - accurate interrupt time spread over usr/sys/idle
– Solaris 10 Update 1 - accurate interrupt time in sys only
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 43
44. CPU time measurements
• Biased sample CPU measurements
– See 1998 Paper quot;Unix CPU Time Measurement Errorsquot;
– Microstate measurements are accurate, but are platform and tool specific.
Sampled metrics are more inaccurate at low utilization
• CPU time is sampled by the 100Hz clock interrupt
– sampling theory says this is accurate for an unbiased sample
– the sample is very biased, as the clock also schedules the CPU
– daemons that wakeup on the clock timer can hide in the gaps
– problem gets worse as the CPU gets faster
• Increase clock interrupt rate? (Solaris)
– set hires_tick=1 sets rate to 1000Hz, good for realtime wakeups
– harder to hide CPU usage, but slightly higher overhead
• Use measured CPU time at per-process level
– microstate accounting takes timestamp on each state change
– very accurate and also provides extra information
– still doesn’t allow for interrupt overhead
– Prstat -m and the pea.se command uses this accurate measurement
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 44
45. More CPU Measurement Issues
• Load average differences
– Just includes CPU queue (Solaris)
– Includes CPU and Disk (Linux) – which is a broken metric
• Wait for I/O is a misleading subset of idle time
– Metric removed in Solaris 10 – always zero
– Ignore it in all other Unix/Linux releases
– Only makes sense on uni-processor systems
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 45
46. How to plot Headroom
• Measure and report absolute CPU power if you can get it…
• Plot shows headroom in blue, margin in red, total power tracking day/
night workload variation, plotted as mean + two standard deviations.
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 46
47. “Cockcroft Headroom Plot”
• Scatter plot of response time
(ms) vs. Throughput (KB) from
iostat metrics
• Histograms on axes
• Throughput time series plot
• Shows distributions and shape
of response time
• Fits throughput weighted
inverse gaussian curve
• Coded using quot;Rquot; statistics
package
• Blogged development at
http://perfcap.blogspot.com/search?q=chp
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 47
48. How busy is that system again?
• Check your assumptions…
• Record and plot absolute capacity for each measurement interval
• Plot response time as a function of throughput, not just utilization
• SOA response characteristics are complicated…
• More detailed discussion in CMG06 Paper and blog entries
– “Utilization is Virtually Useless as a Metric” - Adrian Cockcroft - CMG06
http://perfcap.blogspot.com/search?q=utilization
http://perfcap.blogspot.com/search?q=chp
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 48
49. CPU
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 49
50. CPU Capacity Measurements
• CPU Capacity is defined by CPU type and clock rate, or a
benchmark rating like SPECrateInt2000
• CPU throughput - CPU scheduler transaction rate
– measured as the number of voluntary context switches
• CPU Queue length
– CPU load average gives an approximation via a time
decayed average of number of jobs running and ready to run
• CPU response time
– Solaris microstate accounting measures scheduling delay
• CPU utilization
– Defined as busy time divided by elapsed time for each CPU
– Badly distorted and undermined by virtualization……
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 50
51. Controlling and CPUs in Solaris
• psrinfo - show CPU status and clock rate
• Corestat - show internal behavior of multi-core CPUs
• psradm - enable/disable CPUs
• pbind - bind a process to a CPU
• psrset - create sets of CPUs to partition a system
– At least one CPU must remain in the default set, to run kernel services
like NFS threads
– All CPUs still take interrupts from their assigned sources
– Processes can be bound to sets
• mpstat shows per-CPU counters (per set in Solaris 9)
CPU minf mjf xcal intr ithr csw icsw migr smtx srw syscl usr sys wt idl
0 45 1 0 232 0 780 234 106 201 0 950 72 28 0 0
1 29 1 0 243 0 810 243 115 186 0 1045 69 31 0 0
2 27 1 0 235 0 827 243 110 199 0 1000 75 25 0 0
3 26 0 0 217 0 794 227 120 189 0 925 70 30 0 0
4 9 0 0 234 92 403 94 84 1157 0 625 66 34 0 0
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 51
52. Monitoring CPU mutex lock statistics
• To fix mutex contention change the application workload or upgrade to a newer
OS release
• Locking strategies are too complex to be patched
• Lockstat Command
– very powerful and easy to use
– Solaris 8 extends lockstat to include kernel CPU time profiling
– dynamically changes all locks to be instrumented
– displays lots of useful data about which locks are contending
# lockstat sleep 5
Adaptive mutex spin: 3318 events
Count indv cuml rcnt spin Lock Caller
-------------------------------------------------------------------------------
601 18% 18% 1.00 1 flock_lock cleanlocks+0x10
302 9% 27% 1.00 7 0xf597aab0 dev_get_dev_info+0x4c
251 8% 35% 1.00 1 0xf597aab0 mod_rele_dev_by_major+0x2c
245 7% 42% 1.00 3 0xf597aab0 cdev_size+0x74
160 5% 47% 1.00 7 0xf5b3c738 ddi_prop_search_common+0x50
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 52
55. TCP - A Simple Approach
• Capacity and Throughput Metrics to Watch
• Connections
– Current number of established connections
– New outgoing connection rate (active opens)
– Outgoing connection attempt failure rate
– New incoming connection rate (passive opens)
– Incoming connection attempt failure rate (resets)
• Throughput
– Input and output byte rates
– Input and output segment rates
– Output byte retransmit percentage
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 55
56. Obtaining Measurements
• Get the TCP MIB via SNMP or netstat -s
• Standard TCP metric names:
– tcpCurrEstab: current number of established connections
– tcpActiveOpens: number of outgoing connections since boot
– tcpAttemptFails: number of outgoing failures since boot
– tcpPassiveOpens: number of incoming connections since boot
– tcpOutRsts: number of resets sent to reject connection
– tcpEstabResets: resets sent to terminate established
connections
– (tcpOutRsts - tcpEstabResets): incoming connection failures
– tcpOutDataSegs, tcpInDataSegs: data transfer in segments
– tcpRetransSegs: retransmitted segments
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 56
57. Internet Server Issues
• TCP Connections are expensive
– TCP is optimized for reliable data on long lived connections
– Making a connection uses a lot more CPU than moving data
– Connection setup handshake involves several round trip
delays
– Each open connection consumes about 1 KB plus data buffers
• Pending connections cause “listen queue” issues
• Each new connection goes through a “slow start” ramp up
• Other TCP Issues
– TCP windows can limit high latency high speed links
– Lost or delayed data causes time-outs and retransmissions
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 57
58. TCP Sequence Diagram for HTTP Get
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 58
59. Stalled HTTP Get and Persistent HTTP
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 59
61. Memory Capacity Measurements
• Physical Memory Capacity Utilization and Limits
– Kernel memory, Shared Memory segment
– Executable code, stack and heap
– File system cache usage, Unused free memory
• Virtual Memory Capacity - Paging/Swap Space
– When there is no more available swap, Unix stops working
• Memory Throughput
– Hardware counter metrics can track CPU to Memory traffic
– Page in and page out rates
• Memory Response Time
– Platform specific hardware memory latency makes a difference, but
hard to measure
– Time spent waiting for page-in is part of Solaris microstate
accounting
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 61
62. Page Size Optimization
• Systems may support large pages for reduced overhead
– Solaris support is more dynamic/flexible than Linux at present
• Intimate Shared Memory locks large pages in RAM
– No swap space reservation
– Used for large database server Shared Global Area
• No good metrics to track usage and fragmentation issues
• Solaris ppgsz command can set heap and stack pagesize
• SPARC Architecture
– Base page size is 8KB, Large pages are 4MB
• Intel/AMD x86 Architectures
– Base page size is 4KB, Large pages are 2MB
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 62
63. Cache principles
• Temporal locality - “close in time”
– If you need something frequently, keep it near you
– If you don’t use it for a while, put it back
– If you change it, save the change by putting it back
• Spacial locality - “close in space - nearby”
– If you go to get one thing, get other stuff that is nearby
– You may save a trip by prefetching things
– You can waste bandwidth if you fetch too much you don’t use
• Caches work well with randomness
– Randomness prevents worst case behaviour
– Deterministic patterns often cause cache busting accesses
• Very careful cache friendly tuning can give great speedups
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 63
64. The memory go round - Unix/Linux
• Memory usage flows between subsystems
Kernel System V
Memory Shared
Buffers Memory
shm_unlink
kernel
kernel shmget
alloc
free
Head
delete
exit
Free read
brk RAM write
pagein List mmap
reclaim reclaim
Process Filesystem
Stack and Ta il Cache
Heap
pageout pageout
scanner scanner
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 64
65. The memory go round - Solaris 8 and Later
• Memory usage flows between subsystems
Kernel System V
Memory Shared
Buffers Memory
shm_unlink
kernel
kernel shmget
alloc
free
Head
Free RAM List
exit
read
brk
delete write
pagein
mmap
Filesystem
reclaim Cache
Process
Stack and Ta il
Heap
pageout
scanner
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 65
66. Solaris Swap Space
• Swap is very confusing and badly instrumented!
# se swap.se
ani_max 54814 ani_resv 19429 ani_free 37981 availrmem 13859 swapfs_minfree 1972
ramres 11887 swap_resv 19429 swap_alloc 16833 swap_avail 47272 swap_free
49868
Misleading data printed by swap -s
134664 K allocated + 20768 K reserved = 155432 K used, 378176 K available
Corrected labels:
134664 K allocated + 20768 K unallocated = 155432 K reserved, 378176 K available
Mislabelled sar -r 1
freeswap (really swap available) 756352 blocks
Useful swap data:
Total swap 520 M available 369 M reserved 151 M Total disk 428 M Total RAM 92 M
# swap -s
total: 134056k bytes allocated + 20800k reserved = 154856k used, 378752k available
# sar -r 1
18:40:51 freemem freeswap
18:40:52 4152 756912
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 66
67. Disk
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 67
68. Disk Capacity Measurements
• Detailed metrics vary by platform
• Easy for the simple disk cases
• Hard for cached RAID subsystems
• Almost Impossible for shared disk subsystems and SANs
– Another system or volume can be sharing a backend
spindle, when it gets busy your own volume can saturate,
even though you did not change your own workload!
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 68
69. Storage Utilization
• Storage virtualization broke utilization metrics a long time ago
• Host server measures busy time on a quot;diskquot;
– Simple disk, quot;single serverquot; response time gets high near 100%
utilization
– Cached RAID LUN, one I/O stream can report 100% utilization, but
full capacity supports many threads of I/O since there are many
disks and RAM buffering
• New metric - quot;Capability Utilizationquot;
– Adjusted to report proportion of actual capacity for current workload
mix
– Measured by tools such as Ortera Atlas (http://www.ortera.com)
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 69
70. Solaris Filesystem Issues
ufs - standard, reliable, good for lots of small files
ufs with transaction log - faster writes and recovery
tmpfs - fastest if you have enough RAM, volatile
NFS
NFS2 - safe and common, 8KB blocks, slow writes
NFS3 - more readahead and writebehind, faster
default 32KB block size - fast sequential, may be slow random
default TCP instead of UDP, more robust over WAN
NFS4 - adds stateful behavior
cachefs - good for read-mostly NFS speedup
Veritas VxFS - useful on old Solaris releases
Solaris 8 UFS Upgrade
ufs was extended to be more competitive with VxFS
transaction log unbuffered direct access option and snapshot backup capability
now available “for free” with Solaris 8
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 70
71. Solaris 10 ZFS - What it doesn't have....
• Nice features
– No extra cost - its bundled in a free OS
– No volume manager - its built in
– No space management - file systems use a common pool
– No long wait for newfs to finish - create a 3TB file system in a second
– No fsck - its transactional commit means its consistent on disk
– No slow writes - disk write caches are enabled and flushed reliably
– No random or small writes - all writes are large batched sequential
– No rsync - snapshots can be differenced and replicated remotely
– No silent data corruption - all data is checksummed as it is read
– No bad archives - all the data in the file system is scrubbed regularly
– No penalty for software RAID - RAID-Z has a clever optimization
– No downtime - mirroring, RAID-Z and hot spares
– No immediate maintenance - double parity disks if you need them
• Wish-list
– No way to know how much performance headroom you have!
– No clustering support
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 71
72. Linux Filesystems
• There are a large number of options!
– http://en.wikipedia.org/wiki/Comparison_of_file_systems
• EXT3
– Common default for many Linux distributions
– Efficient for CPU and space, small block size
– relatively simple for reliability and recovery
– Journalling support options can improve performance
– EXT4 came out of development at the end of 2008
• XFS
– Based on Silicon Graphics XFS, mature and reliable
– Better for large files and streaming throughput
– High Performance Computing heritage
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 72
73. Disk Configurations
• Sequential access is ~10 times faster than random
– Sequential rates are now about 50-100 MB/s per disk
– Random rates are 166 operations/sec, (250/sec at 15000rpm)
– The size of each random read should be as big as possible
• Reads should be cached in main memory
– “The only good fast read is the one you didn’t have to do”
– Database shared memory or filesystem cache is microseconds
– Disk subsystem cache is milliseconds, plus extra CPU load
– Underlying disk is ~6ms, as its unlikely that data is in cache
• Writes should be cached in nonvolatile storage
– Allows write cancellation and coalescing optimizations
– NVRAM inside the system - Direct access to Flash storage
– Solid State Disks based on Flash are the quot;Next Big Thingquot;
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 73
77. Simple Disks
• Utilization shows capacity usage
Measured using iostat %b
• Response time is svc_t
svc_t increases due to waiting in the queues caused by bursty
loads
• Service time per I/O is Util/IOPS
Calculate as(%b/100)/(rps+wps)
Decreases due to optimization of queued requests as load
increases
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 77
78. Single Disk Parameters
• e.g. Seagate 18GB ST318203FC
– Obtain from www.seagate.com
– RPM = 10000 = 6.0ms = 166/s
– Avg read seek = 5.2ms
– Avg write seek = 6.0ms
– Avg transfer rate = 24.5 MB/s
– Random IOPS
• Approx 166/s for small requests
• Approx 24.5/size for large requests
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 78
79. Mirrored Disks
• All writes go to both disks
• Read policy alternatives
– All reads from one side
– Alternate from side to side
– Split by block number to reduce seek
– Read both and use first to respond
• Simple Capacity Assumption
– Assume duplicated interconnects
– Same capacity as unmirrored
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 79
80. Concatenated and Fat
Stripe Disks
• Request size less than interlace
• Requests go to one disk
• Single threaded requests
– Same capacity as single disk
• Multithreaded requests
– Same service time as one disk
– Throughput of N disks if more than N threads are evenly
distributed
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 80
81. Striped Disks
• Request size more than interlace
• Requests split over N disks
– Single and multithreaded requests
– N = request size / interlace
– Throughput of N disks
• Service Time Reduction
– Reduced size of request reduces service time for large
transfers
– Need to wait for all disks to complete - slowest dominates
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 81
82. RAID5 for Small
Requests
log
• Writes must calculate parity
– Read parity and old data blocks
– Calculate new parity
– Write log and data and parity
– Triple service time
– One third throughput of one disk
• Read performs like stripe
– Throughput of N-1, service of one
– Degraded mode throughput about one
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 82
83. RAID5 for Large
Requests
log
• Write full stripe and parity
• Capacity similar to stripe
– Similar read and write performance
– Throughput of N-1 disks
– Service time for size reduced by N-1
– Less interconnect load than mirror
• Degraded Mode
– Throughput halved and service similar
– Extra CPU used to regenerate data
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 83
84. Cached RAID5
• Nonvolatile cache
– No need for recovery log disk
• Fast service time for writes
– Interconnect transfer time only
• Cache optimizes RAID5
– Makes all backend writes full stripe
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 84
85. Cached Stripe
• Write caching for stripes
– Greatly reduced service time
– Very worthwhile for small transfers
– Large transfers should not be cached
– In many cases, 128KB is crossover point from small to large
• Optimizations
– Rewriting same block cancels in cache
– Small sequential writes coalesce
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 85
86. Capacity Model Measurements
• Derived from iostat outputs
extended disk statistics
disk r/s w/s Kr/s Kw/s wait actv svc_t %w %b
sd9 33.1 8.7 271.4 71.3 0.0 2.3 15.8 0 27
• Utilization U = %b / 100 = 0.27
• Throughput X = r/s + w/s = 41.8
• Size K = Kr/s + Kw/s / X = 8.2K
• Concurrency N = actv = 2.3
• Service time S = U / X = 6.5ms
• Response time R = svc_t = 15.8ms
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 86
87. Cache Throughput
• Hard to model clustering and write cancellation
improvements
• Make pessimistic assumption that throughput is unchanged
• Primary benefit of cache is fast response time
• Writes can flood cache and saturate back-end disks
– Service times suddenly go from 3ms to 300ms
– Very hard to figure out when this will happen
– Paranoia is a good policy….
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 87
88. Concluding Summary
Walk out of here with the most useful content fresh in your mind!
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 88
89. Quick Tips #1 - Disk
• The system will usually have a disk bottleneck
• Track how busy is the busiest disk of all
• Look for unbalanced, busy or slow disks with iostat
• Options: timestamp, look for busy controllers, ignore idle disks:
% iostat -xnzCM -T d 30
Tue Jan 21 09:19:21 2003 extended device statistics
r/s w/s Mr/s Mw/s wait actv wsvc_t asvc_t %w %b device
141.0 8.6 0.6 0.0 0.0 1.5 0.0 10.0 0 25 c0
3.3 0.0 0.0 0.0 0.0 0.0 0.0 6.5 0 2 c0t0d0
137.7 8.6 0.6 0.0 0.0 1.5 0.0 10.1 0 74 c0t1d0
Watch out for sd_max_throttle limiting throughput when set too low
Watch out for RAID cache being flooded on writes, causes sudden very
large increase in write service time
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 89
90. Quick Tips #2 - Network
• If you ever see a slow machine that also appears to be idle, you should
suspect a network lookup problem. i.e. the system is waiting for some
other system to respond.
• Poor Network Filesystem response times may be hard to see
– Use iostat -xn 30 on a Solaris client
– wsvc_t is the time spent in the client waiting to send a request
– asvc_t is the time spent in the server responding
– %b will show 100% whenever any requests are being processed, it does NOT
mean that the network server is maxed out, as an NFS server is a complex
system that can serve many requests at once.
• Name server delays are also hard to detect
– Overloaded LDAP or NIS servers can cause problems
– DNS configuration errors or server problems often cause 30s delays as the
request times out
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 90
91. Quick Tips #3 - Memory
• Avoid the common vmstat misconceptions
– The first line is average since boot, so ignore it
• Linux, Other Unix and earlier Solaris Releases
– Ignore “free” memory
– Use high page scanner “sr” activity as your RAM shortage indicator
• Solaris 8 and Later Releases
– Use “free” memory to see how much is left for code to use
– Use non-zero page scanner “sr” activity as your RAM shortage indicator
• Don’t panic when you see page-ins and page-outs in vmstat
• Normal filesystem activity uses paging
solaris9% vmstat 30
kthr memory page disk faults cpu
rbw swap free re mf pi po fr de sr f0 s0 s1 s6 in sy cs us sy id
0 0 0 2367832 91768 3 31 2 1 1 0 0 0 0 0 0 511 404 350 0 0 99
0 0 0 2332728 75704 3 29 0 0 0 0 0 0 0 0 0 508 537 410 0 0 99
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 91
92. Quick Tips #4 - CPU
• Look for a long run queue (vmstat procs r) - and add CPUs
– To speedup with a zero run queue you need faster CPUs, not more of them
• Check for CPU system time dominating user time
– Most systems should have lots more Usr than Sys, as they are running
application code
– But... dedicated NFS servers should be 100% Sys
– And... dedicated web servers have high Sys as well
– So... assume that lots of network service drives Sys time
• Watch out for processes that hog the CPU
– Big problem on user desktop systems - look for looping web browsers
– Web search engines may get queries that loop
– Use resource management or limit cputime (ulimit -t) in startup scripts to
terminate web queries
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 92
93. Quick Tips #5 - I/O Wait
• Look for processes blocked waiting for disk I/O (vmstat procs b)
– This is what causes CPU time to be counted as wait not idle
– Nothing else ever causes CPU wait time!
• CPU wait time is a subset of idle time, consumes no resources
– CPU wait time is not calculated properly on multiprocessor machines
on older Solaris releases, it is greatly inflated!
– CPU wait time is no longer calculated, zero in Solaris 10
– Bottom line - don’t worry about CPU wait time, it’s a broken metric
• Look at individual process wait time using microstates
– prstat -m or SE toolkit process monitoring
• Look at I/O wait time using iostat asvc_t
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 93
94. Quick Tips #6 - iostat
• For Solaris remember “expenses” iostat -xPncez 30
• Add -M for Megabytes, and -T d for timestamped logging
• Use 30 second interval to avoid spikes in load. Watch
asvc_t which is the response time for Solaris
• Look for regular disks over 5% busy that have response
times of more than 10ms as a problem.
• If you have cached hardware RAID, look for response
times of more than 5ms as a problem.
• Ignore large response times on idle disks that have
filesystems - its not a problem and the cause is the fsflush
process
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 94
95. Recipe to fix a slow system
• Essential Background Information
– What is the business function of the system?
– Who and where are the users?
– Who says there is a problem, and what is slow?
– What changed recently and what is on the way?
• What is the system configuration?
– CPU/RAM/Disk/Net/OS/Patches, what application software is in use?
• What are the busy processes on the system doing?
– use top, prstat, pea.se or /usr/ucb/ps uax | head
• Report CPU and disk utilization levels, iostat -xPncezM -T d 30
– What is making the disks busy?
• What is the network name service configuration?
– How much network activity is there? Use netstat -i 30 or nx.se 30
• Is there enough memory?
– Check free memory and the scan rate with vmstat 30
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 95
96. Further Reading - Books
General Solaris/Unix/Linux Performance Tuning
– System Performance Tuning (2nd Edition) by Gian-Paolo D. Musumeci and Mike
Loukides; O'Reilly & Associates
Solaris Performance Tuning Books
– Solaris Performance and Tools, Richard McDougall, Jim Mauro, Brendan Gregg;
Prentice Hall
– Configuring and Tuning Databases on the Solaris Platform, Allan Packer; Prentice Hall
– Sun Performance and Tuning, by Adrian Cockcroft and Rich Pettit; Prentice Hall
Sun BluePrints™
– Capacity Planning for Internet Services, Adrian Cockcroft and Bill Walker; Prentice Hall
– Resource Management, Richard McDougall, Adrian Cockcroft et al. Prentice Hall
Linux
– Linux Performance Tuning and Capacity Planning by Jason R. Fink and Matthew D.
Sherer
– Google has a Linux specific search mode http://www.google.com/linux
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 96
97. Questions?
(The End)
2009 Solaris/Linux Performance Measurement and Tuning 5/1/09 Slide 97